Investigating The Role of Noise Correlations in Learning

Overview

A fundamental problem in neuroscience is how the brain computes with noisy neurons. An advantage of population codes is that downstream neurons can pool across multiple neurons to reduce the impact of noise. However, this benefit depends on the noise associated with each neuron being independent. Noise correlations refer to the covariance of noise between pairs of neurons, and such correlations can limit the advantages gained from pooling across large neural populations. Indeed, a large body of theoretical work argues that positive noise correlations between similarly tuned neurons reduce the representational capacity of neural populations and are thus detrimental to neural computation. Despite this apparent disadvantage, such noise correlations are observed across many different brain regions, persist even in well-trained subjects, and are dynamically altered in complex tasks. The investigators have advanced the hypothesis that noise correlations may be a neural mechanism for reducing the dimensionality of learning problems. The viability of this hypothesis has been demonstrated in neural network simulations where noise correlations, when embedded in populations with fixed signal-to-noise ratio, enhance the speed and robustness of learning. Here the investigators aim to empirically test this hypothesis, using a combination of computational modeling, fMRI and pupillometry. Establishing a link between noise correlations and learning would open the door to an investigation into how brains navigate a tradeoff between representational capacity and the speed of learning.

Mammalian brains represent information using distributed population codes which provide a number of advantages from robustness to high representational capacity. However, for downstream readout neurons such codes pose formidable high-dimensional learning problems as a very large number of synaptic connections must be adjusted during learning in search of a suitable readout. Our recent theoretical work hypothesized that these high-dimensional learning problems can be simplified by inductive biases implemented through stimulus-independent noise correlations which express the degree to which a pair of neurons covary in their trial-to-trial fluctuations. While noise correlations have traditionally been viewed as providing constraints on representational capacity our recent work demonstrates that they simultaneously constrain readout learning. In some biologically relevant cases, they could theoretically speed learning by shaping the geometry of the underlying neural space to focus the gradient of learning onto task-relevant dimensions. However, this hypothesized role of noise correlations in shaping learning has not yet been empirically tested. Here the investigators elaborate an experimental framework to test the predicted role of noise correlations, as measured through covariation in fMRI multi-voxel BOLD activity patterns for a given stimulus, on learning in both familiar and novel contexts. In familiar contexts, useful noise correlations may be induced by top-down inputs from the prefrontal cortex that signal relevant task dimensions. Thus, the strength of noise correlations in task-relevant dimensions would predict faster learning about task-relevant features. On the other hand, in novel contexts when the relevant task dimensions are unknown, noise correlations may force gradients onto task-irrelevant dimensions and thus impair learning. Therefore, suppressing noise correlations, which might be achieved through neuromodulatory signaling, may speed learning by reducing bias early during learning or after a change in the task-relevant stimulus. Across our Aims, the investigators develop a plan to test the most basic predictions of our computational model using fMRI to characterize the geometry of noise correlations and pupillometry as a proxy for neuromodulatory signaling in human subjects. The planned research will provide the first empirical test of the role of noise correlations in learning.