Abstract
Traditional models of decision-making, such as the Drift Diffusion Model (DDM), characterize evidence accumulation as a single scalar variable. However, neural recordings reveal a dichotomy: "integrator" neurons in monkeys exhibit ramping firing rates, while "sequential" neurons in rodents fire in a temporal chain. This paper proposes a unified mathematical framework based on the Laplace transform. By maintaining a distributed representation of the distance to a decision bound, this model reconciles these disparate neural signatures, suggesting a canonical computation for time, space, and evidence in the brain.
Introduction
For decades, the Drift Diffusion Model has served as the "Gold Standard" for understanding simple perceptual decisions. The DDM posits that the brain accumulates evidence as a running total until a threshold is reached. While this model successfully explains behavioral data, it struggles to account for the diversity of neural responses observed across species. In primates, neurons in the lateral intraparietal (LIP) area show firing rates that ramp up over time, consistent with a scalar integrator. Conversely, in rodents performing similar tasks, posterior parietal cortex (PPC) neurons exhibit sequential activation, firing at specific points along a decision trajectory. This paper introduces a Laplace-domain framework to bridge this gap, proposing that these two patterns are different manifestations of the same underlying computation.
Methods
The proposed framework extends a model previously used to explain neural representations of time and space. The computation involves two primary stages:
- The Laplace Transform ($F(s)$): A population of leaky integrator neurons maintains the Laplace transform of the distance to a decision boundary. Rather than a single scalar, multiple neurons track evidence using different rate constants ($s$). A simple differential equation updates this transform in real-time based on incoming evidence $\alpha(t)$.
- The Inverse Transform ($\tilde{f}(x^*)$): A second population performs an approximate inversion of the transform using the Post approximation. This step maps the abstract Laplace domain back into a "decision axis," creating receptive fields that fire at specific distances from the decision bound.
This distributed representation allows the brain to maintain a robust, compressed estimate of evidence.
Results
Computational simulations demonstrate that the model qualitatively replicates key experimental findings from both monkey and rodent studies.
As shown in the figure above, the $F(s)$ units (panel a) mimic the coherence-dependent ramping and response-alignment characteristic of monkey LIP neurons. Crucially, the model predicts that this ramping is exponential rather than linear, a subtle distinction that may have been obscured by noise in previous datasets. In panel b, the $\tilde{f}(x^*)$ units produce a sequence of activations that tile the decision axis, matching the sequential firing observed in rodent virtual reality navigation tasks. This "hook" pattern in the heatmaps indicates that neurons are activated in a specific order as the animal approaches a decision.
Discussion
The Laplace-domain framework suggests a "unified field theory" for cognitive representations. By using the same machinery for time (time cells), space (place cells), and evidence (decision cells), the brain may employ a canonical computation across various cortical regions.
One significant implication is the prediction of Weber-Fechner scaling: neurons should be distributed logarithmically, providing higher resolution for evidence levels near the decision bound. While the model aligns well with existing data, future work is needed to experimentally verify the exponential nature of ramping and to identify the specific synaptic architectures that implement the inverse transform. Scaling this framework to multi-alternative decisions and more complex hierarchical reasoning remains an open challenge.
Conclusion
By shifting from a scalar to a distributed Laplace-domain representation, we can unify the seemingly contradictory observations of integrator and sequential neurons. This framework not only provides a more biologically plausible account of decision-making but also points toward a fundamental principle of neural organization.
Summary updated