P308 Brain-Inspired Recurrent Neural Network Featuring Dendrites for Efficient and Accurate Learning in Classification Tasks
Eirini Troullinou*1,2, Spyridon Chavlis1, Panayiota Poirazi1
1Institute of Molecular Biology and Biotechnology, Foundation for Research, and Technology-Hellas, Heraklion, Greece
2Institute of Computer Science, Foundation for Research, and Technology-Hellas, Heraklion, Greece
*Email: eirini_troullinou@imbb.forth.gr
Introduction
Artificial neural networks (ANNs) have achieved substantial advancements in addressing complex tasks across diverse domains, including image recognition and natural language processing. These networks rely on a large number of parameters to attain high performance; however, as the complexity of ANNs increases, the challenge of training them efficiently also escalates [1]. In contrast, the biological brain, which has served as a fundamental inspiration for ANN architectures [2], exhibits remarkable computational efficiency by processing vast amounts of information with minimal energy consumption [3]. Moreover, biological neural networks demonstrate robust generalization capabilities, often achieving effective learning with limited training samples, a phenomenon known as few-shot learning.
Methods
In an effort to develop a more biologically plausible computational model, we propose a sparse, brain-inspired recurrent neural network (RNN) that incorporates biologically motivated connectivity principles. This approach is driven by the computational advantages of dendritic processing [4], which have been extensively studied in biological neural networks. Specifically, our model enforces structured connectivity constraints that emulate the physical relationships between dendrites, neuronal somata, and inter-neuronal connections. These biologically inspired connectivity rules are implemented via structured binary masking, thereby regulating the network's architecture based on empirical neurophysiological observations.
Results
To assess the efficacy of the proposed model, we conducted a series of experiments on benchmark image and time-series datasets. The results indicate that our brain-inspired RNN attains the highest accuracy achieved by a conventional (vanilla) RNN while utilizing fewer trainable parameters. Furthermore, when the number of trainable parameters is increased, our model surpasses the peak performance of the vanilla RNN by a margin of 3–20%, depending on the dataset. In contrast, the conventional RNN exhibits overfitting tendencies, leading to significant performance degradation.
Discussion
In summary, we present a biologically inspired RNN architecture that incorporates dendritic processing and sparse connectivity constraints. Our findings demonstrate that the proposed model outperforms traditional RNNs in both image and time-series classification tasks. Additionally, the model achieves competitive performance with fewer parameters, highlighting the potential role of dendritic computations in machine learning. These results align with experimental evidence suggesting the critical contribution of dendrites to efficient neural processing, thereby offering a promising direction for future ANN development.
AcknowledgementsThis work was supported by the NIH (GA: 1R01MH124867-04), the TITAN ERA Chair project under Contract 101086741 within the Horizon Europe Framework Program of the European Commission, and the Stavros Niarchos Foundation and the Hellenic Foundation for Research and Innovation under the 5th Call of Science and Society "Action Always strive for excellence – Theodoros Papazoglou" (DENDROLEAP 28056).
References
[1] Abdolrasol, M. G, et al. (2021). Artificial neural networks based optimization techniques: A review. Electronics, 10(21), 2689.
[2] Sejnowski, T. J. (2020). The unreasonable effectiveness of deep learning in artificial intelligence. PNAS, 117(48), 30033-38.
[3] Attwell, D., & Laughlin, S. B. (2001). An energy budget for signaling in the grey matter of the brain. J Cereb Blood Flow Metab, 21(10), 1133-1145.
[4] Poirazi, P., & Papoutsi, A. (2020). Illuminating dendritic function with computational models. Nat Rev Neurosci, 21(6), 303-21.
