Resonant Learning in Scale-free Networks

Samuel Goldman, Maximino Aldana, Philippe CluzelbioRxiv 2021 · DOI

TL;DR

This paper introduces "resonant learning," a novel strategy where oscillating the "hub" (most connected node) of a scale-free network allows it to learn new behaviors much faster than traditional methods. This approach leverages the timing of the hub's oscillations to induce distinct network states, enabling the network to learn multiple tasks efficiently without needing complex modular structures.

What Problem Does This Paper Address?

Understanding how complex networks, whether biological (like gene regulatory networks) or artificial, learn and adapt to new behaviors is a fundamental challenge. While scale-free network topologies (where a few "hubs" connect to many nodes) are common, and recent experiments show these hubs can oscillate, the role of these oscillations in network learning remains unclear. The paper aims to bridge this gap by investigating if and how periodic oscillations of network hubs can facilitate learning new dynamical behaviors, particularly in scale-free networks.

Methods

The researchers employed Boolean threshold networks (RBTN) with scale-free topologies as their model system. These networks are composed of nodes that can be either "on" or "off," and their state transitions are governed by a simple threshold rule based on incoming connections.

Key methodological aspects include:

  • Forced Hub Oscillation: The most connected node (the hub) was externally forced to oscillate between "on" and "off" states with a specific period (T).
  • Attractor Landscape Analysis: The study analyzed how these forced oscillations altered the network's "attractor cycles" – stable patterns of network states the system eventually settles into. They quantified the difference between these new attractors and "ground-state" attractors (where the hub is fixed) using a "height" metric.
  • Evolutionary Learning Algorithm: A genetic algorithm was used to train networks. Networks were mutated and selected based on how well their output node's time series matched a predefined target function.
  • Resonant Learning Definition: The paper defines "resonant learning" as the phenomenon where the network's ability to learn new dynamical behaviors is dependent on the chosen period of the hub's oscillations.
  • Time-Division Multiplexing: The researchers tested the network's ability to learn multiple target functions simultaneously by applying different input periods to the hub.

Datasets

Not discussed in the paper. The study uses simulated networks and generated target functions for training and evaluation.

Key Insights and Ideas

  • Hub Oscillations as a Learning Mechanism: The most striking insight is that forcing the hub node to oscillate can fundamentally change the network's dynamics, creating new, robust attractor cycles. This suggests that the timing of a few key nodes can orchestrate complex network-level behaviors.
  • Period-Dependent Learning: The specific period of the hub's oscillation directly influences the resulting network attractors. Different periods lead to distinct attractors, which in turn can be trained to learn different target functions. This is the core of "resonant learning."
  • Faster Learning: Networks employing resonant learning converge to target behaviors significantly faster (over an order of magnitude) than standard learning procedures or networks with a fixed hub state.
  • Learning Multiple Tasks: The network can learn multiple distinct target functions by simply changing the input period to the hub. This "time-division multiplexing" capability is a powerful demonstration of the hub's control.
  • Timescale Over Pattern: The timescale of the hub's oscillation is more critical for learning than the specific pattern of the oscillation (e.g., square wave vs. repeated pattern). White noise, lacking a defined timescale, leads to degraded learning.
  • Alternative to Modularity: Resonant learning offers an alternative design principle for learning multiple tasks, distinct from the previously understood need for modular network architectures.

Key Results

  • Distinct hub oscillation periods induce robust and distinct attractor cycles, with lengths dependent on the oscillation period.
  • Networks can learn multiple target functions by sequentially changing the hub's oscillation period.
  • Resonant learning leads to convergence towards target behaviors over an order of magnitude faster than standard learning procedures.
  • Networks can learn up to seven distinct target functions by varying the hub's oscillation period.
  • Learning performance is significantly degraded when using white noise input functions lacking a defined timescale, compared to periodic inputs (square waves or repeated patterns).
  • Networks with resonant learning outperform networks with a fixed hub state by a significant margin, even when the fixed hub is allowed to update dynamically.

Limitations and Future Work

  • Model Simplification: The study uses idealized Boolean threshold networks. Future work could explore how these principles apply to more complex, continuous-valued network models.
  • Reservoir Computing Integration: The authors suggest investigating how oscillating hubs could be integrated into existing machine learning frameworks like reservoir computing, particularly with scale-free topologies.
  • Biological Relevance: While inspired by biological systems, the direct mapping of these findings to specific biological mechanisms requires further research. The authors posit that frequency-based coordination might be a general mechanism in real biological systems.
  • Generalizability: While the study demonstrates learning multiple tasks, further exploration of the limits and nuances of learning complex, long, or highly varied target functions is warranted.

Summary updated

Abstract

ABSTRACT Over the last decades, analyses of the connectivity of large biological and artificial networks have identified a common scale-free topology, where few of the network elements, called hubs, control many other network elements. In monitoring the dynamics of networks’ hubs, recent experiments have revealed that they can show behaviors oscillating between ‘on’ and ‘off’ states of activation. Prompted by these observations, we ask whether the existence of oscillatory hubs states could contribute to the emergence of specific network dynamical behaviors. Here, we use Boolean threshold networks with scale-free architecture as representative models to demonstrate how periodic activation of the network hub can provide a network-level advantage in learning specific new dynamical behaviors. First, we find that hub oscillations with distinct periods can induce robust and distinct attractors whose lengths depend upon the hub oscillation period. Second, we determine that a given network can exhibit series of different attractors when we sequentially change the period of hub pulses. Using rounds of evolution and selection, these different attractors could independently learn distinct target functions. We term this network-based learning strategy resonant learning , as the emergence of new learned dynamical behaviors depends on the choice of the period of the hub oscillations. Finally, we find that resonant learning leads to convergence towards target behaviors over an order of magnitude faster than standard learning procedures. While it is already known that modular network architecture contributes to learning separate tasks, our results reveal an alternative design principle based on forced oscillations of the network hub. SIGNIFICANCE Large networks of interconnected components such as genes or machines can coordinate complex behavioral dyamics. One outstanding question has been to identify the design principles that allow networks to learn new behaviors. Here, in simulating learning cycles, we randomly modify the interactions between components and select networks that exhibit a desired behavior. Surprisingly, we find that networks can learn new behaviors faster when the state of the most connected network component is forced to oscillate during learning. Remarkably, using distinct periods of oscillations allows a given network to learn distinct behaviors. While it is known that modular network architecture contributes to learning separate tasks, our results reveal an alternative design principle for which modules are not needed.
BibTeX
@misc{goldman2021resonant,
  author = {Samuel Goldman and Maximino Aldana and Philippe Cluzel},
  title = {{Resonant Learning in Scale-free Networks}},
  year = {2021},
  doi = {10.1101/2021.11.10.468065},
  howpublished = {bioRxiv},
}

Loading related content...

Figure