Low-Energy Real-Time Self-Organizing Classifier Architecture for Edge SNN Applications
37th International Conference on Microelectronics-ICM, Cairo, Mısır, 14 - 17 Aralık 2025, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icm66518.2025.11322489
- Basıldığı Şehir: Cairo
- Basıldığı Ülke: Mısır
- Orta Doğu Teknik Üniversitesi Adresli: Evet
Özet
The increasing demand for real-time edge machine learning is driven by the emergence of AI sensors with advanced classification requirements. This work builds upon a previously developed low-energy Spiking Neural Network (SNN) architecture, enhancing autonomy, processing time, and energy efficiency through the integration of dynamic preprocessing filters with self-organizing logic. Key features include a serial streaming interface optimized for pipelining image rows, a simple downsampling mechanism, a filter selection unit, and adaptive filter blocks capable of dynamically monitoring multiple classes and making adjustments for sampled variations within each class. Crucial design hyperparameters, defined as Filters per Class (FpC) and Common Features per Class Filter (CFpCF), are optimized for the MNIST classification problem. This optimization results in a reduction of total energy consumption by more than an order of magnitude compared to the offline preprocessing version. Furthermore, the proposed system achieves superior accuracy compared to recent FPGA-based approaches, with a reduction in energy consumption per image exceeding two orders of magnitude.