Short Answer Questions - Nursing Essays

Reading 1(Neural inhibition enables selection during

language processing)

Reading 2(Neuroscience Needs Behavior:

Correcting a Reductionist Bias)

Questions

Neural inhibition enables selection during language processing(Reading 1 questions)

1. What type of neurons are involved in inhibition in this model?

2. What basic psychological process is the main focus of this paper – i.e., what is the primary cognitive construct of interest? Describe how this process is relevant in everyday life? How is this process relevant to psychological disorders?

3. (Fig.1) Looking at panel a, in your own words, explain why “scissors” and “cut” have a lower retrieval demand than “giraffe” and “eat”. Why do “scissors” and “cut” have a lower selection demand than “cat” and “purr, meow, lick”.

4. (Fig.1)In your own words, describe the two dimensions that the authors varied in the task.

5. (Fig.1)Based on the pattern of results displayed in panels b and c here, and the associated statistics described in the text, describe the main effect of selection demand in your own words. Then describe the main effect of retrieval demand. After you’ve described both of those main effects, clearly describe the interaction between selection demand and retrieval demand that is observed here.

6. (Fig. 2) In which layer of the model is selection (using inhibition) occuring? Use meaningful layer labels rather than layer numbers.

7. Summarize the correspondence between the predictions of the model (fig 2c) and the human behavioral data (fig 1c).

8. In your own words, describe whether competitive inhibition in the model helps or hurts selection. Explain how.

9. Which neurotransmitter system does midazolam act on, and what impact does it have on activity in this system?

10. Describe the effects of anxiety on neural activity shown here, as measured by fMRI. What are the implications for VLPFC’s role in selection?

11. What is the relationship between anxiety and selection costs described here, and between anxiety and verb generation? What about midazolam – What are the effects of midazolam on selection costs here, and on verb generation?

12. Although not explicitly addressed in the article, why was it important to study the effects of anxiety on selection with a non-emotional task?

Neuroscience Needs Behavior: Correcting a Reductionist Bias (Reading 2)

1. Describe reductionismin the context of cognitive neuroscience. If you’re not familiar with the term, read on, then return to this question as you formulate your answer. Which of Marr’s levels would a reductionist primarily focus on?

2. These authors are advocating for a larger role of which of Marr’s levels of description in cognitive neuroscience?

3. In your opinion, which of the hypothetical mappings in Figure 1 would provide the strongest evidence for this argument, and why?

4. (Fig. 1) Which of the hypothetical mappings here illustrates the point that “we have no prior knowledge of what the relevant level of brain organization is for any given behavior”, and how does it illustrate that point?

5. (Fig.1) This panel illustrates a potential problem with neuroscience involving only animal models. Why is this problem less of a threat for cognitive neuroscience with human subjects?

6. (Fig. 1) How do the hypothetical problems depicted in these two panels fundamentally undermine the logic used in cognitive neuroscience? Ground your answer in a concrete example, if you find it helps you formulate your answer.

7. In your own words, describe a possible example of the substitution bias in cognitive neuroscience. Make sure to clearly specify the initial question, as well as the simpler question that is substituted.

8. What does “methodological decathlon” mean in this context? Include concrete examples in your explanation.

9. In your own words, explain the difference between description and understanding, as described here. Use the chess analogy, the mirror neuron example, or your own analogy (clearly explained), in your answer.

10. Explain how figures 1A and 1D illustrate this point.

11. After reading the rest of this section, come back to this prompt to describe your understanding of the concept of emergence in this context as clearly as possible.

12. Pick an example human behavior that you don’t already see included in comments here.

Using the flight analogy depicted in Figure 2, explain the importance of studying the behavior at all three of Marr’s levels to gain real understanding of how it works.

