Researchers at New York University have developed a computational neural-circuit model that explains how the brain learns to sort continuous sensory inputs into discrete categories—think “car vs. motorcycle” or “dog vs. cat.” Published in Nature Communications, the study clarifies neural mechanisms of category learning and provides testable predictions about the role of feedback signals in shaping sensory representations.
Categorization is a fundamental cognitive ability: it helps animals and humans quickly identify edible versus inedible objects, form concepts, and organize relations among classes of things (for example, hierarchical groupings within the animal kingdom). The NYU team makes clear that their model addresses categorization of relatively simple visual features and that further work will be needed to determine whether the same principles apply to more complex or abstract categories.
The study was led by Xiao-Jing Wang, Global Professor of Neural Science, Physics, and Mathematics at NYU and NYU Shanghai, in collaboration with Tatiana Engel (then a postdoctoral associate), doctoral candidate Jah Chaisangmongkon, and experimental collaborator David Freedman at the University of Chicago. Freedman’s prior experimental paradigms—recording single-neuron activity that correlates with category membership of visual stimuli—provided critical data for testing the model.
The proposed cortical circuit model reflects current knowledge about cortical organization and physiology. It describes how lower-level sensory circuits carry information about an analog stimulus feature (for example, the direction of a cloud of moving dots) to a higher-level circuit where that continuous feature is classified into two categories (A or B). The model reproduces a wide range of experimental observations and makes precise predictions that the authors confirmed by analyzing single-neuron recordings collected during a category-learning task.
A key and surprising insight from the model is that learning an accurate category boundary—the point that divides a continuous feature space into category A versus category B—depends on top-down feedback from category-selective neurons back to neurons that code features. In other words, category signals do not simply arise by pooling noisy sensory inputs; instead, feedback from higher-level category representations actively reshapes sensory coding during learning.
For decades, beginning with influential work by J. Anthony Movshon, William Newsome and others, neuroscientists have observed that sensory neurons often carry probabilistic signals related to an animal’s categorical choice, a phenomenon quantified as choice probability. The prevailing interpretation has been that stochastic variability in sensory neurons propagates upward through feedforward pathways to influence categorical decisions.
The NYU model offers a different perspective: much of the observed choice-correlated activity in sensory areas may result from category-to-sensory, top-down signaling. As category representations consolidate through learning, top-down feedback alters both the tuning of intermediate neurons and their correlated variability, producing the choice-related signatures recorded in sensory cortex.
This research was supported by the National Institute of Mental Health (R01MH092927) and the Swartz Foundation. The model combines reward-dependent plasticity and circuit-level dynamics to explain how stable category representations emerge in neurons that lie between sensory and decision-making layers. According to the authors, plasticity of top-down projections from decision neurons gives rise to both choice probability and task-specific interneuronal correlations.
Contact: James Devitt – NYU
Source: NYU press release
Image Source: The image is credited to geralt and is in the public domain.
Original Research: Open-access research article in Nature Communications: “Choice-correlated activity fluctuations underlie learning of neuronal category representation” by Tatiana A. Engel, Warasinee Chaisangmongkon, David J. Freedman and Xiao-Jing Wang. Published online March 11, 2015 (doi:10.1038/ncomms7454).
Choice-correlated activity fluctuations underlie learning of neuronal category representation
To investigate the neural mechanisms of categorization, the authors built a cortical-circuit model capable of learning a motion-categorization task through reward-dependent plasticity. The model demonstrates how stable category representations develop in neurons positioned between sensory and decision layers when those neurons exhibit choice-correlated activity fluctuations (choice probability). It shows that choice probability and task-specific correlated activity naturally arise from plastic top-down projections originating in decision-making neurons. The model’s predictions were tested using single-neuron recordings from monkey parietal cortex and confirmed experimentally: neurons there show a mix of directional and categorical tuning, and neurons with stronger category selectivity also display higher choice probability. Beyond providing a circuit mechanism for categorization, the study highlights a pivotal role for adaptive top-down feedback in shaping both neural tuning properties and correlated variability during learning.