Single-unit activations confer inductive biases for emergent circuit solutions to cognitive tasks DOI Creative Commons
Pavel Tolmachev, Tatiana A. Engel

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 24, 2024

Trained recurrent neural networks (RNNs) have become the leading framework for modeling dynamics in brain, owing to their capacity mimic how population-level computations arise from interactions among many units with heterogeneous responses. RNN are commonly modeled using various nonlinear activation functions, assuming these architectural differences do not affect emerging task solutions. Contrary this view, we show that single-unit functions confer inductive biases influence geometry of population trajectories, selectivity, and fixed point configurations. Using a model distillation approach, find representations reflect qualitatively distinct circuit solutions cognitive tasks RNNs different disparate generalization behavior on out-of-distribution inputs. Our results seemingly minor provide strong solutions, raising question about which architectures better align mechanisms execution biological networks.

Language: Английский

Timescales of learning in prefrontal cortex DOI
Jacob A. Miller, Christos Constantinidis

Nature reviews. Neuroscience, Journal Year: 2024, Volume and Issue: 25(9), P. 597 - 610

Published: June 27, 2024

Language: Английский

Citations

4
Predictive routing emerges from self-supervised stochastic neural plasticity DOI Creative Commons

Hamed Nejat,

Jason Sherfey, André M. Bastos

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 31, 2024

Abstract Neurophysiology studies propose that predictive coding is implemented via alpha/beta (8-30 Hz) rhythms prepare specific pathways to process predicted inputs. This leads a state of relative inhibition, reducing feedforward gamma (40-90 and spiking predictable We refer this model as routing. It currently unclear which circuit mechanisms implement push-pull interaction between rhythms. To explore how routing implemented, we developed self-supervised learning algorithm call generalized Stochastic Delta Rule (gSDR). was necessary develop rule because manual tuning parameters (frequently used in computational modeling) inefficient search through non-linear parameter space defines neuronal emerge interact. gSDR train biophysical neural circuits validated the on simple tasks, e.g., membrane potentials firing rates. next applied observed neurophysiology. asked reproduce shift from baseline oscillatory dynamics (∼<20Hz) stimulus induced (∼40-90Hz) recorded macaque monkey visual cortex. gamma-band oscillation during stimulation emerged by self-modulation synaptic weights gSDR. further showed gamma-beta interactions implied could stochastic modulation both local inhibitory circuitry well top-down modulatory inputs network. summarize, based series objectives. succeeded implementing these revealed neuron underlying are processing tasks systems cognitive neuroscience. Significant Statement study contributes advancement for modeling behavior complex specifically, models framework. Since an evolutionary does not rely model-based assumptions, it improve autonomous approaches neuroscience network research.

Language: Английский

Citations

0
Single-unit activations confer inductive biases for emergent circuit solutions to cognitive tasks DOI Creative Commons
Pavel Tolmachev, Tatiana A. Engel

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: Nov. 24, 2024

Trained recurrent neural networks (RNNs) have become the leading framework for modeling dynamics in brain, owing to their capacity mimic how population-level computations arise from interactions among many units with heterogeneous responses. RNN are commonly modeled using various nonlinear activation functions, assuming these architectural differences do not affect emerging task solutions. Contrary this view, we show that single-unit functions confer inductive biases influence geometry of population trajectories, selectivity, and fixed point configurations. Using a model distillation approach, find representations reflect qualitatively distinct circuit solutions cognitive tasks RNNs different disparate generalization behavior on out-of-distribution inputs. Our results seemingly minor provide strong solutions, raising question about which architectures better align mechanisms execution biological networks.

Language: Английский

Citations

0