Learning predictive signals within a local recurrent circuit

Toshitake Asabuki, Colleen J. Gillon, Claudia ClopathbioRxiv 2023 · DOI

TL;DR

This paper introduces a novel recurrent spiking neural network model that learns to generate prediction error signals solely from local, bottom-up sensory input, mimicking brain activity observed in mismatch responses. This work sheds light on how synaptic plasticity can enable local circuits to build internal models of their environment and signal deviations from expectations, even without explicit top-down predictions.

What Problem Does This Paper Address?

The brain is thought to predict upcoming sensory information to better process the world. When these predictions are violated, the brain generates "mismatch responses," which are thought to signal "prediction errors." While the predictive coding hypothesis explains this by proposing top-down predictions are compared with incoming sensory data, recent experiments suggest these prediction errors can arise even from local circuits processing bottom-up input alone. This raises a key question: how can local circuits, without explicit top-down signals, learn to predict and generate error signals? Existing models often rely on hierarchical structures or explicit top-down predictions, leaving a gap in explaining purely local, experience-driven prediction error generation. This paper aims to bridge this gap by demonstrating how synaptic plasticity within a local recurrent network can achieve this.

Methods

The authors developed a recurrent spiking neural network model composed of excitatory (E) and inhibitory (I) neurons. The core innovation lies in the plasticity rules applied to the synapses:

  • Excitatory Synapses: These synapses on excitatory neurons are plastic and learn to predict the network's activity based on incoming sensory events. Errors between the internally driven excitation and the neuron's output are used to modulate this plasticity, aiming to minimize prediction errors.
  • Inhibitory Synapses: These synapses on excitatory neurons are also plastic but are tuned to maintain the crucial excitatory-inhibitory (E-I) balance within the network. They do this by predicting the recurrent excitatory potential.

This dual plasticity mechanism allows the network to simultaneously learn predictive patterns and maintain stable network dynamics. The model was trained on sequences of stimuli, and its ability to generate prediction error signals was tested by introducing deviant stimuli or transitions.

Datasets

Not discussed in the paper. The paper focuses on simulated data generated by the authors' model.

Key Insights and Ideas

  • Local Prediction is Sufficient: The most striking insight is that a local recurrent circuit, through experience-driven synaptic plasticity, can learn to predict sensory input and generate prediction error signals without relying on explicit top-down predictions.
  • Biphasic Prediction Errors: The model successfully replicates the biphasic nature (negative and positive components) of mismatch negativity (MMN) observed in human EEG/MEG studies. This suggests a mechanistic explanation for this common neurophysiological signature.
  • Distinct Mechanisms for Biphasic Errors: The paper reveals that the early negative component of the prediction error arises from a disruption in E-I balance due to decreased excitatory currents, while the later positive component stems from a similar disruption caused by decreased inhibitory currents.
  • Context-Dependent Prediction: The model demonstrates that prediction errors can be context-dependent, meaning the same stimulus can elicit different responses based on the preceding context, mirroring experimental findings in visual and auditory cortices.
  • Synaptic Plasticity as the Driver: The study strongly emphasizes the role of synaptic plasticity in shaping prediction errors, showing how learning strengthens connections to predict regularities and how deviations from these learned regularities trigger error signals.

Key Results

  • The trained recurrent network generated prediction error signals that mimicked experimental findings, including a biphasic waveform similar to MMN.
  • The model showed context-specific prediction errors, where responses to expected stimuli were suppressed and unexpected stimuli were enhanced, depending on the preceding context.
  • Analysis of recurrent currents revealed distinct mechanisms for the early (negative) and late (positive) phases of the biphasic prediction error, both involving disruptions in the E-I balance but driven by different current changes.

Limitations and Future Work

  • Global Structure Learning: The current model learns local transition statistics between stimuli but cannot learn predictions over the global structure of stimulus chunks (e.g., non-Markovian statistics).
  • Hierarchical Networks: The authors suggest that extending the model to a hierarchical structure, where lower levels handle local transitions and higher levels capture abstract dynamics, could address the limitation of global structure learning.
  • Neuron Compartmentalization: The paper proposes that implementing neurons with distinct compartments (dendritic and somatic) could provide a more biologically plausible way to encode predicted states and prediction errors within single neurons.
  • Spatial vs. Temporal Information: While the model excels at temporal prediction errors, the question of how prediction error signals encode both spatial and temporal information remains open.

Summary updated

Abstract

Abstract The predictive coding hypothesis proposes that top-down predictions are compared with incoming bottom-up sensory information, with prediction errors signaling the discrepancies between these inputs. While this hypothesis explains the presence of prediction errors, recent experimental studies suggest that prediction error signals can emerge within a local circuit, that is, from bottom-up sensory input alone. In this paper, we test whether local circuits alone can generate predictive signals by training a recurrent spiking network using local plasticity rules. Our network model replicates prediction errors resembling various experimental results, such as a biphasic pattern of prediction errors and context-specific representation of error signals. Our findings shed light on how synaptic plasticity can shape prediction errors and enables the acquisition and updating of an internal model of sensory input within a recurrent neural network.
BibTeX
@misc{asabuki2023learning,
  author = {Toshitake Asabuki and Colleen J. Gillon and Claudia Clopath},
  title = {{Learning predictive signals within a local recurrent circuit}},
  year = {2023},
  doi = {10.1101/2023.06.15.545081},
  howpublished = {bioRxiv},
}

Loading related content...

Figure