13. Which of the mappings depicted in Figure 1 most seriously undermines the logic depicted here in Figure 3, and why?

14. Which of Tinberg’s questions an Aristotle’s causes are most similar to one of Marr’s levels, and why?

Neural inhibition enables selection during
language processing
Hannah R. Snyder, Natalie Hutchison, Erika Nyhus, Tim Curran, Marie T. Banich, Randall C. O’Reilly,
and Yuko Munakata1
Department of Psychology and Neuroscience, University of Colorado, Boulder, CO 80309
Edited by Edward E. Smith, Columbia University, New York, NY, and approved August 10, 2010 (received for review February 24, 2010)
| midazolam | ventrolateral prefrontal cortex | neural
P
eople claim to love the freedom of unlimited choices, but in
reality we are often stymied by too many options and disconcerted by not knowing what the outcomes of our choices will
be. Selecting among multiple options is effortful and time consuming, whether choosing among fruit jams (1), retirement plans
(2), or medical treatments (3). This problem is particularly pervasive during language production, when we must constantly
choose words to express a thought (4–6). People with anxiety
disorders ﬁnd coping with too many options particularly difﬁcult,
and struggle with decision-making problems (7), indecisiveness
(8), and intolerance of uncertainty (9). What mechanisms allow us
to select among multiple options when speaking, and why does
this process break down in people with anxiety? Current psychological theories of selection focus on the importance of cognitive control (5) and prefrontal cortical regions (4, 6, 10), but do
not address questions at the level of speciﬁc neural mechanisms.
We address these questions by implementing a uniﬁed, biologically plausible computational model of selection, and testing
its predictions about both brain and behavior in humans making
choices in a well-controlled language production task.
Our model demonstrates how competitive, inhibitory dynamics
among neurons in prefrontal cortical networks (11) support selection between alternatives. Speciﬁcally, these competitive dynamics serve to sharpen cognitive representations by amplifying
activity in the most active, task-relevant, representations (e.g., the
most appropriate word to complete a sentence) and suppressing
competing representations (e.g., for the many other word possibilities). A tenet of the model is that these critical dynamics occur
via inhibitory, GABAergic interneurons (12–14). Here we test the
predictions of the model regarding selection processes that are
supported by the left ventrolateral prefrontal cortex. We do so for
the following reasons: (i) this region has been implicated in
selecting among competing alternatives during language processing (6, 15–18); (ii) this region shows altered activity in individuals
who suffer from anxiety (particularly anxious apprehension,
characterized by worry, and hereafter referred to as “anxiety” (19);
and (iii) GABAergic function is reduced in individuals with anxiety
(20–23). Hence, our model provides a uniﬁed framework that can
link these formerly disconnected observations.
www.pnas.org/cgi/doi/10.1073/pnas.1002291107
Results
We investigated selection and retrieval of words from semantic
memory with a verb generation task, in which participants are
presented with a noun and say the ﬁrst associated verb that comes
to mind (5, 6). This task allows selection and retrieval demands to
be precisely quantiﬁed and independently manipulated (Fig. 1A
and Methods). We present the basic behavioral effects of selection
and retrieval demands and their simulation in a neural network
model ﬁrst, followed by predictions from the model and empirical
tests of the effects of increasing and decreasing neural inhibition.
For the basic behavioral data, verb generation data were analyzed with a 2 × 2 repeated-measures ANOVA. We replicated
previous results showing participants are slowed by (i) greater
selection demand, measured in terms of the competition between alternative responses (e.g., cat brings to mind many possible competing verbs such as purr, meow, and lick, and so has
Author contributions: H.R.S., T.C., M.T.B., R.C.O., and Y.M. designed research; H.R.S., N.H.,
and E.N. performed research; H.R.S. analyzed data; and H.R.S., T.C., M.T.B., R.C.O., and Y.M.
wrote the paper.
The authors declare no conﬂict of interest.
This article is a PNAS Direct Submission.
1
To whom correspondence should be addressed. E-mail: munakata@colorado.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1002291107/-/DCSupplemental.
PNAS | September 21, 2010 | vol. 107 | no. 38 | 16483–16488
NEUROSCIENCE
|
anxiety choice
network model
Speciﬁcally, our model demonstrates how reduced GABAergic
function can lead to reduced competitive dynamics in prefrontal
cortical networks, allowing nonwinning competitors (alternative
responses that are not selected) to become more active and to
compete over a longer period, which impairs selection. Conversely, increased GABAergic function leads to lower and briefer
activation of these competitors and to improvements in selection.
These basic mechanisms provide a uniﬁed framework for understanding how we make choices in language processing and how
this process relates to worrying about the future, to a degree of precision that allows us to test (and conﬁrm) predictions through neuropharmacological manipulation, links to psychopathology, and levels
of brain activity. Whereas decision-making deﬁcits in persons with
anxiety have previously been shown in complex or affective tasks,
our model predicts that selection deﬁcits that lie at the core of these
problems should be observed even in a simple language-production
task, whereas other cognitive processes should remain intact.
We provide a previously undescribed direct test of this model,
and conﬁrm its predictions. We demonstrate that participants
high in anxiety, which has been linked to reduced GABAergic
function, have more difﬁculty selecting between competing word
options and exhibit reduced VLPFC activation during such selection. In addition, we demonstrate that drugs that increase
GABAergic function improve selection in a nonclinical population. Furthermore, such effects are speciﬁc, as the retrieval of
words from semantic memory is unaffected by GABAergic function when selection demands are low, despite the importance of
retrieval processes in language processing and the role of VLPFC
in retrieval (4).
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
Whether grocery shopping or choosing words to express a thought,
selecting between options can be challenging, especially for people
with anxiety. We investigate the neural mechanisms supporting
selection during language processing and its breakdown in anxiety.
Our neural network simulations demonstrate a critical role for
competitive, inhibitory dynamics supported by GABAergic interneurons. As predicted by our model, we ﬁnd that anxiety (associated with reduced neural inhibition) impairs selection among
options and associated prefrontal cortical activity, even in a simple,
nonaffective verb-generation task, and the GABA agonist midazolam (which increases neural inhibition) improves selection, whereas
retrieval from semantic memory is unaffected when selection
demands are low. Neural inhibition is key to choosing our words.
Low
Retrieval Demand
High
Retrieval Demand
A
C
Output
Low
Selection
Demand
Scissors
Cut
Giraffe
eat
Posterior
Cortex
Cat
Hold
Meow
Shelf
Lick
8.0
7.9
7.8
7.7
7.6
7.5
7.4
7.3
7.2
7.1
7.0
Store
C
Low
High
High
Low
Selection Selection Selection Selection
Demand Demand Demand Demand
Low Retrieval
High Retrieval
Demand
Demand
Selection Cost (log RT Diff., +/- SE)
log Reaction Times (+/- SE)
B
Dust
0.25
0.2
0.15
0.1
0.05
0
Low
High
Retrieval Retrieval
Demand Demand
Fig. 1. Design and basic behavioral ﬁndings for the verb generation task.
(A) Selection demands (high vs. low competition) are crossed with retrieval
demands (high vs. low association strength) (SI Methods 2.1). (B) Participants
take longer to generate a response when retrieval demands are high and
when selection demands are high. (C) Selection costs (RT difference between
high and low selection demand conditions) are greater when retrieval
demands are low than when they are high. All error bars are SEs.
high selection demand, whereas scissors generally brings to mind
the single verb cut); and (ii) greater retrieval demand, measured
as the weakness of the association strength between the noun
and the appropriate response (e.g., scissors is strongly associated
with cut and so has low retrieval demand, whereas giraffe is only
weakly associated with any verb). Speciﬁcally, reaction times
were longer in the high selection demand [log reaction time (RT)
mean (M) = 7.70, SE = 0.03; RT 2208 ms, SE = 68] than low
selection demand conditions (log RT M = 7.53, SE = 0.03; RT
1,863 ms, SE = 57) [F(1,82) = 215.9, P < 0.001], and longer in the high retrieval demand (log RT M = 7.78, SE = 0.04; RT 2,392 ms, SE = 98) than low retrieval demand (log RT M = 7.45, SE = 0.03; RT 1,720 ms, SE = 52) conditions [F(1,82) = 387.9, P < 0.001] (Fig. 1B and Table S1). In addition, we found an interaction between selection and retrieval demands: selection costs were greater under low retrieval demands [log RT difference (diff.) M = 0.20, SE = 0.02; RT diff. 352 ms] than under high retrieval demands (log RT diff. M = 0.13, SE = 0.02; RT diff. 324 ms) [F(1,82) = 12.1, P = 0.001] (Fig. 1C). Our model simulates and provides a framework for understanding these ﬁndings. The model uses a powerful framework that simulates the electrophysiological properties of neurons and can use networks of such neurons to simulate human behavior, including language and cognitive control (SI Methods 2.2–2.3 and Tables S2 and S3 provides details of the modeling framework, architecture, and simulations). The model contains layers (simulated brain areas) that simulate the following: (i) presentation of noun stimuli, (ii) activation of associated verb responses in the posterior cortex, (iii) selection of responses in the ventrolateral prefrontal cortex (VLPFC), and (iv) output of a response (Fig. 2A). The strength of connections between nouns and associated 16484 | www.pnas.org/cgi/doi/10.1073/pnas.1002291107 B Model Response Time (Cycles, +/- SE) Purr High Selection Demand VLPFC Input Selection Cost (Model Cycles Diff., +/- SE) A 140 30 25 20 15 10 5 0 Low Retrieval Demand High Retrieval Demand 120 100 80 60 40 20 0 High Low Selection Selection Demand Demand Low Retrieval Demand High Low Selection Selection Demand Demand High Retrieval Demand Fig. 2. Neural network model. (A) Network architecture. (B) Model simulates human performance, showing independent effects of selection demand (driven by competition between active representations) and retrieval demand (driven by synaptic weight strength). (C) Model simulates interaction between selection and retrieval (driven by beneﬁt of spreading activation when retrieval demands are high). All error bars are SEs. verb responses and between alternative verb responses were set according to the known association strengths observed in humans (24); these connections support spreading activation between related semantic representations like that observed in posterior cortex. Simulated neurons in the posterior cortex layer then activate verb representations in the VLPFC layer, which implements competitive lateral inhibition, selecting one response for output. Like people, the model takes longer to respond when retrieval or selection demands are high (SI Methods 2.3, SI Results 3.1, Table S3, Fig. 2B, and Figs. S1 and S2). The effects of retrieval demand are a direct consequence of the strength of the synaptic weights between a stimulus and its response representation in the posterior cortex layer (25); weaker weights cause a slower buildup of activation, requiring more time to reach the threshold for generating a response. Selection demand increases when multiple alternative responses become simultaneously active and competition must be resolved to select a single response. In the model, this resolution is accomplished through strong lateral inhibition in the VLPFC layer, simulating the effects of GABAergic interneurons. In addition, the model replicates the interaction between selection and retrieval demands, and provides insight into why such an interaction occurs (Fig. 2C). When responses are easily retrieved, activating multiple responses serves only to generate competition, imposing a large selection cost. However, when it is difﬁcult to retrieve any response, activating multiple responses aids retrieval, as spreading activation between these weakly associated alternatives (e.g., between hold and store when generating a response for shelf) boosts their activation levels. Thus, when retrieval demands are high, selection costs are partially offset by the advantage multiple responses confer on retrieval (SI Results 3.1 and Fig. S2). Manipulations of competitive inhibition in the VLPFC layer of the model enabled us to generate novel predictions regarding the effects of reduced GABAergic function associated with anxiety, Snyder et al. 2 Reduced Inhibition Standard Inhibition 1.5 1 0.5 0 Low Retrieval Demand Selection Cost (z RT Diff., +/- SE) B High Retrieval Demand High Anxiety Low Anxiety 0.5 0.4 0.3 0.2 0.1 0.0 Low Retrieval Demand High Retrieval Demand Fig. 3. Effects of reduced neural inhibition (A) Model predictions: Reduced competitive neural inhibition in the VLPFC layer, simulating increasing anxiety, impairs selection (i.e., increases selection cost) under high and low retrieval demands, and suggests that effects of anxiety on selection may be most robust under low retrieval demands. (B) Empirical results. Higher anxiety participants show impaired selection under high and low retrieval demands (model ﬁt is further discussed in Results). All error bars are SEs. Snyder et al. Simulation Selection Cost (Z Model Cycle Diff., +/- 1.4 Standard Inhibition 1.2 Increased Inhibition 1 0.8 0.6 0.4 0.2 0 Low Retrieval Demand