Learning prediction error neurons in a canonical interneuron circuit

Loreen Hertäg, Henning Sprekeler

TL;DR

This paper presents a computational model of the mouse visual cortex that explains how "prediction error" neurons, crucial for learning and adapting to sensory input, can emerge through experience-dependent plasticity. The model highlights the critical role of inhibitory interneurons in balancing excitation and inhibition to fine-tune these error signals, offering testable predictions for future experimental research.

What Problem Does This Paper Address?

Our brains are constantly trying to predict what sensory information we'll receive next. When the actual sensory input deviates from our predictions, it generates a "prediction error." These errors are vital for learning and updating our internal models of the world. While we know these prediction error signals exist in various brain regions, the precise neural circuits and mechanisms responsible for computing them, especially in the visual cortex, remain largely mysterious. Specifically, how the complex interplay of different neuron types, particularly inhibitory interneurons, contributes to the formation and refinement of these prediction error signals is not well understood.

Methods

The authors developed a computational model of layer 2/3 of the mouse primary visual cortex. This model is "rate-based," meaning it simulates the average firing rate of neurons rather than their precise spike timing. The network includes excitatory pyramidal cells (PCs) and three types of inhibitory interneurons: parvalbumin-expressing (PV), somatostatin-expressing (SOM), and vasoactive intestinal peptide-expressing (VIP) neurons.

The core of the model's learning mechanism lies in experience-dependent inhibitory plasticity. This means that the strength of connections between inhibitory interneurons and PCs changes based on the network's activity over time. Specifically, they used learning rules that aim to balance excitation and inhibition in PCs, effectively minimizing prediction errors. They also explored different input configurations to these neuron types and simulated "optogenetic manipulations" (virtually activating or inactivating specific neuron types) to see how these changes affected the network's behavior and to distinguish between different circuit variants.

Datasets

Not discussed in the paper. The study relies on a computational model and simulated data.

Key Insights and Ideas

  • Inhibitory Interneurons are Key Players: The model strongly suggests that the precise balance of excitation and inhibition, orchestrated by PV, SOM, and VIP interneurons, is fundamental for learning prediction error signals. It's not just about the excitatory neurons; the inhibitory network is doing a lot of the heavy lifting.
  • Experience Shapes Prediction Circuits: The model demonstrates that "training" the network with realistic sensorimotor experiences (predicting visual input based on motor commands) leads to the emergence of prediction error neurons. This highlights the plasticity and adaptability of neural circuits.
  • Multiple Circuit Variants Exist: The same functional outcome (prediction error neurons) can be achieved through different circuit configurations, depending on which interneuron types receive which type of input (visual vs. motor prediction). This suggests a degree of flexibility in how the brain implements this computation.
  • Optogenetics as a Diagnostic Tool: Simulated optogenetic experiments proved to be powerful in distinguishing between these different circuit variants, suggesting that such manipulations could be used experimentally to probe the underlying circuitry in real brains.
  • Biologically Plausible Learning: The paper shows that these prediction error circuits can be learned using plasticity rules that are biologically plausible, meaning they rely on local information at the synapse, rather than requiring global "backpropagation" signals.

Key Results

  • The model successfully generated negative prediction-error (nPE) neurons that respond specifically when sensory input is weaker than predicted.
  • These nPE neurons emerged through inhibitory synaptic plasticity that balanced excitation and inhibition in pyramidal cells.
  • Simulated optogenetic inactivation/activation of PV, SOM, and VIP neurons could reliably distinguish between different input configurations of the learned circuits, aligning with experimental observations.
  • The fraction of nPE neurons that developed was dependent on the distribution of visual and motor inputs to SOM and VIP interneurons.
  • The model's experience-dependence (e.g., reduced nPE neuron formation under "non-coupled training") mirrored findings in mouse visual cortex.
  • The study identified specific constraints on synaptic strengths and neuron types required for nPE neuron formation in various network configurations.

Limitations and Future Work

  • Simplified Neuron Models: The model uses rate-based neurons and a simplified two-compartment model for PCs, neglecting detailed spiking dynamics and dendritic computations.
  • Lack of Feature Selectivity: The model did not incorporate feature selectivity (e.g., orientation tuning) in neurons, which is a hallmark of sensory areas. The authors suggest future work could explore how feature selectivity impacts prediction error circuits.
  • Focus on nPE Neurons: While the principles are expected to apply to positive prediction-error (pPE) neurons, the study primarily focused on nPE neurons due to more extensive experimental data. Future work could investigate the simultaneous development of both.
  • Neuromodulation: The authors suggest that neuromodulatory effects could play a role in context-dependent responses and the fine-tuning of prediction error circuits, an area for future exploration.
  • Experimental Validation: The paper makes several testable predictions, particularly regarding the effects of optogenetic manipulations, which await experimental verification.

Summary updated

Abstract

Sensory systems constantly compare external sensory information with internally generated predictions. While neural hallmarks of prediction errors have been found throughout the brain, the circuit-level mechanisms that underlie their computation are still largely unknown. Here, we show that a well-orchestrated interplay of three interneuron types shapes the development and refinement of negative prediction-error neurons in a computational model of mouse primary visual cortex. By balancing excitation and inhibition in multiple pathways, experience-dependent inhibitory plasticity can generate different variants of prediction-error circuits, which can be distinguished by simulated optogenetic experiments. The experience-dependence of the model circuit is consistent with that of negative prediction-error circuits in layer 2/3 of mouse primary visual cortex. Our model makes a range of testable predictions that may shed light on the circuitry underlying the neural computation of prediction errors.
BibTeX
@inproceedings{hertag2020learning,
  author = {Loreen Hertäg and Henning Sprekeler},
  title = {{Learning prediction error neurons in a canonical interneuron circuit}},
  booktitle = {eLife},
  year = {2020},
  doi = {10.7554/elife.57541},
}
Figure