Coordinated drift of receptive fields during noisy representation learning

Abstract

Long-term memories and learned behavior are conventionally associated with stable neuronal representations. However, recent experiments showed that neural population codes in many brain areas continuously change even when animals have fully learned and stably perform their tasks. This representational "drift" naturally leads to questions about its causes, dynamics, and functions. Here, we explore the hypothesis that neural representations optimize a representational objective with a degenerate solution space, and noisy synaptic updates drive the network to explore this (near-)optimal space causing representational drift. We illustrate this idea in simple, biologically plausible Hebbian/anti-Hebbian network models of representation learning, which optimize similarity matching objectives, and, when neural outputs are constrained to be nonnegative, learn localized receptive fields (RFs) that tile the stimulus manifold. We find that the drifting RFs of individual neurons can be characterized by a coordinated random walk, with the effective diffusion constants depending on various parameters such as learning rate, noise amplitude, and input statistics. Despite such drift, the representational similarity of population codes is stable over time. Our model recapitulates recent experimental observations in hippocampus and posterior parietal cortex, and makes testable predictions that can be probed in future experiments.

The geometry of representational drift in natural and artificial neural networks

“…Representational drift has also been studied at the computational/theoretical levels [26][27][28][29]. In particular, Ref.…”

Network mechanisms underlying representational drift in area CA1 of hippocampus

“…Other computational modeling work has also proposed that RD may arise due to plasticity-driven perturbations [44][45][46] . In 44 the authors studied a network with prescribed synaptic weights which allow for the storage of a large number of sequential patterns.…”

Long-term stability of single neuron activity in the motor system

“…While the observed slow drift in neural and behavioral space in expert animals suggests that the changes in neural circuits occur in directions of neural state space that affect motor output, it remains to be seen whether this behavioral drift constitutes a learning process that optimizes a utility function such as energy expenditure (Srinivasan & Ruina, 2006) or magnitude of the control signal (Todorov & Jordan, 2002). Alternatively, it could reflect a random walk in a degenerate motor space that preserves task performance (Chaisanguanthum et al, 2014; Qin et al, 2021). Previous work has also suggested that motor variability could be explicitly modulated to balance exploration and exploitation as a function of past performance and task uncertainty (Dhawale et al, 2019).…”

“…Two main theories have been put forth to explain the apparent paradox of stable memories in plastic circuits. In the commonly held view that motor control is governed by low-dimensional dynamics (Gallego et al, 2017; Jensen et al, 2021; Shenoy et al, 2013; Vyas et al, 2020), the paradox can be resolved by having a degenerate subspace in which neural activity can change without affecting behavior (Rokni et al, 2007) or task performance (Qin et al, 2021). While this would do away with the requirement for stable activity at the level of single neurons (Figure 1A) (Clopath et al, 2017), it requires any drift in population activity to occur exclusively in the degenerate subspace.…”

“…A different way to maintain stable motor output is by constraining the changes in neural circuits such that they do not affect single neuron activity associated with already established behaviors (Duncker et al, 2020; Kao, Jensen, et al, 2021; Qin et al, 2021). In this case, the activity patterns of individual neurons locked to the behavior would remain constant or highly similar over time (Figure 1A) (Clopath et al, 2017).…”