Invariant neural subspaces maintained by feedback modulation

Laura Naumann, Joram Keijser, Henning Sprekeler

TL;DR

This paper proposes a novel mechanism for achieving context-invariant sensory perception using feedback modulation in neural networks. They demonstrate that by dynamically adjusting feedforward connections through slow, diffuse feedback, networks can maintain stable representations of stimuli despite changing environmental contexts, a feat not achievable by feedforward-only models. This invariance emerges at the population level, not the individual neuron level, highlighting the importance of population-level analysis.

What Problem Does This Paper Address?

Our sensory systems are remarkably adept at recognizing objects and sounds regardless of variations in lighting, background noise, or other contextual factors. While traditional models attribute this "context invariance" to hierarchical processing in feedforward neural networks that extract increasingly complex features, these models often neglect the abundant feedback connections in the brain. This paper investigates how top-down feedback, rather than solely feedforward processing, can establish and maintain context-invariant representations, particularly in dynamic environments where the context itself changes over time.

Methods

The authors employ a series of increasingly complex computational models to explore their hypothesis. They begin with a simple linear feedforward network modulated by a recurrent neural network (LSTM) acting as a "modulator." This modulator learns to adjust the feedforward weights based on sensory input and network output to solve a dynamic blind source separation task. They then extend this model to incorporate:

  • Slow feedback modulation: By introducing a temporal filter, they investigate how the timescale of feedback affects performance.
  • Spatially diffuse feedback: They model feedback as a spatially broad gain modulation applied to a population of neurons, mimicking neuromodulatory signals.
  • Hierarchical Dalean network: This model incorporates biologically plausible constraints like positive firing rates and Dale's principle (excitatory and inhibitory neurons). Feedback here acts as a gain modulation on a higher-level population.

The core task used is a dynamic blind source separation problem, where the network must disentangle mixed signals (e.g., voices) whose mixing proportions (the context) change over time. Performance is evaluated using "signal clarity," a measure of how well the separated sources match the original ones.

Datasets

The paper does not use external datasets. Instead, it generates synthetic data for training and testing its models. The primary synthetic data consists of:

  • Source signals: Compositions of sine waves, and in some experiments, sawtooth and square waves.
  • Contexts: Time-varying linear mixing matrices that define how the source signals are combined to create sensory stimuli. These contexts are generated by drawing random weights from a uniform distribution.

The models are trained on a set of contexts and then tested on a separate, independent set of contexts to assess generalization.

Key Insights and Ideas

  • Feedback as the key to dynamic invariance: The study strongly suggests that feedback, not just feedforward processing, is crucial for achieving context invariance in dynamic environments.
  • Invariance emerges at the population level: Individual neurons show context-dependent responses, but the population's activity, when viewed in a specific subspace, remains invariant to context changes.
  • Dynamic reorientation of neural manifolds: Feedback modulation doesn't just "correct" errors; it actively reorients the entire population's activity manifold to maintain the invariant subspace.
  • Slow and diffuse feedback is sufficient: The mechanism works even with slow and spatially diffuse feedback signals, aligning with biological neuromodulatory systems.
  • Nonlinear context representation in modulatory units: While the higher-level population represents context linearly, the modulatory units themselves encode context in a nonlinear fashion, which is crucial for the task.

Key Results

  • The modulated feedforward network successfully performed dynamic blind source separation, achieving high "signal clarity" even when contexts changed.
  • Performance degraded only when the feedback modulation timescale was comparable to the context change timescale.
  • Spatially diffuse feedback signals were sufficient for robust performance, with only a few distinct feedback signals needed.
  • Individual neurons showed context-dependent activity, but linear decoding from the modulated population could accurately recover the sources across contexts.
  • A key experiment showed that when feedback was frozen after a context change, the population representation shifted, but reintroducing feedback reoriented the representation into a new invariant subspace.
  • In a hierarchical Dalean network, feedback-driven gain modulation also enabled context invariance at the population level, with individual neurons remaining context-dependent.
  • Decoding of context from modulatory units revealed a nonlinear representation, which was better captured by nonlinear decoding methods.

Limitations and Future Work

  • Simplistic stimuli: The use of sine waves as stimuli is a simplification. Generalizing to complex natural stimuli like speech or visual scenes presents significant computational challenges, especially for training recurrent networks with long temporal dependencies.
  • Training methodology: While the authors used supervised learning to train the modulator, they suggest that other methods like reinforcement learning or unsupervised priors could also be used.
  • Biological plausibility of computation: The specific computation of feedback signals by the modulator was implemented using a standard LSTM, which is not necessarily biologically plausible. Future work could explore how these signals are computed and learned by more biologically realistic mechanisms.
  • Role of feedforward vs. feedback: The paper focuses on feedback but acknowledges that feedforward and feedback processes likely work in concert. Further studies are needed to delineate their relative contributions to different types of invariance.
  • Dynamics of feedback: The authors highlight that feedback mechanisms take time to establish invariance. Investigating these dynamics further and comparing them to feedforward network dynamics is a promising avenue.

Summary updated

Abstract

Sensory systems reliably process incoming stimuli in spite of changes in context. Most recent models accredit this context invariance to an extraction of increasingly complex sensory features in hierarchical feedforward networks. Here, we study how context-invariant representations can be established by feedback rather than feedforward processing. We show that feedforward neural networks modulated by feedback can dynamically generate invariant sensory representations. The required feedback can be implemented as a slow and spatially diffuse gain modulation. The invariance is not present on the level of individual neurons, but emerges only on the population level. Mechanistically, the feedback modulation dynamically reorients the manifold of neural activity and thereby maintains an invariant neural subspace in spite of contextual variations. Our results highlight the importance of population-level analyses for understanding the role of feedback in flexible sensory processing.
BibTeX
@inproceedings{naumann2022invariant,
  author = {Laura Naumann and Joram Keijser and Henning Sprekeler},
  title = {{Invariant neural subspaces maintained by feedback modulation}},
  booktitle = {eLife},
  year = {2022},
  doi = {10.7554/elife.76096},
}
Figure