TL;DR
This paper introduces a novel, anatomically-organized, multi-scale computational model of the corticostriatal circuit that bridges low-level neural activity with high-level cognitive functions like decision-making. Remarkably, this "zero-trained" model not only replicates known neural phenomena but also discovers a new neural code for predicting erroneous behaviors, which was subsequently confirmed in empirical data.
What Problem Does This Paper Address?
Neuroscience faces a significant challenge in connecting the intricate details of neural activity (spiking, field potentials, biochemistry) to complex cognitive abilities (decision-making, working memory) and associated disorders. Existing computational models often excel at either mechanistic accuracy or emergent cognition, but rarely both. There's a need for models that are biologically realistic at the micro-scale, can generate complex cognitive phenomena, and can make novel predictions beyond their training data. This paper aims to bridge this gap by creating a biomimetic model that directly links neural signals to cognitive functions and can discover new neural codes.
Methods
The authors developed an anatomically-organized, multi-scale model of the corticostriatal circuit. This model is built from "biomimetic computational primitives" (BCPs), which are assemblies of simulated neurons organized according to known anatomical structures and physiological operations. These BCPs, such as lateral inhibition and action selection, are not hardcoded but emerge from the model's architecture. The model simulates spiking neurons, local field potentials, phase synchronies, and synaptic changes. A key aspect is its "zero-trained" nature, meaning it wasn't trained on specific experimental data but rather built from established neuroscientific principles. The model was validated against extensive macaque electrophysiology and behavioral data.
Datasets
The model was validated using extensive experimental macaque (non-human primate, NHP) electrophysiology and behavioral data. The paper does not introduce new datasets but uses existing, well-established datasets for validation.
Key Insights and Ideas
- Bridging Scales: The model successfully integrates low-level neural dynamics (spiking) with high-level cognitive functions (category learning, decision-making) within a single, unified framework.
- Emergent Cognition from Primitives: Complex cognitive operations and neural codes emerge naturally from the interaction of simpler, biologically-inspired computational primitives, rather than being explicitly programmed.
- Predictive Power of "Incongruent Neurons": The simulation uncovered a previously unknown neural code – "incongruent neurons" (ICNs) – that specifically predict upcoming erroneous behaviors. This discovery was a significant "aha moment" and was subsequently confirmed in empirical data.
- Zero-Training for Discovery: The model's ability to make novel predictions without prior training on specific datasets highlights the power of biomimetic modeling for genuine scientific discovery.
Key Results
- The model accurately replicated known physiological patterns and behaviors observed in macaque brains during a category learning task.
- It successfully learned the visual categorization task, demonstrating synaptic changes that led to improved accuracy over trials.
- The model predicted the existence of "incongruent neurons" (ICNs) that selectively fire before an incorrect choice, a finding later confirmed in empirical macaque data. These ICNs were found to be statistically significant predictors of erroneous behavior.
- The model showed comparable learning rates to NHPs and replicated cortico-striatal synchrony effects.
Limitations and Future Work
The authors acknowledge that this initial rendition of the model uses "reasonably simplified population-level abstractions" of physiological mechanisms and anatomical constraints. While the model successfully simulates many aspects of NHP behavior, there are minor differences, such as the simulated brain showing a more pronounced decrease in spiking after stimulus offset compared to empirical data. Future work could explore the potential applicability of this model to a broader range of challenging questions related to predictive coding, reinforcement learning, and neurological disorders. The discovery of ICNs also opens up new avenues for understanding decision-making mechanisms and avoiding local minima in complex environments.
Summary updated