Neural circuit models for evidence accumulation through choice-selective sequences

Lindsey S. Brown, Jounhong Ryan Cho, Scott S. Bolkan, Edward H. Nieh, Manuel Schottdorf, David W. Tank, Carlos D. Brody, Ilana B. Witten, Mark S. GoldmanbioRxiv 2023 · DOI

Abstract

Decision making is traditionally thought to be mediated by neurons that accumulate evidence through persistent activity. However, recent decision-making experiments in rodents have observed neurons across the brain that fire sequentially, rather than persistently, with the subset of neurons in the sequence depending on the animal's choice. We developed two new candidate circuit models in which neurons are active sequentially and transfer evidence faithfully to the next active population. One model encodes evidence in the relative firing of two competing chains of neurons, and the other in the network location of a stereotyped pattern ("bump") of neural activity. Neural recordings from four brain regions during an evidence accumulation task revealed that different regions displayed evidence tuning consistent with different candidate models. This work provides a mechanistic explanation for how graded information may be precisely accumulated within and transferred between neural populations, and suggests that different brain regions may accumulate evidence through different circuit mechanisms.

Introduction

For decades, the dominant theory of decision-making has been the Drift Diffusion Model (DDM), where neurons accumulate evidence through "ramping" activity—persistently increasing or decreasing their firing rates. However, recent large-scale recordings in rodents performing the "Accumulating Towers" task have challenged this view. In this task, mice navigate a virtual maze and count visual cues to decide whether to turn left or right. Instead of persistent ramping, researchers observed neurons firing in choice-selective sequences: specific groups of neurons fire at distinct positions along the maze, with the sequence itself depending on the animal's eventual choice.

The accumulating towers task (A), observed neural sequences (B), and two proposed circuit models: Competing Accumulator (C) and Bump Attractor (D).

Proposed Circuit Models

To explain how a circuit can accumulate graded evidence and faithfully "hand it off" to subsequent neurons in a sequence, the authors proposed two distinct classes of models:

  1. Mutually Inhibitory Competing Chains: This model features two parallel feedforward chains (one for "left" evidence, one for "right"). Evidence is encoded in the amplitude of the firing rate (rate coding). A "position-gating" signal allows activity to flow between links in the chain as the animal moves. Mutual inhibition between the chains ensures the dominant evidence source suppresses the alternative.
  2. Position-Gated Bump Attractor: In this model, evidence is represented by the location of a "bump" of neural activity on a 2D plane. One axis represents the animal's position, while the other represents the accumulated evidence. As the animal moves and cues appear, the bump shifts along the corresponding axes. Here, evidence is encoded by which neurons are active (place coding) rather than their firing rate.

Key Insights and Experimental Predictions

Both models rely on an external "position-gating" signal, likely originating from regions like the hippocampus or basal ganglia, to coordinate the transfer of information across the network. However, they offer different computational strategies and experimental signatures:

  • Tuning Curves: The Competing Chains model predicts monotonic tuning, where a neuron's firing rate increases steadily with more evidence. Conversely, the Bump Attractor model predicts non-monotonic tuning, where a neuron is "tuned" to a specific evidence value (e.g., exactly +3 towers).
  • Perturbation Effects: The models make distinct predictions for optogenetic stimulation. In the Chains model, stimulating a neuron increases activity in all subsequent links. In the Bump model, stimulation may decrease activity in some neurons as the activity "bump" is forced to a new location.

Results and Validation

Simulations showed that both models successfully perform the evidence accumulation task, matching the psychometric performance of real mice.

Model simulations showing psychometric performance (B) and the generation of choice-selective neural sequences (C) that match experimental observations.

Analysis of neural data from multiple brain regions, including the hippocampus (HPC), primary visual cortex (V1), and dorsal medial striatum (DMS), revealed functional diversity. Notably, the hippocampus showed non-monotonic tuning, strongly supporting the Bump Attractor mechanism in that region. This suggests the hippocampus may construct a "cognitive map" of evidence. Other regions may utilize the Chains-based mechanism, which is more efficient for encoding graded values like confidence or reward.

Conclusion

This work provides a mechanistic framework for how the brain accumulates and transfers information through time and space. By moving beyond static models of decision-making, these sequential circuit models offer a bridge to the dynamic neural activity observed in modern large-scale recordings, suggesting that different brain regions may employ diverse circuit mechanisms for evidence accumulation.

Summary updated

Abstract

Decision making is traditionally thought to be mediated by neurons that accumulate evidence through persistent activity. However, recent decision-making experiments in rodents have observed neurons across the brain that fire sequentially, rather than persistently, with the subset of neurons in the sequence depending on the animal's choice. We developed two new candidate circuit models in which neurons are active sequentially and transfer evidence faithfully to the next active population. One model encodes evidence in the relative firing of two competing chains of neurons, and the other in the network location of a stereotyped pattern ("bump") of neural activity. Neural recordings from four brain regions during an evidence accumulation task revealed that different regions displayed evidence tuning consistent with different candidate models. This work provides a mechanistic explanation for how graded information may be precisely accumulated within and transferred between neural populations, and suggests that different brain regions may accumulate evidence through different circuit mechanisms.
BibTeX
@misc{brown2023neural,
  author = {Lindsey S. Brown and Jounhong Ryan Cho and Scott S. Bolkan and Edward H. Nieh and Manuel Schottdorf and David W. Tank and Carlos D. Brody and Ilana B. Witten and Mark S. Goldman},
  title = {{Neural circuit models for evidence accumulation through choice-selective sequences}},
  year = {2023},
  doi = {10.1101/2023.09.01.555612},
  howpublished = {bioRxiv},
}
Figure