Task Priors in Neural Decision Making

• Task priors are a mechanism where prior knowledge is embedded in synaptic weights rather than transient neural activity.
• They are crucial in perceptual decision making tasks, improving choices when sensory evidence is weak and environmental statistics shift.
• Hybrid models combining reinforcement learning, feedforward networks, and Bayesian updates reveal that learned priors drive behavioral adaptation while remaining weakly decodable from single-trial neural signals.

Searching arXiv for the specified paper to ground the article and citations. {"^^^^^^^^2query2 \2^^^^^^^^"Encoding priors in the brain: a reinforcement learning model for mouse decision making\"","max_results":5,"sort_by":"relevance"} "Encoding priors in the brain: a reinforcement learning model for mouse decision making" examines how prior knowledge may be represented in neural decision systems during perceptual choice. In the setting of two-alternative forced choice behavior, the paper addresses a standard hypothesis—that the prior is stored in neural activity—and proposes an alternative: the prior is stored in synaptic strengths. The model is developed for the International Brain Laboratory task, in which a mouse reports whether a grating appears on the right or left side of a screen by moving a wheel, under unsignaled blocks in which the right-side probability is either PRESERVED_PLACEHOLDER_^^^^2query2^^^^ or PRESERVED_PLACEHOLDER_^^^^2(Krishnagopal et al., 2021) \2^^^^, with blocks of about 5^^^^2query2^^^^ trials. The paper formulates this as a reinforcement learning problem with a feedforward neural network trained by policy gradient, augmented by an internal state that stores an estimate of the grating and confidence, follows Bayesian updates, and can switch between engaged and disengaged states to mimic animal behavior (&&&^^^^2query2^^^^&&&).

^^^^2(Krishnagopal et al., 2021) \2^^^^. Conceptual question and theoretical stance

The central question is how priors are encoded in the brain during perceptual decision making. In the task studied here, prior knowledge is behaviorally useful because when sensory evidence is weak, a subject can improve performance by favoring the more likely alternative. The paper states this in the context of two-alternative forced choice tasks operating near the psychophysical threshold, where one choice may be much more likely than the other and can therefore be selected when evidence is weak (&&&^^^^2query2^^^^&&&).

The paper contrasts two hypotheses. The common hypothesis is that the prior is stored in neural activity. The proposed alternative hypothesis is that the prior is stored in synaptic strengths. This distinction matters because it changes the expected neural signature of prior use. If priors are stored in activity, one might expect a trial-resolved neural variable carrying block identity. If priors are stored in weights, the behavioral effect of the prior could be substantial even when block structure is only weakly expressed in momentary activity. This suggests a representational dissociation between what shapes behavior and what is readily decodable from single-trial population states.

2. Behavioral task and probabilistic structure

The behavioral setting is the International Brain Laboratory task. A grating appears on either the right or left side of a screen, and a mouse has to move a wheel to bring the grating to the center. The task is made difficult because the grating is often low in contrast. The prior probability that the grating appears on the right is either $80\%$ or $20\%$, in unsignaled blocks of about 5^^^^2query2^^^^ trials (&&&^^^^2query2^^^^&&&).

These unsignaled block switches are the source of the task prior. Because the block identity is not cued directly, the subject must infer the current prior from recent trial history. The paper therefore treats the task as one in which weak sensory evidence and slowly changing latent environmental statistics jointly determine choice. A plausible implication is that the model must combine sensory inference with a mechanism for integrating recent evidence about block structure.

The paper emphasizes the psychometric consequence of this structure. The main experimental finding reproduced by the model is that the psychometric curve with respect to contrast shifts after a block switch in about ^^^^2(Krishnagopal et al., 2021) \2query2^^^^ trials. In the paper’s framing, this shift is the key behavioral signature that a prior has been acquired and updated (&&&^^^^2query2^^^^&&&).

3. Reinforcement learning architecture

The model treats the task as a reinforcement learning problem. A feedforward neural network maps states to actions, and the network weights are adjusted to maximize reward using policy gradient. This is the computational substrate through which the paper instantiates the claim that priors can be stored in synaptic strengths (&&&^^^^2query2^^^^&&&).

The model includes an internal state with three specified functions. First, it stores an estimate of the grating. Second, it stores confidence. Third, it follows Bayesian updates. This internal state is therefore not a generic hidden vector; it is structured to track both sensory content and uncertainty. The inclusion of Bayesian updates is important because it ties the latent-state dynamics to normative inference over uncertain evidence rather than only to reward-driven trial-and-error adaptation.

