Modeling the formation of prediction-error neurons in complex neural networks
Final Report Abstract
Recent research reveals that neuronal networks continuously compare sensory input—what we see, hear, smell, or feel—with predictions based on past experiences and expectations. This process is thought to rely on specialized prediction-error neurons, which signal mismatches between actual and expected inputs by increasing their activity. These neurons are thought to play a crucial role in refining our internal model of the world and shaping neural plasticity in hierarchical networks, thereby enabling learning and adaptation. While evidence of prediction-error activity has been observed throughout the brain, the precise mechanisms driving prediction errors at the circuit level remain poorly understood. How do neural networks compute these mismatches? What makes prediction-error neurons emerge, and how do they remain robust yet adaptable? This project explored the hypothesis that prediction-error neurons arise from a finely tuned interplay between excitatory cells and diverse inhibitory interneurons. In particular, the balance of excitation and inhibition across different compartments of excitatory cells appears essential for detecting deviations from predicted sensory inputs. Using mathematical analyses and network simulations, this study demonstrates that different types of prediction-error neurons can emerge within the same recurrent network when excitation and inhibition are balanced across multiple excitatory and inhibitory pathways. This balance, shaped by various inhibitory interneurons— including PV, SOM, VIP, and NDNF interneurons—enables the brain to generalize learning and adapt to novel stimuli. Furthermore, the findings reveal under which conditions homeostatic inhibitory plasticity alone is sufficient to generate fully functional prediction-error circuits. Additionally, this work addresses how the brain manages uncertainty in sensory processing. The results demonstrate that a hierarchical prediction-error circuit can estimate both sensory and prediction uncertainty through the activity of prediction-error neurons. The model confirms that when sensory input is noisy and the environment remains stable, the brain relies more heavily on predictions—aligning with theories of multi-sensory integration. By combining computational and theoretical approaches, this work shed light on the fundamental principles that give neural circuits their remarkable ability to predict sensory inputs and detect mismatches. Understanding these mechanisms has broad implications for neuroscience, from advancing artificial intelligence to providing insights into neurological conditions where predictive processes go awry, such as schizophrenia or autism.
Publications
-
Prediction-error neurons in circuits with multiple neuron types: Formation, refinement, and functional implications. Proceedings of the National Academy of Sciences, 119(13).
Hertäg, Loreen & Clopath, Claudia
-
A Cortical Microcircuit for Region-Specific Credit Assignment in Reinforcement Learning. openRxiv.
Chevy, Quentin; Szadai, Zoltán; Hertäg, Loreen; Moll, Morgan; Gibson, Elizabeth T.; Costa, Rui Ponte & Kepecs, Adam
-
Transcriptomic Correlates of State Modulation in GABAergic Interneurons: A Cross-Species Analysis. The Journal of Neuroscience, 44(44), e2371232024.
Keijser, Joram; Hertäg, Loreen & Sprekeler, Henning
-
Layer-specific control of inhibition by NDNF interneurons. Proceedings of the National Academy of Sciences, 122(4).
Naumann, Laura Bella; Hertäg, Loreen; Müller, Jennifer; Letzkus, Johannes J. & Sprekeler, Henning
-
Self-supervised predictive learning accounts for cortical layer-specificity. Nature Communications, 16(1).
Nejad, Kevin Kermani; Anastasiades, Paul; Hertäg, Loreen & Costa, Rui Ponte
-
Uncertainty estimation with prediction-error circuits. Nature Communications, 16(1).
Hertäg, Loreen; Wilmes, Katharina A. & Clopath, Claudia