Selection Cost (Z RT Diff., +/- SE) B High Retrieval Demand 0.6 0.5 Saline Midazolam 0.4 0.3 0.2 0.1 0 Low Retrieval Demand High Retrieval Demand Fig. 4. Effect of increased neural inhibition. (A) Model predictions. Increased neural inhibition (simulating GABA agonist drugs) improves selection (i.e., reduces selection cost) only when retrieval demands are low. (B) Empirical results. Midazolam improves selection only when retrieval demands are low. RTs are z transformed to remove baseline differences between conditions. All error bars are SEs. costs were equivalent across the groups [F(1,57) = 0.72, P = 0.4]. Also as predicted, left VLPFC activity correlated with anxiety during selection when retrieval demands were low (r = −0.663, P = 0.004, n = 17) (Fig. 5 and Fig. S4), but not during retrieval when selection demand was low (r = −0.181, P = 0.5, n = 16, Fisher’s z = −1.60, P = 0.05, one-tailed). (The brain activity thus conﬁrms the model’s prediction of larger effect of anxiety on selection when retrieval demands are low, whereas the behavioral data show similar effects of anxiety on selection under high and low retrieval demands, perhaps reﬂecting a lower sensitivity of RT measures or the role of other brain mechanisms or compensatory strategies.) These ﬁndings suggest that higher anxiety individuals lack sufﬁcient competitive dynamics in VLPFC for efﬁciently selecting between competing options (SI Discussion 1.2). We tested the predicted effects of increased neural inhibition in a double-blind, placebo-controlled study in which participants completed the verb generation task after injection of the GABA agonist midazolam as compared with a saline control in two counterbalanced sessions (Methods, SI Methods 2.6, and SI Discussion 1.3). Data were analyzed with a 2 × 2 × 2 repeated-measures ANOVA (Table S6). There was an interaction between drug condition, selection demand, and retrieval demand [F(1,19) = 5.67, P = 0.028]. As predicted, when retrieval demands were low, midazolam improved selection (with selection costs lower under midazolam, z-transformed RT diff. M = 0.15, SE = 0.06; RT diff. 267 ms, than under saline, z-transformed RT diff. M = 0.37, SE = 0.07; RT diff. 355 ms) [t(19) = −2.95, P = 0.008], whereas midazolam did not improve selection when retrieval demands were high [t(19) = 1.05, P = 0.3] (Fig. 4B, Fig. S5, and Table S7). Also as predicted, there was no effect of midazolam on retrieval when selection demands were low [t(19) = −0.53, P = 0.6]. Discussion In daily life, we often face a tyranny of choice (28), and this problem is ubiquitous when we choose words to express a thought. The current studies demonstrate that competitive neural inhibition, via GABAergic interneurons in prefrontal circuits, likely plays an important role in selecting among alternatives during PNAS | September 21, 2010 | vol. 107 | no. 38 | 16485 NEUROSCIENCE 2.5 A PSYCHOLOGICAL AND COGNITIVE SCIENCES A Simulation Selection Cost (Z Model Cycles Diff., +/- SE) and increased GABAergic function under GABA agonists, which we then tested empirically (SI Methods 2.4–2.6 and SI Results 3.2– 3.4). These simulations showed that competitive inhibition is critical for selection between competing alternatives, and that the effect of competitive inhibition on selection is modulated by retrieval demands. Decreasing competitive inhibition (as in anxiety) impairs selection (Fig. 3A), whereas increasing competitive inhibition (as under GABA agonists) improves selection (Fig. 4A). These effects of competitive inhibition on selection are more robust when retrieval demands are low. When retrieval demands are high, increased neural inhibition increases competitive dynamics that support selection, but also reduces spreading activation that aids retrieval, leading to weaker effects (Figs. 3A and 4A). In contrast, changes in competitive inhibition do not affect retrieval when selection demands are low (i.e., there is one associated response and thus no spreading activation). Thus, the simulations predict that reduced neural inhibition associated with anxiety will impair selection and associated VLPFC activity (26, 27) (SI Discussion 1.2), whereas increased neural inhibition under the GABA agonist midazolam will improve selection. These effects may be more apparent when retrieval demands are low. In addition, retrieval should not be affected by changes in neural inhibition when selection demands are low. These predictions were supported in three empirical investigations. The effects of anxiety on selection were investigated in separate behavioral and functional MRI (fMRI) studies in nonclinical populations that varied in levels of anxious apprehension, which should in turn inﬂuence the level of GABAergic activity. Anxious apprehension was assessed by standard questionnaires in which individuals rated how well statements such as “many situations make me worry” applied to them (Methods). Behavioral data were analyzed with a 2 × 2 × 2 mixed factorial ANOVA (Table S4). As predicted, participants higher in anxiety had larger selection costs (z-transformed RT diff. M = 0.37, SE = 0.03; RT diff. 421 ms) than lower-anxiety participants (z-transformed RT diff. M = 0.26, SE = 0.03; RT diff. 269 ms) [F(1,57) = 6.32 P = 0.015] (Fig. 3B, Fig. S3, and Table S5), but the retrieval A 0.35 Engagement of VLPFC During Selection (% Signal Change) B 0.3 0.25 0.2 0.15 0.1 0.05 0 16 26 36 46 56 -0.05 Anxiety Fig. 5. Engagement of left VLPFC during selection, as a function of anxiety. (A) Anatomically deﬁned region of interest (mid and anterior left VLPFC, shown in blue) was chosen based on prior work establishing its role in selection. Engagement of this region during selection was computed for each participant as the difference in fMRI activation between the high and low selection demand conditions. (B) Higher anxiety participants showed reduced engagement of left VLPFC during selection under low retrieval demands, which may reﬂect reduced activity of GABAergic interneurons (SI Discussion). Even for individuals without anxiety disorders, the difﬁculty of selecting between options has important real-world consequences in domains beyond language production. When people are faced with too many options, they may use suboptimal heuristics to reduce the number of alternatives (34), make a decision they regret (1, 28), or delay making a decision altogether (2, 29), often with negative consequences. For example, the more retirement plans that employees must choose among, the less likely they are to join any plan at all (2). Likewise, when physicians are asked to choose between two similar pain medications, they are less likely to prescribe either (3). In these complex domains, and in language production as well, selection is likely to depend on many processes in addition to neural inhibition, such as assigning values to different options (35), which are supported by additional brain areas. A complete model of selection will thus need to incorporate additional processes (e.g., generating potential response options and dynamically increasing control when there is response competition) supported by a larger network of brain areas [e.g., anterior cingulate cortex, presupplementary motor area (36, 37)], which may also be affected by anxiety. Therefore, an important goal for future research will be to investigate how these processes interact with competitive inhibition processes in ventrolateral prefrontal cortex to support selection in language processing, and whether these processes generalize across other prefrontal regions to support decision making in other domains (38, 39). Future work should also include imaging methods (e.g., SPECT) and modeling frameworks [e.g., detailed neurophysiological models (40)] that provide more direct measures and simulations of GABAergic activity. Whether in the grocery store or speaking a sentence, there is no escaping the necessity of selecting among competing alternatives. In language processing, and perhaps beyond, competitive neural inhibition is critical in helping us to cope with this tyranny of choice. Methods language processing. As predicted by neural network simulations, selection and associated prefrontal activity in a verb-generation task are impaired by anxiety (associated with reduced GABAergic function), whereas selection is improved under the drug midazolam (which increases GABAergic function). Of note, retrieval is unaffected by GABAergic function; instead, other mechanisms [e.g., sustained neuronal activation, enabled by recurrent connections in PFC networks (SI Discussion 1.1)] may support retrieval of weakly active representations. These ﬁndings shed light on why choosing among many options can be difﬁcult for anyone, and why it can be paralyzing for people with anxiety. In complex decision-making tasks, choice-overload is believed to increase when there are many options (1–3), the options are similar (29), or there is conﬂict between equally good options (30). Similarly, in language production, the difﬁculty of selecting among words has been described as a function of the number of alternatives (31), or similarity of activation levels across alternatives (5, 32). Each of these factors can be seen as increasing the amount of time necessary to resolve competition among options through neural inhibition, as each increases the degree to which multiple options are represented with equal strength. Although we are normally able to use cognitive control to overcome these selection difﬁculties, this process becomes more difﬁcult for persons with anxiety. Our modeling and empirical work suggest that the reduced GABAergic function associated with anxiety leads to impaired competitive neural inhibition and contributes to difﬁculty in selection. Although GABA agonists are widely used to treat the affective symptoms of anxiety disorders (33), we demonstrate that midazolam improves the cognitive process of selection in a nonclinical population, suggesting that GABA agonists may also be effective in treating the cognitive control and decision-making deﬁcits in anxiety disorders. 16486 | www.pnas.org/cgi/doi/10.1073/pnas.1002291107 All participants were from the University of Colorado and wider Boulder and Denver communities, spoke English as a ﬁrst language, and did not report any reading disorder. All participants gave informed consent and were treated in accordance with procedures approved by the University of Colorado institutional review board. Experiment 1: Basic Behavioral Effects and Effects of Anxiety. Participants Participants were 85 young adults (52 female and 33 male); a subset (n = 60) also completed anxiety and depression measures. An additional eight participants were excluded for not following task directions (n = 4), self-reported reading disorders (n = 2), and equipment failure (n = 2). In addition, two outliers with negative selection and/or retrieval effects were excluded from analysis, because the basic effects of the task manipulations are very robust, occurring for the vast majority of subjects, making it difﬁcult to interpret individual differences in cases in which there is a clear manipulation failure. With the inclusion of these subjects, all signiﬁcant effects remain signiﬁcant. Design, procedure, and analysis. Verb generation stimuli were 100 nouns in a 2 × 2 design: high and low retrieval demand (association strength between nouns and possible verb responses) crossed with high and low selection demand (degree of competition among alternative responses). Association strength and competition were calculated as in previous work (5), using latent semantic analysis (24). High and low association strength conditions were matched on competition, whereas high and low competition conditions were matched on association strength. The full stimulus set is available upon request. Participants were instructed to say the ﬁrst verb that came to mind when presented with a noun (e.g., “meow” or “feed” for “cat”), and were given an example and eight practice trials before completing the task. A ﬁxation-cross appeared for 500 ms, followed by a noun. Participants responded by speaking into a microphone that recorded voice-triggered reaction times (RTs), and advanced the computer to the next trial. Trial order was randomized for each participant. When the microphone was accidentally triggered (e.g., by a cough) or an error made (a nonverb), the trial was eliminated from analysis. The data were trimmed to remove trials with RTs less than 200 ms or greater than three SDs above each participant’s mean. For within-subject analyses, RTs were log transformed to normalize the data. For individual differences analyses, RTs were further z transformed to remove baseline differences. Snyder et al. Neural Network Model. We used a biologically plausible neural network modeling framework, Leabra (47, 48) (SI Methods), implemented in Emergent (http://grey.colorado.edu/emergent/). On each trial, one noun input unit was turned on, activating the relevant verb response units in the posterior cortex layer, which then became active in the VLPFC layer, which implemented strong k-winners-take-all (kWTA) lateral inhibition to select one verb and biased the posterior cortex layer representations toward this response. Multiple simulations were run with differing levels of kWTA lateral inhibition in the VLPFC layer. The level of inhibition was reduced to test the effects of decreased neural inhibition (that is likely associated with anxious apprehension) and increased to test the effect of increased neural inhibition (under midazolam) (SI Methods and SI Results). Each trial ended when the network settled, producing a response in the output layer. Thus, the VLPFC layer of this model implements mechanisms similar to the leaky, competing accumulator model of perceptual choice (49, 50), in which lateral inhibition between units allows one response to emerge as the winner, whereas decay of unit activations is counteracted by recurrent excitatory connections. Experiment 3: Effects of Increased Neural Inhibition under Midazolam. Participants. Participants were 24 young adults (10 female and 14 male). One ACKNOWLEDGMENTS. We thank members of the P50 center on Executive Function and Dysfunction including Wendy Heller, whose ideas have inﬂuenced the present study. In addition we thank Michael Frank, Peter Todd, Susan JungGrant, Rosi Kaiser, Ranjani Prabhakaran, and Mark Whisman for discussions, Paula Villar, Kirsten Orcutt, and Luka Ruzic for assistance with the fMRI study, and the staff of the University of Colorado Boulder Clinical and Translational Research Center for medical assistance with the midazolam study. This research was supported by grants from the National Institutes of Health: The midazolam study was supported by RO1-MH64812 and by Clinical Translational Research Center Grants (M01-RR0051 and UL1-RR025780), and all other research by P50MH079485 and 1F31-MH087073-01. 