P214 A Canonical Microcircuit of Predictive Coding Under Efficient Coding Principles
Elnaz Nemati*1, Catherine E. Davey1, Hamish Meffin1,2, Anthony N. Burkitt1,2
1Department of Biomedical Engineering, The University of Melbourne, Victoria, Australia.
2Graeme Clark Institute, The University of Melbourne, Victoria, Australia.
*Email: fnemati@student.unimelb.edu.au
Introduction
Predictive coding describes how the brain integrates sensory inputs with expectations by minimizing expectation errors [1]. Studies show increased neural activity in cortical L2/3 during sensory mismatches [2], offering insights into disorders like autism [3] and perceptual phenomena such as visual illusions [4]. Canonical microcircuit models [5, 6] advance understanding but often overlook spiking dynamics, detailed inhibitory mechanisms, and Dale’s law adherence. They also neglect the distinct roles of cortical layers, especially L4 and L5/6 [7]. The Deneve framework [8] provides another perspective, modeling neurons as decoders where spikes are triggered if the membrane potentials, representing reconstruction errors, exceeds a threshold.
Methods
This study extends Deneve’s predictive coding framework [7] by assigning Gabor receptive fields to layer 4 neurons, creating a V1-like biologically inspired feature extractor. It introduces two-compartment neurons in layer 2/3 for prediction error signaling within a balanced E/I network. Our hierarchical model mirrors canonical circuits, using spiking neurons and simplified inhibitory populations: Parvalbumin (PV) and Somatostatin (SOM) inspired by Hertäg and Clopath [8]. Layer 5/6 contains similar neurons, generating predictions balanced by these populations (Fig.1a,b). Employing spiking neurons with Leaky Integrate-and-Fire dynamics, the model processes whitened images in ON/OFF channels, as in experimentally observed LGN responses.
Results
The model successfully results in L4 neurons displaying balance (Fig.1c), and orientation and phase selectivity (Fig.1d,e), thereby demonstrating biologically realistic V1 feature extraction. Layer 2/3 neurons robustly signal prediction errors across matched (FF=FB), mismatched (FF≠FB), feedforward-only (FF>FB), and feedback-only (FB>FF) conditions. Neuronal responses matched experimental evidence, where matched inputs minimized activity, while mismatched inputs elicited strong prediction-error signaling (Fig.1d). Critically, layer 5/6 neurons effectively integrated prediction errors from layer 2/3, significantly reducing sensory reconstruction errors and validating their predictive coding function.
Discussion
The model proposes that predictive coding effectively described cortical function through specific feedback interactions within canonical cortical circuits. It highlights the essential roles played by distinct neuronal compartments and inhibitory inter-neuron populations, specifically PV and SOM neurons, in modulating the balance. The close alignment of theoretical predictions with experimental observations supports the model's validity and enhances our understanding of cortical dynamics. Additionally, the model provides a robust foundation for future research in perceptual neuroscience, the development of neuromorphic systems, and the exploration of clinical interventions for disorders involving disrupted predictive coding mechanisms.
Figure 1. Fig 1: (a) Predictive coding microcircuit representation. (b) Detailed circuitry within each layer and connectivity. (c) Display of excitatory, inhibitory, and net currents showing balanced currents (d) Orientation Bias Index and (e) Phase Bias Index of Layer 4 excitatory populations. (f) Spike responses in Layer 2/3 to various feedforward and feedback inputs.
Acknowledgements
-
References
1.https://doi.org/10.1038/4580
2.https://doi.org/10.1016/j.neuron.2020.09.024
3.https://doi.org/10.1152/jn.00543.2015
4.https://doi.org/10.1016/j.neunet.2021.08.024
5.https://doi.org/10.1016/j.neuron.2012.10.038
6.https://doi.org/10.1016/j.neuron.2018.10.003
7.https://doi.org/10.1073/pnas.2115699119
8.https://doi.org/10.1038/nn.4243
