P165 Incorporation of Neuromodulation into Predictive-Coding Spiking Neural Networks
Yelim Lee1, Dongmyeong Lee1, Hae-Jeong Park*1,2,3,4
1Department of Nuclear Medicine, Graduate School of Medical Science, Brain Korea 21 Project, Yonsei University College of Medicine, Seoul, Republic of Korea
2 Department of Nuclear Medicine, Severance Hospital,Seoul, Republic of Korea
3Department of Cognitive Science, Yonsei University, Seoul, Republic of Korea
4Center for Systems and Translational Brain Sciences, Institute of Human Complexity and Systems Science, Yonsei University, Seoul, Republic of Korea
*Email: parkhj@yonsei.ac.kr
Introduction
Neuromodulation is often considered to enhance the selection mechanism in the brains of living organisms, prioritizing processing of inputs relevant to their goals. By modifying effective synaptic strength and altering firing properties, neuromodulators engage various cellular mechanisms, leading to a dynamic reconfiguration of neural circuits. Adjusting a target neuron’s excitability is one mechanism for enabling attentional effects. This research explores how this mechanism enhances predictive coding and learning in a spiking neural network (SNN) with two-compartment neurons, focusing on classification ability and internal representations in hidden layers with top-down signals.
Methods
The network includes one input layer, one output layer, and three fully connected hidden layers with feedback and feedforward connections. The dynamics of the hidden neurons are based on the Adaptive Leaky-Integrate-and-Fire (ALIF) model from Zhang and Bohte's previous work. The dendritic compartment of the hidden neurons integrate inputs from higher regions, and the somatic compartment integrates input from lower areas. To implement the neuromodulation effect on the hidden neurons, we introduced a new top-down attention connection from the higher layer to the lower layer. This adjustment enables modifying the target neuron’s excitability by dynamically altering the baseline firing threshold. We used a spiking MNIST image as input data, modifying the original MNIST dataset to provide spiking input over time. Additionally, we created multiple variations of the MNIST dataset, introducing noise or making occluded or overlapped images, to provide the ambiguous context.
Results
We performed image classification tasks with the MNIST dataset, achieving a high accuracy for the original set and highly noisy test data set. We analyzed the uncertainty of output neurons by tracking their membrane potential for each digit class, noting increased firing for the correct class despite initial uncertainty. To assess predictive coding, we evaluated each hidden layer's internal representation by decoding the spiking activity. This involved no inputs or half-occluded inputs while clamping the output neuron’s membrane potential to a specific class. The results showed successful digit representation in spiking activities, especially with applied modulation weights, compared to the previous model.
Discussion
Clarifying important information in uncertain contexts improves with appropriate attention and prediction. This study suggests that neuromodulation enhances hierarchical encoding and learning in SNN during ambiguous scenarios. The model maintained high classification accuracy even in noisy and occluded conditions, and the internal representation, along with reduced uncertainty of output neurons, aligns with predictive coding principles, where top-down modulation refines internal representations.
Acknowledgements
This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NO. 2023R1A2C200621711)
References
1.Desimone, R., & Duncan, J. (1995). Neural mechanisms of selective visual attention. Annual review of neuroscience, 18(1), 193-222.
2.Marder, E. (2012). Neuromodulation of neuronal circuits: back to the future. Neuron, 76(1), 1-11.
3.Thiele, A., & Bellgrove, M. A. (2018). Neuromodulation of attention. Neuron, 97(4), 769-785.
4.Zhang, M., & Bohte, S. M. (2024). Energy Optimization Induces Predictive-coding Properties in a Multicompartment Spiking Neural Network Model. bioRxiv, 2024-01.