1. Iyengar SS, Lepper MR (2000) When choice is demotivating: Can one desire too much of a good thing? J Pers Soc Psychol 79:995–1006. 2. Sethi-Iyengar S, Huberman G, Jiang W (2004) Pension design and structure: New lessons from behavioral ﬁnance, eds Mitchell OS, Utkus S (Oxford University Press, Oxford), pp 83–95. 3. Redelmeier DA, Shaﬁr E (1995) Medical decision making in situations that offer multiple alternatives. JAMA 273:302–305. 4. Badre D, Wagner AD (2007) Left ventrolateral prefrontal cortex and the cognitive control of memory. Neuropsychologia 45:2883–2901. 5. Snyder HR, Munakata Y (2008) So many options, so little time: The roles of association and competition in underdetermined responding. Psychon Bull Rev 15:1083–1088. 6. Thompson-Schill SL, D’Esposito M, Aguirre GK, Farah MJ (1997) Role of left inferior prefrontal cortex in retrieval of semantic knowledge: A reevaluation. Proc Natl Acad Sci USA 94:14792–14797. 7. Sachdev PS, Malhi GS (2005) Obsessive-compulsive behaviour: A disorder of decisionmaking. Aust N Z J Psychiatry 39:757–763. 8. Abramowitz JS, Foa EB (1998) Worries and obsessions in individuals with obsessivecompulsive disorder with and without comorbid generalized anxiety disorder. Behav Res Ther 36:695–700. 9. Starcevic V, Berle D (2006) Cognitive speciﬁcity of anxiety disorders: A review of selected key constructs. Depress Anxiety 23:51–61. 10. Barch DM, Braver TS, Sabb FW, Noll DC (2000) Anterior cingulate and the monitoriing of response conﬂict: Evidence from an fMRI study of overt verb generation. J Cogn Neurosci 12:298–309. 11. Herd SA, Banich MT, O’Reilly RC (2006) Neural mechanisms of cognitive control: An integrative model of stroop task performance and FMRI data. J Cogn Neurosci 18: 22–32. 12. Phillips WA, Silverstein SM (2003) Convergence of biological and psychological perspectives on cognitive coordination in schizophrenia. Behav Brain Sci 26:65–82, discussion 82–137. 13. Bagary M, et al. (2000) Is benzodiazepine-induced amnesia due to deactivation of the left prefrontal cortex? Psychopharmacology (Berl) 150:292–299. 14. Jansen JF, et al. (2006) Functional MRI reveals declined prefrontal cortex activation in patients with epilepsy on topiramate therapy. Epilepsy Behav 9:181–185. 15. Kan IP, Kable JW, Van Scoyoc A, Chatterjee A, Thompson-Schill SL (2006) Fractionating the left frontal response to tools: Dissociable effects of motor experience and lexical competition. J Cogn Neurosci 18:267–277. 16. Kan IP, Thompson-Schill SL (2004) Effect of name agreement on prefrontal activity during overt and covert picture naming. Cogn Affect Behav Neurosci 4:43–57. 17. Persson J, et al. (2004) Selection requirements during verb generation: Differential recruitment in older and younger adults. Neuroimage 23:1382–1390. 18. Thompson-Schill SL, et al. (1998) Verb generation in patients with focal frontal lesions: A neuropsychological test of neuroimaging ﬁndings. Proc Natl Acad Sci USA 95:15855–15860. 19. Engels AS, et al. (2007) Speciﬁcity of regional brain activity in anxiety types during emotion processing. Psychophysiology 44:352–363. 20. Sen S, et al. (2004) Serotonin transporter and GABAA alpha 6 receptor variants are associated with neuroticism. Biol Psychiatry 55:244–249. 21. Smoller JW, et al. (2001) Genetic association analysis of behavioral inhibition using candidate loci from mouse models. Am J Med Genet 105:226–235. 22. Zai G, et al. (2005) Evidence for the gamma-amino-butyric acid type B receptor 1 (GABBR1) gene as a susceptibility factor in obsessive-compulsive disorder. Am J Med Genet B Neuropsychiatr Genet 134B:25–29. 23. Kalueff AV, Nutt DJ (2007) Role of GABA in anxiety and depression. Depress Anxiety 24:495–517. 24. Landauer TK, Dumais ST (1997) A solution to Plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychol Rev 104:211–240. 25. Wagner AD, Paré-Blagoev EJ, Clark J, Poldrack RA (2001) Recovering meaning: Left prefrontal cortex guides controlled semantic retrieval. Neuron 31:329–338. 26. Buzsáki G, Kaila K, Raichle M (2007) Inhibition and brain work. Neuron 56:771–783. 27. Logothetis NK (2008) What we can do and what we cannot do with fMRI. Nature 453: 869–878. 28. Schwartz B (2004) The tyranny of choice. Sci Am 290:70–75. Snyder et al. PNAS | September 21, 2010 | vol. 107 | no. 38 | 16487 NEUROSCIENCE Experiment 2: Effects of Anxiety on VLPFC Activity during Selection. Experiment 2 was part of a larger fMRI study focused on prefrontal organization. Here we focus on the critical test of our theory, the effect of anxiety on VLPFC activity. Participants. Participants were 18 right-handed young adults (9 female and 9 male). Three additional participants were excluded because of excessive (>25%) error rates.
Design, procedure, and analysis. The verb generation task was identical to experiment 1, except that, to adapt the procedure for fMRI scanning, trail length
was ﬁxed to 4,000 ms rather than being self-paced. Verbal responses were
recorded with a ﬁberoptic microphone (Optoacoustics Ltd.). A rapid eventrelated paradigm and standard blood oxygen level-dependent (BOLD) imaging techniques were used to collect functional data in one run lasting ∼9
min (SI Methods). Participants completed the PSWQ and MASQ outside of the
scanner (refer to experiment 1).
Nonverb errors were removed from analysis, and instead included as
a nuisance covariate in the GLMs. (RTs are not available because the scanner
environment was too noisy to accurately determine voice-onset RTs.) First,
GLMs were conducted estimating signal change for contrasts of high vs. low
selection demand and high vs. low retrieval demand conditions (SI Methods).
Next, an anatomical region of interest (ROI) consisting of left inferior frontal
gyrus pars triangularis and pars orbitalis (mid and anterior VLPFC) was deﬁned using the Harvard–Oxford Cortical Structures (http://www.fmrib.ox.ac.
uk/fsl/data/atlas-descriptions.html) and Duvernoy atlases (46), and signal
change in each contrast within this ROI was extracted for each participant and
correlated with anxiety and depression scores. Outlier analyses were conducted (Cook’s D > 3 SD above mean), resulting in the exclusion of no more
than two participants from each correlation analysis.
additional participant was excluded for not completing the second session. In
addition, four outliers with negative selection and/or retrieval effects under
saline were excluded from analysis, as in experiment 1, because it is difﬁcult to
interpret drug effects in cases in which there is a clear manipulation failure.
Including these subjects does not change the overall pattern of results; the
signiﬁcant three-way interaction becomes marginally signiﬁcant (P = 0.06),
and all other signiﬁcant effects remain signiﬁcant.
Design, procedure, and analysis. A double-blind, placebo-controlled design was
used. Participants completed counterbalanced sessions, 1 wk apart, under
midazolam (0.03 mg/kg body weight diluted to a total volume of 10 mL) and
placebo (10 mL saline solution) (SI Methods). The verb generation task was
identical to that in experiment 1, except that the stimulus set was divided at
random into two lists (with an additional 50 ﬁller items), counterbalanced
across sessions and drug conditions. RT data were trimmed and transformed
as in experiment 1.
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
Participants completed four standardized questionnaires to assess anxious
apprehension: (i) NEO Five Factor Inventory (NEO-FFI) neuroticism subscale
(41), (ii) Lehrer Woolfolk Anxiety Symptom Questionnaire cognitive factor
(LWASQ) (42), (iii) Penn State Worry Questionnaire (PSWQ) (43), and (iv)
Behavioral Inhibition Scale/Behavioral Activation Scale (BIS/BAS) (44) Behavioral Inhibition subscale. These questionnaires were combined into
a summary score (SI Methods). Participants were classiﬁed as characterized
by high or low anxious apprehension using a median split (SI Results
describes converging results from a continuous analysis). In addition, participants completed the Mood and Anxiety Symptom Questionnaire (MASQ)
(45) to control for depression and anxious arousal symptoms (SI Discussion
and SI Results).
29. Fasolo B, Hertwig R, Huber M, Ludwig M (2009) Size, entropy, and density: What is
the difference that makes the difference between small and large real-world
assortments? Psychol Mark 26:254–279.
30. Tversky A, Shaﬁr E (1992) Choice under conﬂict: The dynamics of deferred decision.
Psychol Sci 3:358–361.
31. Desmond JE, Gabrieli JD, Glover GH (1998) Dissociation of frontal and cerebellar
activity in a cognitive task: Evidence for a distinction between selection and search.
Neuroimage 7:368–376.
32. Thompson-Schill SL, Botvinick MM (2006) Resolving conﬂict: A response to Martin and
Cheng (2006). Psychon Bull Rev 13:402–408, discussion 409–411.
33. Mula M, Pini S, Cassano GB (2007) The role of anticonvulsant drugs in anxiety
disorders: A critical review of the evidence. J Clin Psychopharmacol 27:263–272.
34. Tversky A, Kahneman D (1974) Judgment under uncertainty: Heuristics and biases.
Science 185:1124–1131.
35. Hare TA, Camerer CF,, Rangel A (2009) Self-control in decision-making involves
modulation of vmPFC valuation system. Science 324:646–648.
36. Botvinick MM, Braver TS, Barch DM, Carter CS, Cohen JD (2001) Conﬂict monitoring
and cognitive control. Psychol Rev 108:624–652.
37. Lau HC, Rogers RD, Haggard P, Passingham RE (2004) Attention to intention. Science
303:1208–1210.
38. O’Reilly RC (2010) The What and How of prefrontal cortical organization. Trends
Neurosci 33:355–361.
39. Sumner P, Edden RAE, Bompas A, Evans CJ, Singh KD (2010) More GABA, less distraction:
A neurochemical predictor of motor decision speed. Nat Neurosci 13:825–827.
16488 | www.pnas.org/cgi/doi/10.1073/pnas.1002291107
40. Deco G, Rolls ET, Horwitz B (2004) Integrating fMRI and single-cell data of visual
working memory. Neurocomputing 58:729–737.
41. Costa PT, McCrae RR (1992) Revised NEO Personality Inventory (NEO-PI-R) and NEO
Five-Factor Inventory (NEO-FFI) professional manual (Psychological Assessment
Resources, Odessa, FL).
42. Lehrer PM, Woolfolk RL (1982) Self-report assessment of anxiety: Semantic, cognitive
and behavioral modalities. Behav Assess 4:167–177.
43. Meyer TJ, Miller ML, Metzger RL, Borkovec TD (1990) Development and validation of
the Penn State Worry Questionnaire. Behav Res Ther 28:487–495.
44. Carver CS, White TL (1994) Behavioral inhibition, behavioral activation, and affective
responses to impending reward and punishment. J Pers Soc Psychol 67:319–333.
45. Watson D, et al. (1995) Testing a tripartite model: I. Evaluating the convergent and
discriminant validity of anxiety and depression symptom scales. J Abnorm Psychol 104:
3–14.
46. Duvernoy HM (1999) The Human Brain: Surface, Blood Supply, and ThreeDimensional Sectional Anatomy (Springer-Verlag Wien, New York).
47. O’Reilly RC (1998) Six principles for biologically based computational models of
cortical cognition. Trends Cogn Sci 2:455–462.
48. O’Reilly RC, Munakata Y (2000) Computational explorations in cognitive neuroscience:
Understanding the mind by simulating the brain (MIT Press, Cambridge, MA).
49. Bogacz R, Usher M, Zhang J, McClelland JL (2007) Extending a biologically inspired
model of choice: Multi-alternatives, nonlinearity and value-based multidimensional
choice. Philos Trans R Soc Lond B Biol Sci 362:1655–1670.
50. Usher M, McClelland JL (2001) The time course of perceptual choice: The leaky,
competing accumulator model. Psychol Rev 108:550–592.
Snyder et al.
Neuron
Perspective
Neuroscience Needs Behavior:
Correcting a Reductionist Bias
John W. Krakauer,1,* Asif A. Ghazanfar,2 Alex Gomez-Marin,3 Malcolm A. MacIver,4 and David Poeppel5,6
1Departments of Neurology, and Neuroscience, Johns Hopkins University, Baltimore, MD 21287, USA
2Princeton Neuroscience Institute, Departments of Psychology and Ecology & Evolutionary Biology, Princeton University, Princeton,
NJ 08540 USA
3Instituto de Neurociencias, Consejo Superior de Investigaciones Cientı́ficas & Universidad Miguel Hernández, Sant Joan d’Alacant,
03550 Alicante, Spain
4Neuroscience and Robotics Laboratory, Department of Neurobiology, Department of Mechanical Engineering, Northwestern University,
Evanston, IL 60208, USA
5Department of Psychology, New York University, New York, NY 10003, USA
6Neuroscience Department, Max-Planck Institute for Empirical Aesthetics, 60322 Frankfurt, Germany
*Correspondence: jkrakau1@jhmi.edu
http://dx.doi.org/10.1016/j.neuron.2016.12.041
There are ever more compelling tools available for neuroscience research, ranging from selective genetic targeting to optogenetic circuit control to mapping whole connectomes. These approaches are coupled with a
deep-seated, often tacit, belief in the reductionist program for understanding the link between the brain and
behavior. The aim of this program is causal explanation through neural manipulations that allow testing of
necessity and sufficiency claims. We argue, however, that another equally important approach seeks an
alternative form of understanding through careful theoretical and experimental decomposition of behavior.
