Adaptive population codes for flexible visually-guided behaviors

University Of California, San Diego, La Jolla CA

Investigators

Linked publications & trials

Abstract

Summary/Abstract Traditional models of visual information processing propose that stable perceptual and mnemonic representations are supported by persistent activity in neurons that are selective for relevant stimulus features. However, there is an emerging focus on the importance of reducing redundancy in neural codes to achieve more efficient processing. For example, the visual system adapts to overall lightness levels and to frequently encountered stimuli, allowing neural codes to maintain a useful dynamic range and to expend less energy encoding expected stimuli. In addition to dynamics induced by recent stimulus history, recent work in machine learning and in animal model systems reveals that codes also drift over short and long time scales, often while preserving population-level geometric relationships between different stimulus representations. While many factors likely contribute to this set of phenomena, here we focus on the general hypothesis that dynamics induced by stimulus history and other factors support efficient yet error-tolerant codes. For example, recent work in our lab suggests that the perceptual distortions induced by adaptation (e.g. the classic motion after-effect) are offset by a flexible decoding scheme that utilizes a prior for temporal stability to compensate. Short-term dynamics, particularly during visual working memory, at least partially reflect a reduction in the complexity of representations so that only the most behaviorally relevant aspects of sensory information are retained to guide behavior. Finally, long-term dynamics, or representational drift, may reflect a refining of neural codes to support increasingly sparse information processing that is still robust to interference. In the present proposal, we build on this recent empirical and theoretical work to evaluate the functional role of neural dynamics driven by different factors across different time scales. We first test how the flexibility of Bayesian read-out rules allows the visual system to counter the effects of stimulus history to produce continuous and robust representation even in the face of internal and external noise. We then use recurrent neural networks (RNNs) with different processing constraints to show that a drive toward sparser, and more energy efficient, representations is sufficient to induce short and long-term drift in neural codes. Guided by these modelling efforts, we will evaluate the impact of drift on neural codes as measured in human visual cortex using functional magnetic resonance imaging (fMRI) and psychophysics, with a focus on testing hypotheses about the impact of drift on the dimensionality and robustness of feature-selective activation patterns. Collectively, this work will challenge traditional theories of perceptual inference, attentional selection, and working memory that are based on the notion of stable neural codes by providing novel insights into the functional role that pervasive neural dynamics play in visual information processing.