Neuronal mechanisms of learning-evoked stimulus orthogonalization

In nature, an animal's survival often depends on learning to discriminate between objects with similar features. Our recent work has established that this learning is associated with changes in activity across the population of neurons in primary visual cortex (V1). Learning causes neurons with weak preferences for stimuli to be suppressed by stimuli informative to the animal, while neurons with strong stimulus preferences are only mildly affected. This learning-evoked effect results in a sparse number of neurons responding strongly to informative stimuli (i.e. sparsening), reducing the number of neurons responding to more than one informative stimulus. This reduction in the similarity of the neuronal responses to informative stimuli may allow regions of the brain that receive visual information from V1 to generate different stimulus-specific behaviors. However, the neurobiological mechanisms underlying this learning effect and how it impacts downstream brain regions involved in decision-making are unknown. To answer these questions, we propose three complementary research objectives.

Our first objective is to identify classes of V1 neurons involved in stimulus-response sparsening. We will carry out large-scale recordings of neuronal activity combined with our laboratory's powerful genetic analysis technique. By probing the gene expression of thousands of recorded neurons, we can classify them based on their expression profiles, thus allowing for the correlation of activity in specific neuronal classes to stimulus sparsening effects.

Our second objective is to understand the fine-scale temporal dynamics of stimulus-response sparsening across all layers of the V1 brain circuit. Our previous study recorded neuronal activity using microscopy, which is limited in temporal resolution and biased to superficial cortical layers. Thus, we will take advantage of state-of-the-art electrophysiological probes, partly developed in our laboratory, that can sample the activity of large populations of neurons with millisecond resolution and across the entire depth of V1. These experiments will allow us to understand the time course of the sparsening effect and whether it is specific to layers of cortex that send information to downstream brain regions.

Our third objective is to understand the relationship between the learning effects we observe in V1 and activity in downstream brain regions important for making decisions and initiating behaviors. Our recent work showed that learning causes V1 responses to informative stimuli to become more dissimilar, and we hypothesize that this effect allows downstream regions to initiate different behaviors in response to them. We will directly test this by carrying out paired simultaneous electrophysiological recordings in V1 and brain regions implicated in decision-making, to determine whether V1 sparsening correlates with changes in activity in these downstream brain regions.

Completing these objectives will provide crucial insights into the neuronal circuits and brain dynamics underlying learned-evoked effects on V1 and its importance for the planning and generation of visually guided behaviors.

Learning a new task requires changes to brain circuits, which alters the way neuronal activity responds to sensory inputs and produces motor outputs. We recently showed that one rule could explain learning effects in V1: responses to informative stimuli are sparsened, resulting in an orthogonalization of their representations. This orthogonalization should allow downstream regions to produce differentiated responses to informative stimuli, thus enabling stimulus-specific behaviors.
However, while we have identified a robust effect of task learning on V1, several questions remain. First, how do local neuronal circuits implement sparsening? Second, what are the temporal dynamics of sparsening and its layer specificity? Third, what importance does sparsening have for downstream brain regions implicated in decision making? We will investigate these questions using a combination of 2-photon imaging, in situ transcriptomics, and Neuropixels 2.0 probe recordings.
Objective 1: We will determine how specific neuron classes contribute to stimulus sparsening by following large-scale 2-photon imaging recordings with in situ transcriptomics to identify the recorded neuron classes post-hoc.
Objective 2: We will determine how sparsening temporally evolves and propagates across the layers of V1 by recording neurons with millisecond resolution using Neuropixels probes
Objective 3: We hypothesize that orthogonalizing stimulus representations improves downstream regions' ability to produce different behaviors. We will test this by carrying out simultaneous Neuropixels probe recordings in V1 and multiple decision-related brain regions to investigate the effects of stimulus-response sparsening on their activity.
Completing these objectives will answer a long-standing question in neuroscience and may have implications not just for neurobiology but also for artificial intelligence, where optimizing stimulus differentiation is critical for developing machine learning systems.