Emergent perceptual biases from state-space geometry in trained spiking recurrent neural networks. Serrano-Fernandez et al.

This dataset contains both the behavior and the single-neuron spike trains from two example recurrent neural networks of spiking LIF neurons (sRNN) performing a time interval discrimination task. In this task, two-time stimuli are presented sequentially separated by a delay period in order to be compared.

Starting from different random initializations of the network parameters, each sRNN was trained in a block of 5,000 training trials, presented in a continuous manner, without perturbing the network activity at the end of the trial.

Once the networks were trained, a series of test simulations were conducted to assess their performance in the task and to collect the attached datasets (see the different "README" documents inside the dataset for further information). Their performance exhibited a perceptual bias known as contraction bias. Due to this bias, the first stimulus is perceived as contracted towards the mean of its distribution. To shed light on this effect, we explored, in the following article, the causes of this bias and how behavior relates to population firing activity:

Serrano-Fernandez, Beiran, and Parga. "Emergent perceptual biases from state-space geometry in trained spiking recurrent neural networks." Cell Reports (2024).
DOI: https://doi.org/10.1016/j.celrep.2024.114412