Task learning increases information redundancy of neural responses in macaque visual cortex

TL;DR

Strong support is found for Bayesian predictions that task learning increased redundancy in neural responses over weeks of training and within single trials, suggesting that sensory processing in the brain reflects a generative rather than discriminative inference process.

Abstract

How does the brain optimize sensory information for decision-making in new tasks? One hypothesis suggests learning reduces redundancy in neural representations to improve efficiency, while another, based on Bayesian inference, predicts learning increases redundancy by distributing information across neurons. We tested these hypotheses by tracking population responses in macaque cortical area V4 as monkeys learned visual discrimination tasks. We found strong support for the Bayesian predictions: task learning increased redundancy in neural responses over weeks of training and within single trials. This redundancy did not reduce information but instead increased the information carried by individual neurons. These insights suggest sensory processing in the brain reflects a generative rather than discriminative inference process.

Figures (32)

• Figure 1: Illustration of experiment and key predictions.a: Response of two example neurons to a horizontal and a vertical stimulus of intermediate signal strength ("coherence"). Each neuron's response is variable (middle), and this variability is correlated. In this example, positive correlations ($\rho = 0.32$, Pearson correlation) align with the task-tuning axis (both neurons prefer 0 degree stimulus), reducing the decoding performance (right). b: Linear Fisher information as a function of population size for two example sessions: during the task early in learning, during the task late in learning, and during passive viewing late in learning. $I_{\rm real}$ (orange) denotes actual (linear) Fisher information, and $I_{\rm shuffle}$ (yellow) represents the linear Fisher information after shuffling trials, thereby destroying noise correlations. c: Differential predictions for changes in redundancy over the course of task learning.
• Figure 2: Experimental design and behavior.a: Task design: after animals fixated the center to start a trial. Choice targets were shown, followed by the stimulus. After stimulus offset, animals indicated their choice by a saccade to one target. Bottom: study timeline. b: Example stimulus images for four orientations and two coherence levels. c: Spatial layout of the fixation point (blue dot), stimulus (dashed circle), and choice targets (black and white dots). The red numbers (not shown during the experiment) indicate the stimulus orientation corresponding to each choice target. d: Psychometric curves of late-learning sessions. e: Orientation kernels of late-learning sessions. Dashed lines: ideal observer predictions. Solid lines: across-sessions average. Shaded areas: the standard error of the mean across sessions. The bars on the right indicate the bias averaged across sessions, and the error bars indicate the standard error. (The temporal kernels in Fig. \ref{['fig:Figure_S18_temporal_kernel_afterlearning']}) f: Time course of the learning index for two animals and two tasks. Solid lines and shading represent medians and 68% confidence intervals across 1000 bootstrap samples. Gray vertical lines separate each epoch into "early-learning" and "late-learning". (Time course with real experiment dates can be found in Fig. \ref{['fig:Figure_S10_timecourse_behavior_realdate']})
• Figure 3: Redundancy increases over the course of learning.a: Time course of $I_{\rm redundancy}$ during the course of learning using sub-sampled populations with consistent number of units (monkey R: cardinal: $n=17$, oblique: $n=18$; monkey G: oblique $n=44$, cardinal $n=55$). Lines and error bars are means and 68% confidence interval across bootstrapped samples. b: Scatter plots between the learning index and $I_{\rm redundancy}$ for two tasks and two animals. Each dot represents a session. Black dashed lines are linear regression fit across sessions. c: We separated sessions into two groups: "early" and "late", using the boundary show in Fig. \ref{['Figure2_exp_behav']}f. The bar plots averaged $I_{\rm redundancy}$ across sessions in each learning stage when the animals were performing the task or passively viewing the stimulus. Error bars indicate the standard error of the mean across sessions. (Significance levels: $^{*}p < 0.05$, $^{**}p < 0.01$, $^{***}p < 0.001$)
• Figure 4: Increase in redundancy is explained by increase in $I_{\rm shuffle}$, not a decrease in $I_{\rm real}$.a: Information sharing during generative inference. The activities of sensory neurons, $r_{i}$, represent posterior beliefs about the intensity of oriented edges in the input stimulus. A decision-making area computes a belief about a decision variable, $p(d)$. This belief, based on the information in all sensory neurons, informs all sensory neurons via feedback signals (blue arrows). b, c: Model haefner2016perceptual simulations at different stages of learning. Errorbars indicate 68% confidence interval across random samples of 32 (panel b) and 8, 32, 256 (panel c) out of a total of 256 model neurons. In panel c, for each population size, the values were normalized by the mean $I_{\rm real}$ before learning. d: Time course of $I_{\rm real}$, $I_{\rm shuffle}$ and behavioral Fisher information $I_\textbf{behav}$ during learning. Shaded areas: 68% across Bootstrap samples. e: Correlation between learning index $I_{\rm real}$ and $I_{\rm shuffle}$, respectively. One dot per session. Black lines are a linear regression fit across sessions, solid for $I_{\rm real}$ and dashed for $I_{\rm shuffle}$. (Significance levels: $^{*}p < 0.05$, $^{**}p < 0.01$, $^{***}p < 0.001$)
• Figure 5: Redundancy increases within a trial. a: Temporal propagation of beliefs during generative inference. $t_1..t_4$ represent times within a single trial. b: Change in $I_{\rm redundancy}$ across eight time bins within a trial for three stages of learning (model). Error bars are 68% confidence interval across 256 random sub-samples (size = 32) of model neurons. c: $I_{\rm redundancy}$ for each 200ms time bin within a trial in empirical data, controlled for population size. Solid line: average across late-learning sessions. Dashed line: average across early-learning sessions. $\beta$ is the regression coefficient: $I_{\rm redundancy}$ vs time bin. Error bars indicate the standard error of the mean. across sessions (Significance levels: $^{*}p < 0.05$, $^{**}p < 0.01$, $^{***}p < 0.001$)