Specifically, the detailed analysis of tasks and of the behavior they elicit is best suited for discovering component processes and their underlying algorithms. In most cases, we argue that study of the neural implementation of behavior is best investigated after such behavioral work. Thus, we advocate a more pluralistic notion
of neuroscience when it comes to the brain-behavior relationship: behavioral work provides understanding,
whereas neural interventions test causality.
Introduction
Advances in technology have allowed the study of neurons,
including their component parts and molecular machinery, to
an unprecedented degree. This work promises to yield much
new information about brain structure and physiology independent of behavior—for example, the biophysics of receptors
or details of spatial summation in dendrites. In addition, new
methods such as optogenetics allow some causal relationships between brain and behavior to be determined. Here we
will argue, however, that detailed examination of brain parts
or their selective perturbation is not sufficient to understand
how the brain generates behavior (Figure 1). One reason is
that we have no prior knowledge of what the relevant level of
brain organization is for any given behavior (Figure 1A).
When this concern is coupled with the brain’s deep degeneracy, it becomes apparent that the causal manipulation
approach is not sufficient for gaining a full understanding of
the brain’s role in behavior (Marom et al., 2009). The same
behavior may result from alternative circuit configurations
(Marder and Goaillard, 2006), from different circuits altogether
or the same circuit may generate different behaviors (Katz,
2016) (Figures 1D and 1E). This concern has been voiced
before in a variety of ways (Anderson, 1972; Marr, 1982/
2010; Oatley, 1978), but we think that it is useful to revisit
and reframe the arguments at a time of understandable excitement about ever more effective interventionist approaches in
neuroscience.
480 Neuron 93, February 8, 2017 ª 2016 Elsevier Inc.
An analogy from computer science that has both historical and
conceptual appeal is the distinction between software and hardware; whereby the software represents ‘‘what’’ the brain (or one
of its modules) is doing, and the hardware represents ‘‘how’’ it is
doing it. Sternberg has stated it as the ‘‘distinction between processors and the processes that they implement’’ (Sternberg,
2011, p. 158). The core question we address here is whether
the processes governing behavior are best inferred from examination of the processors. In a nice irony, the computer science
analogy has come full circle with a provocative study that applied
numerous neuroscience techniques to a single microprocessor
(analogous to a brain) in an attempt to understand how it controls
three classic videogames (analogous to behaviors) (Jonas and
Kording, 2017). Crucial to the experiment was that the answer
was known a priori: the processor’s operations can be drawn
as an algorithmic flow chart. The sobering result was that performing interventionist neuroscience on the processor could
not explain how the processor worked. We have more to say
about this study later.
Neuroscience is replete with cases that illustrate the fundamental epistemological difficulty of deriving processes from
processors. For example, in the case of the roundworm (Caenorhabditis elegans), we know the genome, the cell types, and
the connectome—every cell and its connections (Bargmann,
1998; White et al., 1986). Despite this wealth of knowledge,
our understanding of how all this structure maps onto the
worm’s behavior remains frustratingly incomplete. Thus, it is
Neuron
Perspective
Figure 1. The Multiple Potential Mappings
between Neural Activity Patterns and
Natural Behaviors
C
Neural activity patterns
Natural behaviors
Unnatural
activity pattern
Behavior outside
natural repertoire
Multiple possible
patterns of activity
Single natural
behavior
Single
pattern of activity
Multiple natural
behaviors
D
A
Subset of
possible activity patterns
All natural behaviors
B
E
Smaller subset of
activity patterns
Subset of
natural behaviors
readily apparent that it is very hard to infer the mapping between
the behavior of a system and its lower-level properties by only
looking at the lower-level properties (Badre et al., 2015; Carandini, 2012; Cooper and Peebles, 2015; Gomez-Marin et al.,
2014). When we ask, ‘‘How does the brain generate behavior,’’
we are primarily asking about how putative processing modules
are organized so that they combine to generate behavior in a
particular task environment. Relying solely on the collection of
neural data, with behavior incorporated as an after-thought
(and typically over-constrained; Box 1; Figure 1C), will not
lead to meaningful answers. This is a question best answered
through precise hypotheses articulated in an a priori conceptual
framework, careful task design, and the collection of behavioral data.
How Did We Get Here?
New technologies have enabled the acquisition of massive and
intricate datasets, and the means to analyze them have become
concomitantly more complex. This in turn has led to a need for
experts in computation and data analysis, with a reduced
emphasis on organismal-level thinkers who develop detailed
functional analyses of behavior, its developmental trajectory,
and its evolutionary basis. Deep and thorny questions like
‘‘what would even count as an explanation in this context,’’
‘‘what is a mechanism for the behavior we are trying to understand,’’ and ‘‘what does it mean to understand the brain’’ get
sidelined. The emphasis in neuroscience has transitioned from
these larger scope questions to the development of technologies, model systems, and the approaches needed to analyze
the deluge of data they produce. Technique-driven neuroscience could be considered an example of what is known as the
substitution bias: ‘‘[.] when faced with a difficult question, we
often answer an easier one instead, usually without noticing
the substitution’’ (Kahneman, 2011, p. 12).
(A) Of all the possible activity patterns of a brain in
a dish (big pale blue circle), only a subset of
these (medium dark blue circle) will be relevant in
behaving animals in their natural environment (big
magenta circle).
(B) Designing behavioral tasks that are ecologically
valid (small magenta circle) ensures discovery of
neural circuits relevant to the naturalistic behavior
(small blue circle). Tasks that elicit species-typical
behaviors with species-typical signals are examples (see Box 1).
(C) In order to study animal behavior in the lab,
the task studied (small white circle) might be so
non-ecological it elicits neural responses (small
blue circle) that are never used in natural
behaviors.
(D) Multiple Realizability: different patterns of activity or circuit configurations (small blue circles)
can lead to the same behavior (small magenta
circle).
(E) The same neural activity pattern (small blue
circle) can be used in two different behaviors (two
magenta circles). The circle with dashed perimeter in (B)–(E) is the subset of all possible neural
activity patterns that map onto natural behaviors
(from A).
In an interesting historical parallel to the argument we make
here, the historian of science Lily Kay described how the discipline of molecular biology also arose from placing a premium
on technology and its application to simple model systems
(Kay, 1996). She quotes with concern Monod’s line, ‘‘What
is true for the bacterium is true for the elephant’’ (Kay, 1996,
p. 5). Here we caution similarly against the idea that what is
true for the circuit is true for the behavior. Monod’s line has
echoed through to the present day with the argument that molecular biology and its techniques should serve as the model
for understanding in neuroscience (Bickle, 2016). We disagree
with this totalizing reductionist view but take it as evidence that
excessive faith in molecular and cellular biology may be partially
to blame for the current dominance of interventionist explanations in neuroscience. We fully acknowledge the crucial role
that technology plays in advancing biological knowledge and
the value of interventionist approaches, but this tool-driven trend
is not sufficient for understanding the brain-behavior relationship. Neural data obtained from new methods cannot substitute
for developing new conceptual frameworks that provide the
mapping between such neural data and behavior in an algorithmic sense (and not just a correlative or even causal way). Accomplishing this task requires hypotheses and theories based
on careful dissection of behavior into its component parts or subroutines (Cooper and Peebles, 2015). The behavioral work needs
to be as fine-grained as work at the neural level. Otherwise one is
imperiled by a granularity mismatch between levels that prevents
substantive alignment between different levels of description
(Poeppel and Embick, 2005).
The first step for developing conceptual frameworks that
meaningfully relate neural circuits to behavioral predictions is
to design hypothesis-based behavioral experiments. Despite
this pure behavioral step being of critical importance and highly
informative in its own right, it has increasingly been marginalized
Neuron 93, February 8, 2017 481
Neuron
Perspective
Box 1. Behavior
Tinbergen defined behavior as ‘‘The total movements made by the intact animal’’ (Tinbergen, 1955). Authors of a recent survey
designed to investigate how working scientists define behavior came up with the following attempt at an updated definition,
‘‘Behavior is the internally coordinated responses (actions or inactions) of whole living organisms (individuals or groups) to internal
and/or external stimuli, excluding responses more easily understood as developmental changes’’ (Levitis et al., 2009). The core of
this definition is that behavior is the ‘‘internally coordinated responses . to internal and/or external stimuli.’’ Clearly, however,
some stimuli are more important than others in furthering our understanding of animals and their nervous systems.
Unfortunately, animals do not raise a checkered flag to indicate when they are about to perform a behavior, or signal when it ends.
They are constantly in motion and responding to whatever is around them, which invites the following easy mistake: since an animal is responding to stimuli, and physiological correlates are measurable, one is therefore studying an animal’s behavior (see the
‘‘overly constrained behavior’’ of Figure 1C). If, following the lead of 20th century ethologists, we treat behavior as no less an
evolved entity than is, say, the shape of the humerus (Tinbergen, 1963), then correctly labeling something as behavior is contingent
on the outcome of an investigation into what the animal does to ensure its survival in its native habitat. In this way it would be
discovered, for example, that bats navigate through dense forests in total darkness while hunting insects and that rodents beat
a hasty retreat when they see a hawk diving towards them, while not responding to similarly sized birds flying straight over. It is
therefore a significant confusion to label a coordinated response to a stimulus a ‘‘behavior’’ without first determining the relevance
of the response to the animal’s natural life (Tinbergen’s second question, Figure 4). Not doing so is to conflate a ‘‘stimulus
response’’ with a spatio-temporal pattern that is the product of selection over time. While any perturbation applied to an animal
can lead to productive lines of inquiry, whether or not it is founded in anything ethological, the history of many of the most produc€ ll, 1992) in such a way as
tive moments in neuroscience is a history of having ingeniously abstracted an animal’s Umwelt (Von Uexku
to admit of controlled, repeatable experiments.
In light of the preceding, placing a behaving animal in a situation where it perceives sensory events that are behaviorally relevant,
and can act on them in approximately the same way as if they were embedded in the world, can be enormously useful (Figure 1B).
For example, engaging a subject with the real or simulated presence of another can capture behavioral principles that are common
across species. Doing so with marmoset monkeys revealed vocal turn-taking behavior with similar patterns of phase-locking and
entrainment as in human communication (Takahashi et al., 2013). Eschewing the purely big data approach where behavioral data
are acquired blindly from large numbers of animals through automation and without regard for the individual (Anderson and Perona, 2014), this organismal-level study led to insights into both developmental (Takahashi et al., 2015) and evolutionary (Borjon and
Ghazanfar, 2014) processes, and, subsequently, to computational principles shared across species (Takahashi et al., 2012).
Similar ethological approaches in other species have led to a number of behaviorally driven investigations of neural level mechanisms: fish (Bass and Chagnaud, 2012), frogs (Leininger and Kelley, 2015), and birds (Benichov et al., 2016).
Ultimately, the most effective approach may be to simulate the entire natural task environment in order to elicit the full range of
adaptive behavioral possibilities. Virtual reality (VR) systems developed for a host of model systems, including rodents, flies,
and fish, offer such an approach (Dombeck and Reiser, 2012). VR systems were originally an answer to the problem of how to
mediate the uneasy marriage of the tightly controlled but non-ethological world of laboratory physiology, and the poorly controlled
but ethologically relevant world of the behaviors studied by ethologists (MacIver, 2009). It is critically important to realize, however,
that effective use of VR requires a fine-grained quantitative understanding of the behavior under study as it occurs in the unimpeded animal. Only then can the investigator assess, for example, whether the VR system is in fact able to trick the animal into
believing it is in the world as it would normally operate within. For example, a VR system has been developed by which we can
reliably elicit prey capture behavior in larval zebrafish (which hunt Paramecia) (Bianco et al., 2011). Careful behavioral work on
