Synergistic pathways of modulation enable robust task packing within neural dynamics

TL;DR

This study uses recurrent network models to probe the distinctions between two forms of contextual modulation of neural dynamics, at the level of neuronal excitability and at the level of synaptic strength, and characterize these mechanisms in terms of their functional outcomes, indicating complementarity and synergy in how these mechanisms act.

Abstract

Understanding how brain networks learn and manage multiple tasks simultaneously is of interest in both neuroscience and artificial intelligence. In this regard, a recent research thread in theoretical neuroscience has focused on how recurrent neural network models and their internal dynamics enact multi-task learning. To manage different tasks requires a mechanism to convey information about task identity or context into the model, which from a biological perspective may involve mechanisms of neuromodulation. In this study, we use recurrent network models to probe the distinctions between two forms of contextual modulation of neural dynamics, at the level of neuronal excitability and at the level of synaptic strength. We characterize these mechanisms in terms of their functional outcomes, focusing on their robustness to context ambiguity and, relatedly, their efficiency with respect to packing multiple tasks into finite size networks. We also demonstrate distinction between these mechanisms at the level of the neuronal dynamics they induce. Together, these characterizations indicate complementarity and synergy in how these mechanisms act, potentially over multiple time-scales, toward enhancing robustness of multi-task learning.

Figures (6)

• Figure 1: Memory task representation. Trials begin with fixation around the origin until time $t = T_{Fix}$. Subsequently, a stimulus (location on a circle) is presented remains visible until time $t = T_{TP}$. Following this, a variable delay period ensues. Finally, a response period occurs during which the network produces a context-dependent output.
• Figure 2: Robustness characterization. The ERNN, SRNN and SERNN mechanisms were trained on identical paradigms, with the same level of contextual ambiguity and additive disturbance, i.e. $\sigma = 1$ and $\epsilon = 1$. Both ERNN and SRNN exhibit sensitivity to context ambiguity (left) and input disturbance (right). The combined mechanisms (SERNN) achieve notable robustness gains in both settings, indicating mechanistic synergy. Plots show mean and standard deviations based on $n=150$ independently trained networks.
• Figure 3: Task packing for different mechanisms, number of neurons and tasks. Task packability as a function of the number of embedded contexts and the number of available neurons, represented by the value of the cost function after 1000 epochs. A value smaller than 5000 corresponds to a well-trained network.
• Figure 4: Transferability of the different modulation mechanisms. Networks were trained on four contexts, prior to transfer to a fifth novel context. Transfer learning trajectories for the (A) ERNN, (B) SRNN, and (C) SERNN, where the synergistic performance gain of the combined modulation is reflected in fast acquisition (in terms of the log loss) that exceeds a fully trainable vRNN. Pre- and post-transfer performance of the (D) ERNN, (E) SRNN, (F) SERNN, demonstrating the manifestation of the improved loss trajectory on performance. Pre-transfer performance is durable under all mechanisms.
• Figure 5: Dimensionality reduced visualization of network activity.(A) ERNN, (B) SRNN, (C) SERNN.