상세 보기
A Digital Neural Array IC for Real-Time Neural Network Replication from Spike Activities
- Park, Donghyun;
- Mun, Hajung;
- Jeong, Minhyeok;
- Kang, Dahee;
- Lee, Jongmin;
- ... Lee, Yoonmyung
SCOPUS
0초록
Growing demand to deepen understanding of the human brain has accelerated efforts to identify the structure of biological neural networks from neuronal activities. This paper presents a fully digital, tile-able neural array integrated circuit(IC) that, to our knowledge, is the first hardware platform for network reconstruction - inferring synaptic connectivity directly from spike-train data generated by a biological (ground-truth) network. Designed with the overarching goal of emulation of biological networks, the architecture employs repeatable digital neuron-module tiles to ensure the scalability, flexibility and verifiability. Two chip-level run-time interfaces are integrated: a writable spike-forcing path for injecting biological spike pulses into selected IC neurons for synchronized co-firing, and a dedicated monitoring path for streaming spike events, synaptic weights, and membrane potentials. Scalability is further enabled by single-timer Δt capture and a piecewise-linear STDP Δw generator shared per neuron, avoiding complex LUTs/multipliers while preserving biological plausibility. The platform is realized in silicon and tested with an FPGA-based setup. Using only spike activity, the system reconstructs synaptic connectivity with high fidelity across diverse ground-truth networks ranging from simple 2-layer topologies to biologically derived C.elegans head network, as well as networks with bimodal and trimodal weight distributions. Accuracy was comprehensively quantified from multiple perspectives for both connectivity and spike-train similarity, confirming faithful replication. These results demonstrate that the proposed platform can recover the synaptic structure from spikes alone and provide a practical tool for predicting network learning responses under varied stimuli, advancing biological neural network research and real-time neuromorphic experimentation.
키워드
- 제목
- A Digital Neural Array IC for Real-Time Neural Network Replication from Spike Activities
- 저자
- Park, Donghyun; Mun, Hajung; Jeong, Minhyeok; Kang, Dahee; Lee, Jongmin; Lee, Yoonmyung
- 발행일
- 2026
- 유형
- Conference Paper
- 저널명
- Proceedings -Design, Automation and Test in Europe, DATE
- 언어
- ENG
- 출판사
- Institute of Electrical and Electronics Engineers Inc.
- ISSN
- P 1530-1591