Choice-correlated activity fluctuations underlie learning of neuronal category representation

Abstract

The ability to categorize stimuli into discrete behaviourally relevant groups is an essential cognitive function. To elucidate the neural mechanisms underlying categorization, we constructed a cortical circuit model that is capable of learning a motion categorization task through reward-dependent plasticity. Here we show that stable category representations develop in neurons intermediate to sensory and decision layers if they exhibit choice-correlated activity fluctuations (choice probability). In the model, choice probability and task-specific interneuronal correlations emerge from plasticity of top-down projections from decision neurons. Specific model predictions are confirmed by analysis of single-neuron activity from the monkey parietal cortex, which reveals a mixture of directional and categorical tuning, and a positive correlation between category selectivity and choice probability. Beyond demonstrating a circuit mechanism for categorization, the present work suggests a key role of plastic top-down feedback in simultaneously shaping both neural tuning and correlated neural variability.

Figure 1. Categorization task and the neural circuit model.

(a) A set of 12 motion direction stimuli is divided into two categories, C1 and C2 (red and blue arrows), separated by a category boundary (black dashed line). On each trial, one randomly chosen motion stimulus is presented, and the model learned through trial and error to indicate its category membership. (b) Schematic of the circuit model. The network comprises a sensory (MT), an association (LIP) and a decision neural circuits. Neurons in the sensory circuit are tuned to motion directions (indicated by arrows). They receive directional bottom-up inputs and provide inputs to the association neurons through feedforward synapses (c^S→A). The decision circuit (C₁ and C₂ populations) pools activity of association neurons through feedforward synapses (c^A→D) and generates a category decision through competitive attractor dynamics. The model has feedback connections from the decision to association neurons (c^D→A). All synaptic connections between the local circuits undergo Hebbian plasticity modulated by a reward prediction error signal. (c) An example network activity before categorization training. A motion direction stimulus (195°) is presented for 1 s (grey bar). The sensory and association neurons show direction-tuned responses in their spatiotemporal activity patterns (lower and middle panels, respectively). x axis, time; y axis, neurons labelled by the preferred direction; firing rate is colour-coded. The decision circuit generates categorical choice through a winner-take-all competition between the C₁ and C₂ populations (upper panel). Firing rates of the C₁ and C₂ populations are shown for two trials, where C₁ (red line) and C₂ (blue line) choice was made for the same stimulus.

Figure 2. Behavioural performance of the network models during training on the motion categorization task.

Performance of the networks with feedback (green), without feedback (purple) and with fixed tuning of association neurons (grey) is shown at three stages of learning: short (a,d), intermediate (b,e) and long (c,f). (a–c) Overall percent correct responses as a function of the number of trials performed. (d–f) Psychometric functions evaluated at the end of each training epoch: percent correct responses for stimuli close to (15°) and farther from (45° and 75°) category boundary. At the short training stage, performance improved equally in all three models. As the training progressed, performance of the network with feedback steadily improved especially for the near-boundary stimuli, while performance of the network without feedback gradually deteriorated and eventually dropped to the chance level. Performance of the network with fixed tuning of association neurons remained at the level attained by the end of the short training stage. Shaded area in a–c and error bars in d–f indicate s.d. across five independent realizations of each network type.

Figure 3. Mixed direction and category tuning emerges in association neurons through learning.

(a) Tuning profiles of association neurons after short (6 × 10³ trials, upper row) and long (420 × 10³ trials, lower row) periods of training for the network without feedback (left column) and with feedback (right column). x axis, stimulus motion direction; y axis, neurons arranged and labelled by their preferred direction before learning; firing rate is colour-coded. After extensive categorization training, motion tuning deteriorates in the network without feedback, whereas categorical tuning develops in the network with feedback. (b) The average category-tuning index of association neurons steadily increases over the course of training. (c–e) Mixed direction and category tuning of the association neurons at the intermediate stage of training (65 × 10³ trials). (c) Tuning profiles of two example association neurons before (grey dashed line) and after learning (coloured dots—firing rates, black solid line—best-fitted tuning function). Tuning curves broaden (right panel) and shift towards category centres (left panel) in neurons with initial preferred directions near category centres and category boundaries, respectively. (d) Bimodal distribution of preferred directions in direction-tuned association neurons (Hartigan’s dip test, P=0.002). Majority of neurons are tuned away from the category boundary (indicated by red dashed lines). (e) Multidimensional scaling analysis reveals a circular configuration of motion directions in the representation of sensory neurons (left panel), and an elliptical configuration elongated along the axis perpendicular to the category boundary in the representation of association neurons (right panel).

Figure 4. Mixed direction and category tuning in LIP neurons.

