TL;DR
This paper presents a computational model that explains how two types of "prediction-error" neurons (nPE and pPE) can emerge and coexist in the brain. The model shows that a balanced interplay between excitation and inhibition, mediated by diverse inhibitory interneurons, is crucial for forming these neurons, which are essential for learning and adapting to unexpected sensory information.
What Problem Does This Paper Address?
The brain is remarkably adept at predicting upcoming sensory information. When predictions are accurate, neural responses are minimal. However, when there's a mismatch between what's expected and what's actually perceived, specific neurons, known as prediction-error (PE) neurons, become active. These PE neurons are thought to signal the "surprise" and drive learning. Interestingly, PE neurons seem to come in two flavors: those that respond to overpredictions (nPE) and those that respond to underpredictions (pPE). A key question is how these two types of neurons can develop and function together within the complex circuitry of the brain, especially given the vast diversity of inhibitory interneurons that play a critical role in neural computation. This paper aims to unravel the circuit-level mechanisms that allow for the simultaneous formation and coordinated function of both nPE and pPE neurons.
Methods
The researchers developed a computational network model using a rate-based approach, simulating excitatory pyramidal cells (PCs) and three types of inhibitory interneurons (PV, SOM, and VIP). They incorporated biologically inspired connectivity patterns and synaptic plasticity rules, specifically homeostatic inhibitory plasticity. This plasticity mechanism adjusts synaptic strengths to maintain a target firing rate or, in a more refined version, a target total input for the PCs. The model was analyzed using both mean-field approximations for mathematical tractability and full network simulations. They also introduced perturbations to the network to test the robustness of the PE neurons. Finally, they connected the PE circuit to an attractor network to explore functional implications like biased perception.
Datasets
Not discussed in the paper. The study relies on a simulated neural network model rather than experimental datasets.
Key Insights and Ideas
- E/I Balance is Key: The simultaneous existence of both nPE and pPE neurons hinges on a sophisticated balance of excitation and inhibition (E/I balance) not just in total inputs but across multiple pathways within the circuit.
- Diversity of Inhibitory Interneurons is Crucial: The model highlights that different types of inhibitory interneurons (PV, SOM, VIP) and their specific roles in receiving actual sensory input versus predicted input are essential for generating both nPE and pPE responses.
- Homeostatic Plasticity Learns PE Neurons: Experience-dependent learning, specifically homeostatic inhibitory plasticity, can simultaneously shape both nPE and pPE neurons, even with low baseline firing rates.
- Robustness Through Learning: PE neurons, once formed through this balanced E/I learning, are surprisingly robust to moderate network perturbations, suggesting they are stable computational units.
- PE Neurons Shape Perception: The model demonstrates that PE neurons can influence perception, leading to phenomena like "contraction bias" where uncertain stimuli are perceived as closer to the mean of past experiences.
Key Results
- A simplified mean-field analysis showed that a multi-pathway E/I balance is necessary for both nPE and pPE neurons to coexist. This balance requires at least two distinct sources of somatic inhibition and dendritic inhibition driven by actual sensory inputs.
- In a heterogeneous, plastic network, homeostatic inhibitory plasticity with a zero target rate for PCs allowed for the simultaneous development of nPE and pPE neurons.
- Modifying the plasticity rule to target total input (rather than just firing rate) resulted in PE neurons that generalized better to unseen stimuli and were more robust to perturbations.
- The initial connectivity and the distribution of actual and predicted sensory inputs onto interneurons significantly influenced the ratio of nPE to pPE neurons formed.
- The model successfully reproduced "contraction bias" in perception, demonstrating that PE neurons can act as internal cues to switch between different perceptual states or adapt to changing stimulus statistics.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Linearity Assumption: The model primarily uses linear input-output transfer functions for neurons, which is an approximation. Future work could explore how non-linearities affect PE neuron formation and robustness.
- Simplified Dendritic Structure: The model uses a simplified two-compartment model for pyramidal cells. More complex dendritic structures could be investigated.
- Stimulus Representation: Sensory stimuli were modeled as simple step currents. Future research could incorporate more complex temporal dynamics and feature selectivity.
- Excitatory Plasticity: While inhibitory plasticity was central, the role of excitatory plasticity in PE neuron formation was not explored and is suggested as a future avenue.
- Experimental Predictions: The paper makes several testable predictions regarding the robustness of PE neurons to perturbations, their role in biased perception, and their generalization capabilities, which can guide future experimental validation.
Summary updated