P041 Balanced inhibition allows for robust learning of input-output associations in feedforward networks with Hebbian plasticity
Gloria Cecchini*1, Alex Roxin1
1Centre de Recerca Matemàtica, Barcelona, Spain
*Email: gcecchini@crm.cat
Introduction
In neural networks, post-synaptic activity depends on multiple pre-synaptic inputs. Hebbian plasticity allows sensory inputs to be associated with internal states, as seen in the CA1 region of the hippocampus. By modifying synaptic weights, Hebbian rules enable sensory inputs to elicit correlated outputs, allowing for efficient memory storage. When input and output patterns are uncorrelated, numerous associations can be encoded. However, if output patterns weakly correlate with input patterns, Hebbian learning reinforces shared synapses across patterns, leading to reduced network flexibility and impaired associative learning.
Methods We analyzed the effects of Hebbian plasticity in a feedforward network model where input-output correlations emerge due to intrinsic connectivity. Using numerical simulations, we examined how weak correlations between inputs and outputs shape synaptic weight dynamics over time. We then introduced a balanced inhibition mechanism, inspired by in-vivo cortical circuits [1], to assess its impact on synaptic weight distribution and the network’s ability to store diverse associations. Network performance was evaluated by measuring output pattern variability.
Results Our results show that when weak correlations exist between input and output patterns, Hebbian learning selectively strengthens synapses shared across patterns. This reinforcement leads to a rigid network state, where outputs become highly correlated over time. Consequently, the network loses the ability to store multiple distinct associations, significantly reducing its learning capacity. However, introducing balanced inhibition prevents the over-strengthening of shared synapses, allowing output patterns to remain distinct and ensuring a more flexible associative learning process.
Discussion These findings highlight a fundamental limitation of Hebbian learning in feedforward networks when input-output correlations exist. Without a regulatory mechanism, the network structure becomes overly rigid, preventing effective storage of new associations. Balanced inhibition emerges as a simple yet effective strategy to mitigate this issue, preserving learning flexibility by counteracting correlation-driven synaptic reinforcement. Our study underscores the critical role of inhibition in biological neural circuits, offering insights into how the brain maintains efficient and adaptive information processing.
Acknowledgements This project has received funding from Proyectos De Generación De Conocimiento 2021 (PID2021-124702OB-I00). This work is supported by the Spanish State Research Agency, through the Severo Ochoa and Maria de Maeztu Program for Centers and Units of Excellence in R&D (CEX2020-001084-M). We thank CERCA Programme/Generalitat de Catalunya for institutional support. References 1. Bilal Haider, Alvaro Duque, Andrea R. Hasenstaub, David A. McCormick (2006) Neocortical Network Activity In Vivo Is Generated through a Dynamic Balance of Excitation and Inhibition. Journal of Neuroscience, 26 (17) 4535-4545; https://10.1523/JNEUROSCI.5297-05.2006