the free-swimming animal showed that prey detection is marked by the fish’s eyes verging together to point to the prey (Bianco
et al., 2011; Patterson et al., 2013). This eye movement is not seen during any other behaviors. With that knowledge in hand,
we can assess the success of a VR system by how often we can see the eyes move in this manner when we display artificial Paramecia on a small screen in front of the animal, which is otherwise fixed in place by being embedded in a block of agar.
Understanding behavior and its component processes at the level of detail necessary to generate meaningful neural level insights
will require an emphasis on natural behaviors performed by individuals. Although there are new technologies enabling the blind
acquisition of massive behavioral datasets and the application of machine learning algorithms (Anderson and Perona, 2014),
they will not lead to the detailed functional analyses of ethological behavior, its developmental trajectory, and its evolutionary basis
that are necessary for appropriately constrained implementation-level theories.
or at best postponed (Anderson and Perona, 2014). It is disturbingly common for studies to include behavior as simply a hasty
add-on in papers that are otherwise crammed full of multiple
techniques, types of results, and even species. It is as if every
paper needs to be a methodological decathlon in order to be
considered important. Behavior must be seen as something
482 Neuron 93, February 8, 2017
that can stand alone as a foundational phenomenon in its own
right (Box 1).
Why We Still Need Behavior
Perhaps the best-known example of a framework devised to
formalize what it means to understand the link between brain
Neuron
Perspective
A
Figure 2. Marr’s Three Levels of Analysis
B
LEVELS
Computation
1
why (problem)
1 suggests 2
Algorithm
2
what (rules)
2 explains 1
Implementation
3
how (physical)
1
2 predicts 3
2
3
second step
C
epistemological
bias
feathers
1
2
understanding
and behavior is David Marr’s levels of understanding of complex
systems, originally fashioned as a critique of reductionist work in
neurophysiology (Marr, 1982/2010). We do not intend to rehash
or dissect Marr’s particular formulation, as this has been done
many times before, and sometimes arguments over the definitions of his three levels—or whether there should more than
three—obscure the central and deep point made by Marr (and
many others before and after [Anderson, 1972; Bechtel, 2008;
Carandini, 2012; Cooper and Peebles, 2015; Oatley, 1978; Tinbergen, 1963; Badre et al., 2015; Fetsch, 2016; Frank and Badre,
2015; Schall, 2004]): understanding something is not the same
as just describing it or knowing how to intervene to change it.
To most it is not news that description is not understanding,
but too often in neuroscience causal efficacy is taken as equal
to understanding.
Marr took a strong position on the inadequacy of a strictly
neurophysiological approach to understanding: ‘‘.trying to
understand perception by understanding neurons is like trying
to understand a bird’s flight by studying only feathers. It just
cannot be done’’ (Marr, 1982/2010) (Figure 2). Marr’s main
intuition was that it is much more difficult to infer from the neural hardware (or implementation; level 3) what algorithm (level
2) the nervous system is employing as compared to getting to
it via an analysis of the computational problem (level 1) it is
trying to solve. Marr’s main objection to trying to understand
the brain by recording from neurons was that this only leads
to descriptions rather than explanations. A description of neural activity and connections is not synonymous with knowing
what they are doing to cause behavior. Even strong believers
in the work done at the level of neurons and molecules (the implementation level) concede Marr’s point (Bickle, 2015). An
analogy that helps get this point across is understanding of
the game of chess. Understanding the game does not depend
on knowing anything about the material out of which the board
or chess pieces are made. Indeed, Marr suggested that the
details of the nervous system may not even matter. While it
is true that the physical properties of the chess pieces can
impinge on application of the rules—for example, if one inadvertently gives a child a chess set for which all the pieces are
too heavy for her to pick up. The ‘‘therapeutic’’ solution is lighter chess pieces, but this in no way has changed the analysis
3
manipulation
(A) A bird attempts to fly (goal) by flapping
its wings (algorithmic realization) whose aerodynamics depend on the features of its
feathers (physical implementation). Feathers
‘‘have something to do’’ with flight and flapping,
but what level of understanding do we achieve
if we dissect the properties of the feathers
alone? Bats fly but don’t have feathers, and
birds can fly without flapping.
(B) The relationship between the three levels is
not arbitrary; step 1 comes before step 2: the
algorithmic level of understanding is essential to
interpret its mechanistic implementation. Step 2:
implementation level work feeds back to inform
the algorithmic level.
(C) An epistemological bias toward manipulationbased view of understanding induced by technology (black filled arrow).
or understanding of chess. This chess analogy serves to make
an important point: well-designed behavioral experiments
in the absence of work at the neural level can be highly informative on their own. Behavioral experiments often are a
necessary first step before a subsequent mutually beneficial
knowledge loop is set up between implementation and behavioral level work.
A more concrete example of the problems that arise if neural
data are used to infer a psychological process comes from the
debates regarding the behavioral relevance of ‘‘mirror neurons.’’
Mirror neurons, first discovered in the premotor cortex of monkeys, fire whether the monkey itself performs a particular motor
goal or observes another individual doing so (di Pellegrino et al.,
1992). A huge number of variants of these experiments have
been done in both humans and monkeys, but they all have the
same general approach: show a common neuronal firing (or
fMRI or EEG/MEG activation pattern) when a goal is achieved
either in the first person or observed in the third person. Interpretation then has the following logic: as neurons can be decoded
for intention in the first person, and these same neurons decoded
for the same intention in the third person, then activation of the
mirror neurons can be interpreted as meaning that the primate
has understood the intention of the primate it is watching. The
problem with this attempt to posit an algorithm for ‘‘understanding’’ based on neuronal responses is that no independent
behavioral experiment is done to show evidence that any kind
of understanding is actually occurring, understanding that could
then be correlated with the mirror neurons. This is a key error in
our view: behavior is used to drive neuronal activity but no either/
or behavioral hypothesis is being tested per se. Thus, an interpretation is being mistaken for a result; namely, that the mirror
neurons understand the other individual. Additional behavioral
evidence that the participant understands the other individual
is lacking. This tendency to ascribe psychological properties to
single neuron activity that can only be sensibly ascribed to a
whole behaving organism is known as the mereological fallacy—a fallacy that we neuroscientists continue to fall for even
though we’ve known about it since Aristotle’s De Anima (Smit
and Hacker, 2014). Thus, what is needed is a better a priori testable framework for behavioral-level understanding that can lead
to more thoughtfully designed neurophysiological experiments.
Neuron 93, February 8, 2017 483
Neuron
Perspective
Indeed, to the degree that action understanding has been examined in patients, the evidence does not support a role for the putative mirror neuron mechanism (Hickok, 2009).
A recent review exemplifies the neuroscience zeitgeist by stating that the time has come to go from considering individual neurons as the functional units of the nervous system to ensembles
of neurons (circuits, networks) (Yuste, 2015). The main argument
is that ensembles generate states that would never be appreciated by recording one neuron at a time. The claim is then
made that, with the application of new technologies (e.g., twophoton imaging, multielectrode recordings, etc.), identification
of these neural states will help us better understand the link between the brain and behavior (although again behavior is at best
given backseat status) (Yuste, 2015). No overt new theory, however, is offered that will bridge ensemble activity and behavior. It
is therefore unclear how fundamentally different it really is,
conceptually, to move from ‘‘neuron’’ to ‘‘neurons.’’ Modeling
and studying the responses of the neural substrate on any
scale—large or small—will not by itself lead to insights about
how behavior is generated. One reason for this is that the properties of neural tissue may be more diverse than the subset actually exploited for natural behaviors (Figures 1A and 1D).
Without well-characterized behavior and theories that can act
as a constraint on circuit-level inferences, brains and behavior
will be like two ships passing in the night. The field has been
here before. Concerns with the complete-circuit-description
approach were already recognized almost 40 years ago with
the publication—along with extensive accompanying commentary—of an article titled ‘‘Are Central Pattern Generators Understandable?’’ (Selverston, 1980). A plea, similar to those we hear
today, was made for ever more detailed characterization of each
element in the circuit and specification of the synaptic connectivity between these elements. Many of the commentaries pointed
out, however, that increasingly complete descriptions at one
level do not serve as a bridge to the next level. For example,
Sten Grillner wrote that a central pattern generator may be understood better using the intentional stance, borrowing from
the philosopher Daniel Dennett (Dennett, 1989), which is the
view that when an entity is designed for a purpose it is therefore
subject to rational rules that can be determined by studying its
behavior without necessarily having to analyze all its physical
parts (comment by Grillner in Selverston, 1980).
The phenomenon at issue here, when making a case for
recording from populations of neurons or characterizing whole
networks, is emergence—neurons in their aggregate organization cause effects that are not apparent in any single neuron.
Following this logic, however, leads to the conclusion that
behavior itself is emergent from aggregated neural circuits and
therefore should also be studied in its own right. An example of
an emergent behavior that can only be understood at the algorithmic level, which in turn can only be determined by studying
the emergent behavior itself, is flocking in birds. First one has
to observe the behavior and then one can begin to test simple
rules that will lead to reproduction of the behavior, in this
case best done through simulation. The rules are simple—for
example, one of them is ‘‘steer to average heading of neighbors’’
(Reynolds, 1987). Clearly, observing or dissecting an individual
bird, or even several birds could never derive such a rule. Substi-
484 Neuron 93, February 8, 2017
tute flocking with a behavior like reaching, and birds for neurons,
and it becomes clear how adopting an overly reductionist
approach can hinder understanding.
How has neuroscience dealt with this persistent gap between
explanation and description? It has opted to favor interventionist
causal versions of explanation. Unfortunately, there are no shortcuts in the trajectory from psychology, cognition, perception,
and behavior to neurons and circuits. One might argue that techniques now exist that make it possible to manipulate neural
circuits directly, for example, with optogenetics or transcranial
magnetic stimulation, so that causal relations—and not just
correlations—can be discovered (Bickle, 2015). The critical
point, however, is that causal-mechanistic explanations are
qualitatively different from understanding how component modules perform the computations that then combine to produce
behavior.
The distinction between causal claims and understanding via
algorithmic or computational processes should be apparent
from argument alone. That said, the recent study by Jonas and
Kording (Jonas and Kording, 2017) we referred to earlier provides an empirical demonstration of the fundamental difference
between intervening and recording versus understanding how
information flows through processing steps. The study poses
the question of whether a neuroscientist could understand a
microprocessor. They applied numerous neuroscience techniques to a high-fidelity simulation of a classic video game microprocessor (the ‘‘brain’’) in an attempt to understand how it
controls the initiation of three well-known videogames (which
they dubbed as ‘‘behaviors’’) originally programmed to run on
that microprocessor. Crucial to the experiment was the fact
that it was performed on an object that is already fully understood: the fundamental fetch-decode-execute structure of a
microprocessor can be drawn in a diagram. Understanding the
chip using neuroscientific techniques would therefore mean
being able to discover this diagram. In the study, (simulated)
transistors were lesioned, their tuning determined, local field potentials recorded, and dimensionality reduction performed on
activity across all the transistors. The result was that none of
these techniques came close to reverse engineering the standard stored-program computer architecture (Jonas and Kording, 2017).
A number of noteworthy points emerge from this study that
should be further highlighted. The treatment of ‘‘behavior’’
perfectly represents how the neuroscience field typically tends
