Emergent perceptual biases from state-space geometry in spiking recurrent neural networks trained to discriminate time intervals

Abstract

The contraction bias has been extensively studied in psychophysics, but only recently has its origin begun to be investigated with neural activity recordings. Leveraging the fact that it is currently possible to train recurrent networks of spiking neurons to perform cognitive tasks, we investigated the causes of the contraction bias and how behavior relates to population firing activity in networks trained to discriminate temporal intervals. A novel normative model containing potential sources of the bias guided us in investigating how it emerged in the networks. Various geometric features of state-space trajectories encoded compressed representations of the durations of the first interval, which were modulated by sensory history. Furthermore, these compressions conveyed a Bayesian estimate of the durations, thus relating activity and behavior. We conjecture that this occurs in areas of the brain where information about the current and preceding stimuli converges, and that it holds generally in delayed comparison tasks.

Emergent computations in trained artificial neural networks and real brains

“…However, the trained networks do comply with two important biological features (network connectivity and neuron type) and both the network performance and neural activity exhibit many features found in behaving animals. This was recently seen in sRNNs trained to a delayed match-to-sample task [22] and in sRNNs trained to discriminate on the duration of temporal intervals [23].…”

“…To obtain more biologically plausible models of credit assignment, researchers have proceeded in several directions. One approach attempts to bring together the most salient biological features, e.g., spiking neurons and recurrent connectivity [3,5,22,23,[72][73][74]. However, most of these studies neglect the plausibility of the learning paradigm and the learning rule.…”

“…However, most of these studies neglect the plausibility of the learning paradigm and the learning rule. For example, recursive least squares, used by [3,23,72] to train sRNNs, is not a local rule, as required by biology. Also backpropagation-through-time, used in [5,22], lacks biological realism [43].…”

“…In short, there is currently no credit assignment model that simultaneously considers sRNNs, trained with reward-based algorithms and biological plausible learning rules. Note, however, that the optimized sRNNs do bring together two essential properties of cortical networks, and can be used to investigate how they solve cognitive tasks [22,23].…”

“…This is unsatisfactory, as it does not allow identifying how a neural network could solve these tasks; the models thus constructed only reflect the researcher's intuitions about how the tasks could be performed. In contrast, trained networks provide us with a valuable tool to investigate mechanisms that networks could use to perform the tasks [6,7,13,[16][17][18][19][20][21][22][23]. This approach enables to formulate hypotheses, draw predictions and devise analysis methods that could help to guide experiments.…”