Flexible multitask computation in recurrent networks utilizes shared dynamical motifs

Abstract

Flexible computation is a hallmark of intelligent behavior. Yet, little is known about how neural networks contextually reconfigure for different computations. Humans are able to perform a new task without extensive training, presumably through the composition of elementary processes that were previously learned. Cognitive scientists have long hypothesized the possibility of a compositional neural code, where complex neural computations are made up of constituent components; however, the neural substrate underlying this structure remains elusive in biological and artificial neural networks. Here we identified an algorithmic neural substrate for compositional computation through the study of multitasking artificial recurrent neural networks. Dynamical systems analyses of networks revealed learned computational strategies that mirrored the modular subtask structure of the task-set used for training. Dynamical motifs such as attractors, decision boundaries and rotations were reused across different task computations. For example, tasks that required memory of a continuous circular variable repurposed the same ring attractor. We show that dynamical motifs are implemented by clusters of units and are reused across different contexts, allowing for flexibility and generalization of previously learned computation. Lesioning these clusters resulted in modular effects on network performance: a lesion that destroyed one dynamical motif only minimally perturbed the structure of other dynamical motifs. Finally, modular dynamical motifs could be reconfigured for fast transfer learning. After slow initial learning of dynamical motifs, a subsequent faster stage of learning reconfigured motifs to perform novel tasks. This work contributes to a more fundamental understanding of compositional computation underlying flexible general intelligence in neural systems. We present a conceptual framework that establishes dynamical motifs as a fundamental unit of computation, intermediate between the neuron and the network. As more whole brain imaging studies record neural activity from multiple specialized systems simultaneously, the framework of dynamical motifs will guide questions about specialization and generalization across brain regions.

Disentangling with Biological Constraints: A Theory of Functional Cell Types

“…Indeed in a companion paper, we use the same constraints, along with formalising structure/path-integration using group and representations theory, to mathematically understand why grid cells look like grid cells (Dorrell et al, 2022). Similarly, our current understanding is limited to the optimal solution for factorised representations, but we anticipate similar ideas will be applicable to neural dynamics (Driscoll et al, 2022).…”

“…Non-negativity. Recent work on multitask learning in recurrent neural networks (RNNs) (Yang et al, 2019;Driscoll et al, 2022) demonstrated that neural populations, with a nonnegative activation function, partition themselves into task specific modules (Driscoll et al, 2022). Non-negativity is also important in obtaining hexagonal, not square, grid cells (Dordek et al, 2016;Whittington et al, 2021b;Sorscher et al, 2019;.…”

Emergent neural dynamics and geometry for generalization in a transitive inference task

“…While a non-rotational solution may also be possible, constraints relevant to neurobiology (e.g. metabolic efficiency or the dimensionality of outputs) may favor rotational solutions [103, 128, 117, 129]. Notably, we used the generalized idea of rotational transformation during the delay to identify an additional neural prediction distinguishing RNN variants (mean angle change, Fig.…”

“…Given these clues, we next sought to identify the underlying implementation explicitly. Importantly, prior work has found that, in trained RNNs, analysis of network-level neural dynamics with respect to fixed points (FPs) can identify dynamical components that have specific functional roles [92, 115, 116, 117] and are jointly sufficient to perform cognitive tasks.…”

Generalization in Sensorimotor Networks Configured with Natural Language Instructions

“…On each trial, the model receives natural language instructions for the present task embedded through a transformer architecture pre-trained on one of several natural language processing objectives. Both the type of tasks we use and the neural network modeling of such tasks have a rich history in the experimental and computational neuroscience literature respectively [19, 20, 21, 22, 23, 24, 25].…”