to work with this concept. The behavioral data analyzed consisted of 10 s of spontaneous activity with no player actually
playing the game (Jonas and Kording, 2017). This is a fragment
of activity, which we refer to as ‘‘stimulus-response’’ (Box 1) to
distinguish it from behavior, an adaptive pattern of activity (i.e.,
one that enhances fitness). As such, this activity is a starting
point that is unlikely to result in understanding no matter how
advanced the subsequent analysis. But let’s suppose that
instead of a fragment, we have a complete activity map for an
entire game played by a person. Let’s further suppose that a
much better job, using better analysis algorithms, could be
done with the data. We suggest this would similarly lead to no
meaningful insights into the processor’s functional architecture,
since again no behavioral-level hypothesis is being tested; there
Neuron
Perspective
is no conceptual structure in place. Which of an infinite set of
potentially interesting patterns in the data should be selected
for further investigation (see Figure 1C, overly constrained
behavior)? The best way to answer this would be to examine
the game-play—the behavior—itself. An engineer trying to
make a copy of the machine would generate a high-level task
analysis of what the microprocessor needs to do in order to
play its part of the game. For example, she might study the onscreen positions, shapes, and colors of agents over time, the
value of point scores, and the responses to joystick input generated by the player. Then she might ask how is the chip in the machine fulfilling these higher-order needs of game-play. From that
starting point, more productive work on the role of specific portions of the chip would be possible.
Thus, two questions need to be asked. First, is causal-mechanistic explanation at the neural level better in principle than
algorithmic or computational accounts of behavior? Second, is
causal-mechanistic explanation sufficient to explain a behavioral
phenomenon? Clearly knowing the sufficient and necessary
conditions to evoke a behavior as found through an optogenetic
or similar manipulation can fall far short of knowing the rules
needed to instruct a robot or computer to perform the activity
in question. Thus, we would argue that if the question is ‘‘how
does the brain lead to behavior?’’ we are first asking why is the
brain performing this behavior and then asking how is it doing
it. So using the flying analogy (Figure 2), once we agree that
bird flight is an adaptive behavior, we then determine that it flies
by flapping it wings and not by wiggling its feet. Once we have
worked this out, we can start studying the feathers that make
up the wing. Seen this way, understanding that the flapping of
wings is critical to flight aids the subsequent study of feathers.
It is unlikely that, from the outset, studying an ostrich feather in
isolation would lead to the conclusion that there is such a phenomenon as flight or even that feather-like structures would be
useful for flight.
Why Higher-Level Concepts Are Needed to Understand
Neuronal Results: The Nature of ‘‘Mechanism’’
Why is it the case that explanations of experiments at the neural
level are dependent on higher-level vocabulary and concepts?
The answer is that this dependency is intrinsic to the very
concept of ‘‘mechanism.’’ A mechanism can be defined as ‘‘a
structure performing a function in virtue of its component parts,
component operations, and their organization. The orchestrated
functioning of the mechanism is responsible for one or more phenomena’’ (Bechtel, 2008, p. 13). Crucially, the components of a
mechanism do different things than the mechanism organized
as a whole (i.e., emergence) (Bechtel, 2008). A reductionist treatment of the components must be combined with investigation of
how the total mechanism is organized and how it behaves when
embedded in an environment; an approach that unavoidably
spans two levels (Bechtel, 2008) (Box 1). Even the reductionist
idea of causality needs to be qualified. An idea related to emergence is that of ‘‘downward causation.’’ Take, for example, the
cardiac rhythm—a behavior that is the net consequence of the
interplay between a cell’s membrane and the ion channels in it
(Noble, 2012). The conceptual point is that the ion channels do
not cause the cardiac rhythm—instead the rhythm just is the
combination of the higher level of the cell membrane and the
lower level of ion channels. So even when causality claims are
sought they often only make sense when all levels are considered together simultaneously rather than seeing the higher level
as subordinate or collapsible to the lower level. Ion channels do
not beat, heart cells do. Neural circuits do not feel pain, whole
organisms do.
A potential objection to this might be to say, ‘‘Who cares what
philosophers say about the differences between psychology
and neuroscience, or reductionism in general? We are scientists, not philosophers!’’ The answer to this is simple: there is
no escape from philosophy. Every scientist takes a philosophical position, either tacitly or explicitly, whenever they state
that a result is ‘‘important,’’ ‘‘fundamental,’’ or ‘‘interesting.’’
This is because such assertions are always a judgment from
outside of science. There is no ‘‘interesting’’ variable inherent
to the data that can be objectively plotted on a graph—abstract
reasoning and normative claims cannot be substituted by, or
obtained from, data. Tacit awareness that causal manipulative
work, which tests necessity and sufficiency claims, is not the
same as understanding is apparent with common sentences
like, ‘‘The circuit X is involved in behavior Y.’’ This is, however,
just a restatement of the correlation or causal relation and
does no extra explanatory work. The italicized word is known
as a filler term, which indicates the lack of an explicit conceptual
framework for the mapping between circuit and behavior
(Craver, 2008) and so just fills in for it (Figure 3). Importantly,
however, the use of filler terms signals a tacit awareness of
the lack of, and a desire for, a different kind of understanding,
which we argue can be obtained by doing empirical and theoretical behavioral work. This work would complement causal
explanations at the neural and circuit level. A nice statement
of this dual view of understanding in neuroscience was made
by the cognitive scientist Longuet-Higgins, ‘‘In so far as the
neurophysiologist is concerned to understand how the brain
works, he must equip himself with a non-physiological account
of the tasks which the brain and its peripheral organs are able to
perform; only then can he form mature hypotheses as to how
these tasks are carried out by the available ‘hardware’—to
borrow a phrase from computing science’’ (Longuet-Higgins,
1972, p. 256).
Behaviorally Driven Neuroscience Yields More
Complete Insights
Here we describe four examples of what we take to be the essential interplay between computational theories and algorithmic
formulations of behaviors, on the one hand, and their neural implementation, on the other.
Bradykinesia
Bradykinesia is one of the cardinal symptoms of Parkinson’s disease, which is manifest as a lack of movement vigor. Bradykinesia
is causally related to dopamine depletion in the substantia nigra
and is partially and transiently reversed by increasing dopamine
with medication. Do these facts truly help us understand why
dopamine depletion leads to bradykinesia? The answer is clearly
no—loss of a neurotransmitter may explain the necessary and
sufficient starting conditions for bradykinesia, and indeed the
investigation of these starting conditions—dopaminergic cell
Neuron 93, February 8, 2017 485
Neuron
Perspective
What is done
A
C
Most used
Manipulating a circuit
as a way of intervening in behavior
(i) Necessity:
on
o
on
B
on
What is claimed
reveals
involves
regulates
mediates
generates
modulates
shapes
underlies
produces
encodes
induces
enables
ensures
supports
promotes
determines
plays a role in
contributes to
is associated with
Figure 3. The Interventionist Type of
Understanding Is Not Sufficient
The current dominant notion of what constitutes
understanding in neuroscience is based on an
interventionist approach to causality.
(A) Intervening at the neural level (blue) as a way to
explain the behavior (magenta) through necessity
and sufficiency claims.
(B) The result that ‘‘X is necessary and sufficient for
Y to occur’’ allows a causal claim. An additional
explanatory sentence is often added but is merely
the same causal result rearticulated with a ‘‘filler’’
verb.
(C) Common filler verbs.
is the cause of
explanation
+
=
of
verb
death—is a major area of ongoing work at the cellular level. Nevertheless, work restricted to this level cannot explain why dopamine
depletion causes a loss of vigor. This is also true when neural
correlates are found. In contrast, psychophysical studies in patients with Parkinson disease suggested that low vigor in these
patients is caused by a loss of implicit motivation for moving
fast because of a skewed cost function containing effort and accuracy terms (Mazzoni et al., 2007). These human psychophysical
experiments then led to analogously designed experiments in
mice that demonstrated that there are cells in the dorsal striatum
whose activity correlates with movement vigor, and that suppression of these cells optogenetically reduces vigor (Panigrahi et al.,
2015). Note again that these mouse experiments very much
adhere to the causal interventionist approach to explanation.
The complete interpretation of these experiments, however, requires the explanatory framework provided by the initial human
behavioral work.
Sound Localization
The study of sound localization in avian and mammalian brains
provides a nice example of the value of the behaviorally driven
approach we are advocating. In the dark, both avian (e.g., barn
owls) and mammalian (e.g., gerbils) brains must localize sound
sources in the horizontal plane (that is, left or right with respect
to their own body plan). The behavioral problem is thus well
specified: namely, to localize a sound source based only on auditory cues.
One way to solve this problem is with inter-aural time difference cues. A way to calculate sound source location on the basis
of such time-difference cues in a principled way was proposed
by Jeffress (Jeffress, 1948). He showed that the computational
goal could be achieved via an algorithm that combines delay
lines (one waveform from each source, or ear) with coincidence
detectors, enabling a temporal activation pattern to be translated into a place code. In subsequent decades, Jeffress’ influential model stimulated a range of anatomical, physiological,
psychophysical, and formal investigations to understand sound
localization (for an excellent review, see Grothe, 2003). In several
tour-de-force experiments, the predictions of this algorithmic
model were in large part confirmed at the implementation level
486 Neuron 93, February 8, 2017
in barn owls. That is to say, the algorithmically inspired experiments sought to
identify possible implementations of the
procedures, such as delay lines and coincidence detectors.
Given that success, we now have a causal, mechanistic explanation of barn owl sound localization that is embedded in a
computational theory. The specific algorithm follows from the
details of the circuit (e.g., the relevant nuclei of the barn owl
are organized in a manner that directly licenses that algorithm
and not some other formal procedure). But, crucially, to identify
the precise circuitry that underpins delay-line-plus-coincidence
computation, the computational level (behavioral) characterization and the algorithmic hypothesis were necessary prerequisites to identify the neural substrates. It is hard to imagine how
(even extensive) recordings, staining, and anatomical studies
of the barn owl auditory system in the absence of such a behaviorally motivated computational theory would have yielded such
a rich explanatory framework.
It was subsequently discovered that the mammalian auditory
system, such as in gerbils, solves the same computational
problem using a different algorithm with different implementation at the circuit level (Grothe, 2003). This finding is a compelling example of multiple realizability and deep degeneracy,
since different implementational features yield different algorithms, but both solve the same computational problem
(Katz, 2016; Marom et al., 2009). Implementation/circuit A
yields algorithm A0 , which solves computational goal X; and
implementation B yields algorithm B0 , which also solves
computational goal X (Figure 1D). Once again, without formally
specified analyses at the behavioral level as well as explicit
algorithmic hypotheses, these fully developed mechanistic accounts would have been hard, indeed perhaps impossible, to
identify.
Electrolocation
The neurobiology of weakly electric fish is grounded in comparative, behavioral, and evolutionary analyses. Through a series of
studies, Walter Heiligenberg was able to provide perhaps the
most complete understanding of a vertebrate circuit from sensory input through to motor output in his analysis of the jamming
avoidance response (JAR). JAR is a natural behavior (although
possibly initially discovered as a stimulus response [Watanabe
and Takeda, 1963]) in which two weakly electric fish that are discharging an oscillating electric field at the same frequency will
Neuron
Perspective
modify their discharge frequencies so as to avoid ‘‘jamming’’
each other’s electrolocation system (Heiligenberg, 1991). This
is important because their ability to localize objects in the dark
is mediated by sensing distortions to their self-generated field.
Analysis of the behavior was first performed by investigators
who built a simple virtual world (Box 1) for electric fish: a small
tank in which a fish’s own electric field was picked up, then
slightly shifted in frequency and fed back, resulting in the reliable
elicitation of what is now called the JAR (Watanabe and
Takeda, 1963).
As the electric field is controlled by the fish’s motor system, it