The model also switches between engaged and disengaged states to mimic animal behavior. This element is notable because it separates the decision mechanism from a higher-level behavioral state variable. Within the paper’s framework, disengagement is not treated as noise in the same sense as sensory unreliability; it is part of the latent behavioral regime required to reproduce observed patterns.

A concise summary of the model components stated in the paper is given below.

Component	Role
Feedforward neural network	Maps states to actions
Policy gradient learning	Adjusts weights to maximize reward
Internal state	Stores grating estimate and confidence; follows Bayesian updates
Engaged/disengaged switching	Mimics animal behavior

The architecture is conceptually important because it combines three levels that are often treated separately: perceptual belief updating, action selection, and reinforcement-driven synaptic modification. This suggests a mechanistic account in which prior-sensitive behavior can emerge from weight adaptation without requiring a strongly explicit activity code for block identity on each trial.

4. Prior encoding in synaptic strengths

The paper’s defining hypothesis is that the prior is stored in synaptic strengths rather than neural activity. In the model, this means that the effect of block history is expressed through learned network parameters that shape policy, not through a large, directly readable activity difference between right and left blocks on single trials (&&&^^^^2query2^^^^&&&).

This claim is tied to a specific empirical prediction. The paper reports that, as in the experiments, the difference in neuronal activity in the right and left blocks is small. More strongly, it states that it is virtually impossible to decode block structure from activity on single trials if noise is about $2\%$ (&&&^^^^2query2^^^^&&&). This is the main neural argument for a synaptic-weight account: behavior can be prior-sensitive while activity remains only weakly block-informative at the single-trial level.

The paper therefore reorients the interpretation of weak decodability. Under an activity-storage view, weak block decoding might be seen as a challenge to the idea that the subject is using the prior robustly. Under the weight-storage view, weak decoding is expected. The prior can shape the policy through connectivity while leaving only small block-dependent modulations in ongoing activity.

A plausible implication is that behavioral adaptation and neural decodability need not scale together. Strong behavioral use of priors can coexist with weak trial-level neural signatures if the relevant information is embedded in synaptic structure and only indirectly expressed in activity.

5. Behavioral and neural results

The paper reports that the model reproduces the main experimental finding: the psychometric curve with respect to contrast shifts after a block switch in about ^^^^2(Krishnagopal et al., 2021) \2query2^^^^ trials (&&&^^^^2query2^^^^&&&). This result is central because it links the reinforcement-learning model directly to the characteristic timescale of prior updating in the task.

The paper also reports a neural result aligned with experiment: the difference in neuronal activity between right and left blocks is small. Under noise of about $2\%$, block structure is virtually impossible to decode from activity on single trials (&&&^^^^2query2^^^^&&&). Together, these two findings define the paper’s core contribution. The model simultaneously accounts for relatively rapid behavioral adaptation and weak single-trial neural discriminability of block structure.

These results support the paper’s broader interpretive move. The model does not merely fit behavior; it also offers an explanation for why neural activity might fail to show a large explicit prior signal despite strong prior-dependent choice effects. This suggests that the absence of a robust trialwise neural code for block identity does not, by itself, rule out effective prior use.

6. Interpretation, limits, and experimental outlook

The paper presents the hypothesis that priors are stored in weights as difficult to test, but states that the technology to do so should be available in the not so distant future (&&&^^^^2query2^^^^&&&). This is an explicit acknowledgment that the proposal is mechanistic and experimentally ambitious. The model yields a concrete explanatory framework, but the decisive evidence would require methods capable of linking behavioral prior use to synaptic-level changes rather than only to population activity.

A common misconception would be to treat the proposal as denying any role for neural activity in prior-guided choice. The paper does not claim that activity is irrelevant; rather, it argues against the stronger claim that the prior itself must be stored as an explicit activity variable. In the model, activity still carries sensory estimates, confidence-related information, and action-relevant state, while synaptic strengths encode the learned prior structure that biases policy.

Another possible misconception would be to read the model as a purely Bayesian observer account. The paper instead combines Bayesian internal-state updates with reinforcement learning through policy gradient in a feedforward network. The resulting framework is hybrid: Bayesian structure governs the internal estimate and confidence dynamics, while reward maximization governs weight adaptation and action policy (&&&^^^^2query2^^^^&&&).

The broader significance of the work lies in reframing what should count as evidence for prior representation in neural systems. If priors can be embedded in synaptic strengths, then the correct experimental target may be neither a single “prior neuron” nor a strongly decodable block-state signal, but a distributed change in connectivity that alters how weak sensory evidence is converted into action.