Discrete stepping and nonlinear ramping dynamics underlie spiking responses of LIP neurons during decision-making

Abstract

Neurons in macaque area LIP exhibit gradual ramping in their trial-averaged spike responses during sensory decision-making. However, recent work has sparked debate over whether singletrial LIP spike trains are better described by discrete "stepping" or continuous drift-diffusion ("ramping") dynamics. Here we address this controversy using powerful model-based analyses of LIP spike responses. We extended latent dynamical models of LIP spike trains to incorporate non-Poisson spiking, baseline firing rates, and various nonlinear relationships between the latent variable and firing rate. Moreover, we used advanced model-comparison methods, including cross-validation and a fully Bayesian information criterion, to evaluate and compare models. These analyses revealed that when non-Poisson spiking was incorporated into existing stepping and ramping models, a majority of neurons remained better described by stepping dynamics, even when conditioning on evidence level or choice. However, an extended ramping model with a non-zero baseline and a compressive output nonlinearity accounted for roughly as many neurons as the stepping model. The latent dynamics inferred under this model exhibited high diffusion variance for many neurons, making them qualitatively different than slowly-evolving continuous dynamics. These findings generalized to alternative tasks, suggesting that a robust fraction of LIP neurons are better described by each model class.

Efficient sampling and noisy decisions

“…stochastic resonance phenomena), our formal analysis provides a unique view of the necessity of noise to optimize decision behavior in capacity-limited systems that rely on information coding based on binary samples. Interestingly, this notion appears to be consistent with the recent controversial finding that dynamics of LIP neurons likely reflect binary (discrete) coding states to guide decision behavior(18, 27). Based on this potential link between our and their work, our theoretical framework generates testable predictions that could be investigated in future neurophysiological work.…”

“…This result reveals three important aspects of neural function and decision behavior: First, it makes explicit why a system that evolved to code information using our proposed coding scheme (i.e., ones and zeros codes or binary-state states(18, 27)) must be necessarily noisy. That is, we do not attribute the randomness of peoples’ responses to a particular set of stimuli or decision problem to unavoidable randomness of the hardware used to process the information.…”

“…For instance, noise distribution in neural circuits should dynamically adapt according to the prior distribution of inputs and goals of the organism. Consequently, the rate of “step-like” coding in single neurons(27) should also be dynamically adjusted (perhaps optimally) to statistical regularities and behavioral goals.…”

The interaction between elapsed time and decision accuracy differs between humans and rats

“…It remains unclear to us whether and how other classes of models (19,37,(42)(43)(44)(45)) might account for both species' data, but this is an interesting question for future research. It is hoped that exploring this question will lead to model refinements and help distinguish among alternative theories (46).…”

Evidence Accumulation in a Laplace Domain Decision Space

“…It is also possible that work arguing for discrete steps in activation has not considered the appropriate family of continuously-evolving activation. Indeed, recent work has begun to explore non-linear models for activation as a function of time (Zoltowski, Latimer, Yates, Huk, & Pillow, 2018) and has found relatively nuanced results.…”