enabled the complete characterization of a behavior using precisely controlled signals and set the groundwork for Heiligenberg’s subsequent discoveries of the circuits underlying the
behavior (Heiligenberg, 1991). From that period forward, electric
fish were typically studied by applying field distortions across
the entire fish. As it turns out, these ‘‘wide-field’’ signals are
interpreted by the fish as communication signals, since electrolocation targets such as prey result in focal ‘‘narrow-field’’ distortions. This was not fully understood until prey capture behavior
was automatically tracked in three dimensions, and then an
entire suite of empirically validated models was used to compute
the neuronal signal going to the brain over the course of prey
capture behavior (MacIver et al., 2001; Nelson and MacIver,
1999; Snyder et al., 2007). Once investigators started applying
distortions in focal manner to mimic the effect of prey, they unmasked a filtering system within the hindbrain that processed
these signals differently from the wide-field signals that had
been formerly been examined (reviewed in Fortune, 2006). The
point is that even within a behaviorally oriented model system,
understanding key components of this system did not enable a
full understanding of how the brain processes signals related
to predation until careful behavioral and computational simulation work was done.
Motor Learning
A final example of the benefits of a well-motivated behavioral
perspective comes from motor learning. Motor learning is often
studied using adaptation paradigms in which participants are
subjected to an external perturbation, which causes systematic
errors that are then corrected (Shadmehr et al., 2010). A large
body of empirical and theoretical work, originating in large part
from Marr himself, has suggested a crucial role for cerebellar
circuitry for this kind of learning (Therrien and Bastian, 2015). In
essence, Marr initially believed—prior to elaborating his idea of
levels of analysis—that the mechanism of error-based learning
could be reverse-engineered from examination at the circuit
level. Indeed, much work after Marr has made considerable
progress in determining how cerebellar circuitry performs error-based learning. Crucially, however, it is now known from
careful psychophysical work that many distinct learning algorithms operate together to counter the effects of a perturbation
during adaptation, even though phenotypically their summed
behavior can look like pure error-based cerebellar learning
(Huang et al., 2011; Taylor et al., 2014). This is a further example
of multiple realizability (Figure 1D), a long-standing basis for an
objection to reductionism (Sober, 1999). The core argument is
that if there are many ways to neurally generate the same
behavior, then the properties of a single circuit at best are a
particular instantiation and do not reveal a general design
principle.
The distinct algorithms operating in adaptation experiments
include error-based learning, reinforcement learning, and
cognitive strategies (Huberdeau et al., 2015). The relative
weighting of these processes depends on how the task is
framed with respect to the kind of feedback and instructions
given. For example, when subjects are given endpoint feedback and instruction they solve the same visuomotor rotation
task differently to when no instruction is given and they are
provided with continuous visual feedback. This higher-level organization operates at the level of the global task goal, which
can only be identified top-down by observing the summed
behavior and then decomposing it psychophysically, not by
first knowing how each of these components is neurally implemented (Taylor et al., 2014); the adaptation task is perceived
and solved by the whole brain in a body, not just by any given
circuit component.
Two points are illustrated with these four examples. The first
is that experiments at the level of neural substrate are best designed with hypotheses based on pre-existing behavioral work
that has discovered or proposed candidate algorithmic and
computational processes (for a range of arguments why experimental work based on algorithm-level hypotheses is foundational, see Cooper and Peebles, 2015). Second, the explanations
of the results at the neural level are almost entirely dependent on
the higher-level vocabulary and concepts derived from behavioral work. Lower levels of explanation do not ‘‘explain away’’
higher levels.
The Need for More Pluralistic Neuroscience
If we look far into the future of our science, what will it
mean to say we ‘understand’ the mechanism of behaviour? The obvious answer is what may be called the neurophysiologist’s nirvana: the complete wiring diagram of
the nervous system of a species, every synapse labelled
as excitatory or inhibitory; presumably, also a graph, for
each axon, of nerve impulses as a function of time during
the course of each behaviour pattern . Real understanding will only come from distillation of general principles at a
higher level, to parallel for example the great principles of
genetics—particulate inheritance, continuity of germ-line
and non-inheritance of acquired characteristics, dominance, linkage, mutation, and so on . it seems possible
that at higher levels some important principles may be
anticipated from behavioural evidence alone. The major
principles of genetics were all inferred from external evidence long before the internal molecular structure of
the gene was even seriously thought about.—Richard
Dawkins (Dawkins, 1976, pp. 7–8)
Neuroscience has been focused of late on neural circuits. This
is largely due to the recent development and incorporation of
techniques that allow both causal manipulation and the rapid
acquisition of large amounts of data. There seems to be an implicit assumption that implementation-level description will not
only allow causal claims but also somehow lead to algorithmic
and computational understanding (‘‘naive’’ emergence). We
Neuron 93, February 8, 2017 487
Neuron
Perspective
A
final
B
Marr’s
3 levels
computation
(level 1)
efficient
algorithm
(level 2)
Aristotle’s
4 causes
formal
implementation
(level 3)
Pluralistic
explanations
evolution
material
Where we want to be
understanding
Correcting
the bias
causal
manipulation
behavior
Tinbergen’s
4 questions
neurons
levels 1-2
ontogeny
level 3
function
mechanism
Where we are
Figure 4. The Future History of Pluralistic Explanation
(A) That understanding of a phenomenon is multidimensional has long been appreciated. Aristotle posited four kinds of explanation: to explain ‘‘why’’ something
changes, a polyhedric notion of causality is necessary; one that includes not only the material cause (what it is made out of), but also the other three ‘‘whys’’:
formal (what it is to be), efficient (what produces it), and final (what it is for). Tinbergen also devised four questions about behavior: to go beyond its proximate
causation (mechanism) to also considering its evolution, development, and real-world function. Marr’s three levels are also shown.
(B) Three-dimensional space with axes of understanding-manipulation, behavior-neurons, and Marr’s levels. The red box is where we are and the blue is where
we should be.
contend that such an approach is simply not going to yield the
kind of insight and explanation that we ultimately demand from
the neurosciences, at least those parts concerned with developing an understanding of the link between brain and behavior
that goes beyond causality claims.
Since the causal-manipulation view by itself will not lead to understanding, a more pluralistic conception of mechanistic understanding can only help neuroscience. Pluralism in science can be
defined as ‘‘the doctrine advocating the cultivation of multiple
systems of practice in any given field of science’’ (Chang,
2012). Here we have argued that when scientists ask ‘‘how
does the brain generate behavior,’’ they are in fact asking a question best approached through behavioral work, specifically task
analysis, aided by theory, that allows behavior to be decomposed into separable modules and processing operations.
As Woese has argued (Woese, 2004), science is driven by both
technological advances and a guiding vision. The key is to balance their contributions, ‘‘.without the proper technological advances the road ahead is blocked. Without a guiding vision there
is no road ahead’’ (Woese, 2004, p. 173). Insofar as the goal of a
neuroscience research question is to explain some behavior, be
it a phenomenon from vision, communication, motor control,
navigation, language, memory, or decision making, the behavioral research must be considered, for the most part, epistemologically prior (Figure 4). The neural basis of behavior cannot be
properly characterized without first allowing for independent
detailed study of the behavior itself.
ACKNOWLEDGMENTS
We thank the following for invaluable critical discussion about and/or careful
reading of the evolving manuscript: Björn Brembs, Philip Buttery, Rui Costa,
William Dauer, Greg DeAngelis, Shimon Edelman, Stuart Firestein, Adrian
Haith, David Huberdeau, Don Katz, Konrad Körding, David Linden, Zachary
488 Neuron 93, February 8, 2017
Mainen, Tony Movshon, Yael Niv, Bence Ölveczky, Bijan Pesaran, Scott
Rennie, Nick Roy, Scott Small, Aaron Wong, and Jing Xu.
REFERENCES
Anderson, P.W. (1972). More is different. Science 177, 393–396.
Anderson, D.J., and Perona, P. (2014). Toward a science of computational
ethology. Neuron 84, 18–31.
Badre, D., Frank, M.J., and Moore, C.I. (2015). Interactionist Neuroscience.
Neuron 88, 855–860.
Bargmann, C.I. (1998). Neurobiology of the Caenorhabditis elegans genome.
Science 282, 2028–2033.
Bass, A.H., and Chagnaud, B.P. (2012). Shared developmental and evolutionary origins for neural basis of vocal-acoustic and pectoral-gestural
signaling. Proc. Natl. Acad. Sci. USA 109 (Suppl 1 ), 10677–10684.
Bechtel, W. (2008). Mental Mechanisms: Philosophical Perspectives on Cognitive Neuroscience (Routledge).
Benichov, J.I., Benezra, S.E., Vallentin, D., Globerson, E., Long, M.A., and
Tchernichovski, O. (2016). The forebrain song system mediates predictive
call timing in female and male zebra finches. Curr. Biol. 26, 309–318.
Bianco, I.H., Kampff, A.R., and Engert, F. (2011). Prey capture behavior evoked
by simple visual stimuli in larval zebrafish. Front. Syst. Neurosci. 5, 101.
Bickle, J. (2015). Marr and reductionism. Top. Cogn. Sci. 7, 299–311.
Bickle, J. (2016). Revolutions in neuroscience: Tool development. Front. Syst.
Neurosci. 10, 24.
Borjon, J.I., and Ghazanfar, A.A. (2014). Convergent evolution of vocal cooperation without convergent evolution of brain size. Brain Behav. Evol.
84, 93–102.
Carandini, M. (2012). From circuits to behavior: a bridge too far? Nat. Neurosci.
15, 507–509.
Chang, H. (2012). Is Water H2O?: Evidence, Realism and Pluralism
Volume 293 (Springer Science & Business Media).
Neuron
Perspective
Cooper, R.P., and Peebles, D. (2015). Beyond single-level accounts: the role of
cognitive architectures in cognitive scientific explanation. Top. Cogn. Sci. 7,
243–258.
Craver, C.F. (2008). Explaining the brain: mechanisms and the mosaic unity of
neuroscience. Psychol. Med. 38, 899–900.
Marder, E., and Goaillard, J.M. (2006). Variability, compensation and homeostasis in neuron and network function. Nat. Rev. Neurosci. 7, 563–574.
Marom, S., Meir, R., Braun, E., Gal, A., Kermany, E., and Eytan, D. (2009). On the
precarious path of reverse neuro-engineering. Front. Comput. Neurosci. 3, 5.
Marr, D. (1982/2010). Vision: A Computational Approach (MIT Press).
Dawkins, R. (1976). Hierarchical organisation: A candidate principle for
ethology. In Growing Points in Ethology, P. Bateson and R.A. Hinde, eds.
(Cambridge University Press), pp. 7–54.
Mazzoni, P., Hristova, A., and Krakauer, J.W. (2007). Why don’t we move
faster? Parkinson’s disease, movement vigor, and implicit motivation.
J. Neurosci. 27, 7105–7116.
Dennett, D.C. (1989). The Intentional Stance (MIT Press).
di Pellegrino, G., Fadiga, L., Fogassi, L., Gallese, V., and Rizzolatti, G. (1992).
Understanding motor events: a neurophysiological study. Exp. Brain Res. 91,
176–180.
Dombeck, D.A., and Reiser, M.B. (2012). Real neuroscience in virtual worlds.
Curr. Opin. Neurobiol. 22, 3–10.
Fetsch, C.R. (2016). The importance of task design and behavioral control for
understanding the neural basis of cognitive functions. Curr. Opin. Neurobiol.
37, 16–22.
Fortune, E.S. (2006). The decoding of electrosensory systems. Curr. Opin.
Neurobiol. 16, 474–480.
Frank, M.J., and Badre, D. (2015). How cognitive theory guides neuroscience.
Cognition 135, 14–20.
Gomez-Marin, A., Paton, J.J., Kampff, A.R., Costa, R.M., and Mainen, Z.F.
(2014). Big behavioral data: psychology, ethology and the foundations of
neuroscience. Nat. Neurosci. 17, 1455–1462.
Grothe, B. (2003). New roles for synaptic inhibition in sound localization. Nat.
Rev. Neurosci. 4, 540–550.
Nelson, M.E., and MacIver, M.A. (1999). Prey capture in the weakly electric fish
Apteronotus albifrons: sensory acquisition strategies and electrosensory consequences. J. Exp. Biol. 202, 1195–1203.
Noble, D. (2012). A theory of biological relativity: no privileged level of causation. Interface Focus 2, 55–64.
Oatley, K. (1978). Perceptions and Representations: The Theoretical Bases of
Brain Research and Psychology (Methuen).
Panigrahi, B., Martin, K.A., Li, Y., Graves, A.R., Vollmer, A., Olson, L., Mensh,
B.D., Karpova, A.Y., and Dudman, J.T. (2015). Dopamine is required for the
neural representation and control of movement vigor. Cell 162, 1418–1430.
Patterson, B.W., Abraham, A.O., MacIver, M.A., and McLean, D.L. (2013).
Visually guided gradation of prey capture movements in larval zebrafish.
J. Exp. Biol. 216, 3071–3083.
Poeppel, D., and Embick, D. (2005). Defining the relation between linguistics
and neuroscience. In Defining the Relation between Linguistics and Neuroscience Twenty-First Century Psycholinguistics: Fo…

Order an Essay Now & Get These Features For Free:

Turnitin Report

Formatting

Title Page

Citation

Outline

Place an Order