(a) Average response to the 12 motion directions for three example LIP neurons, classified as direction-tuned (upper row), category-tuned (middle row) and mixed direction and category-tuned (lower row). Left column: tuning curves, the average standardized firing rates during stimulus presentation (coloured dots) overlaid by the best-fitted tuning function (black line). Error bars indicate s.e.m. Right column: average firing rate traces in response to stimuli from category C1 (red lines) and C2 (blue lines). (b) The fraction of MT (N=67) and LIP (N=156) neurons that exhibited directional tuning, category tuning, mixed tuning or were not stimulus-selective. (c) Bimodal distribution of preferred directions in direction-tuned LIP neurons (Hartigan’s dip test, P=0.003, N=109), in support of the model prediction (Fig. 3d). (d) Multidimensional scaling analysis reveals a nearly uniform representation of motion directions in the MT population (upper panel), and an elongated representation in the LIP population (lower panel), where the distances are larger between the stimuli from different categories, c.f. the model prediction in Fig. 3e.

Figure 5. Choice probability determines the direction and magnitude of synaptic changes.

(a) Schematic of a toy model: firing rate of a neuron is correlated with choices in a behavioural task, where choice C₁ is rewarded and C₂ is not. Synapse of this neuron is updated according to reward-dependent Hebbian plasticity. (b,c) Firing rate distributions of a toy-model neuron on trials when different choices are made (C₁—red bars, C₂—blue bars). CP is close to 0.5, if the distributions largely overlap, that is, rate fluctuations are not correlated with choices (b). CP deviates from 0.5 towards 1 (or 0) when the distributions are well separated, that is, rate fluctuations are correlated with choices (c). (d) The sign of CP−0.5 determines the direction of synaptic changes: CP>0.5 leads to potentiation and CP<0.5 leads to depression (CP is computed relative to the rewarded choice C₁). The larger is the deviation of CP from 0.5, the faster is the rate of synaptic changes. Synaptic strengths are shown across learning trials for different CP values over an intermediate period of learning (three independent realizations for each CP value). (e) Synaptic strengths are shown across many learning trials for CP≈0.5 (corresponds to yellow traces in d, note the difference in scale of the x axis between d and e). Synaptic changes are random, and any weight becomes equally likely over a long period of learning. Four independent realizations are shown.

Figure 6. Association neurons in the network with feedback, but not in the network without feedback, exhibit choice-correlated fluctuations.

(a) In the network without feedback, CP is close to 0.5 in all association neurons and does not change throughout learning. (b) In the network with feedback, CP is randomly scattered around 0.5 before learning (grey dots), but a bimodal profile of CP develops after a short period of learning (500 trials, green dots), such that CP>0.5 in neurons with preferred directions in category C1, and CP<0.5 in neurons with preferred directions in category C2. CP (y axis) is plotted for all association neurons labeled by their initial preferred direction (x axis) before (grey dots) and after (olive dots) learning. Histograms to the right show the corresponding CP distributions.

Figure 7. The covariance between reward and neural activity drives tuning changes in association neurons.

(a,b) Covariance Cov[R, r_i|θ] between the firing rate and reward (y axis) measured for each of the 12 motion direction stimuli (x axis) after 1,000 trials of learning. The covariance profile is shown for two association neurons with preferred directions near category centre (a) and near category boundary (b), the same sample neurons as in Fig. 3c. The covariance has opposing sign for stimuli in different categories (see equation 3). The magnitude of the covariance is proportional to the product of probability to make a correct response and the probability to make an error (grey-scale code bar above the graph), which is the largest for near-boundary stimuli. (c,d) The expected weight change for the synapses from all sensory neurons (x axis) to a single association neuron; c,d correspond to the association neurons shown in a,b, respectively. The weight changes expected for each stimulus (red lines—stimuli from category C1, blue lines—stimuli from C2, left y axis) add up to produce the total expected weight change (black dashed line, right y axis). (e,f) The weight changes from c,d accumulate over multiple trials. As a result, the synaptic profile from sensory neurons is gradually broadened for the association neuron tuned to the category centre, and shifted towards the category centre for the association neuron tuned near the category boundary. These changes of the synaptic profiles underlie neural tuning changes shown in Fig. 3c.

Figure 8. Model predicts interdependence between the CP, CS and noise correlations.

(a) In the association neurons, a positive correlation between the CP and CS arises through learning. In the scatter plot, each dot represents a CP–CS pair for an association neuron (measured after 500 trials), colour-coded according to its category preference (C1—red, C2—blue, nonselective—grey). (b) A significant positive correlation between the CP and CS was found in LIP neurons recorded from behaving monkeys, similar to the model prediction in a. (c) Histograms of CP in MT and LIP populations. Black shading indicates neurons with individually significant CP. Overall magnitude of CP was significantly greater in LIP than in MT. (d,e) Plasticity of the feedback connections from decision neurons gives rise to task-specific noise correlations. (d) After learning, noise correlations decrease with the difference in preferred directions of two neurons, but are stronger for neurons preferring the same category, than for neurons preferring different categories. Error bars indicate s.d. across neurons. (e) Noise correlations are stronger in pairs of association neurons with more similar and individually larger CS. Scatter plot of noise correlation (y axis) versus the absolute difference between the category sensitivities (x axis). Each dot is coloured according to the average CS strength in the pair (purple-to-orange colour code corresponds to low-to-high CS strength). Histograms show the distributions of noise correlations for neural pairs with the high (>0.4) and low (<0.1) average CS strength (orange and purple, respectively).