TinyTNAS: Time-Bound, GPU-Independent Hardware-Aware Neural Architecture Search for TinyML Time-Series Classification

We present TinyTNAS, a hardware-aware Neural Architecture Search (NAS) framework optimized for efficient execution on CPUs, eliminating the need for costly GPUs. Traditional NAS methods often depend on reinforcement learning or evolutionary algorithms, requiring significant GPU resources and search time, which may be inaccessible to many machine learning researchers and practitioners. TinyTNAS addresses these limitations with an intelligent grid search approach that drastically reduces search time from hours to minutes, operating seamlessly on CPUs. It enables scalable model generation tailored for resource-constrained devices, optimizing neural networks within stringent constraints on RAM, Flash, and MAC operations. TinyTNAS also supports time-bound searches, ensuring rapid and efficient architecture discovery. Experiments on benchmark datasets, including UCIHAR, PAMAP2, WISDM, MIT-BIH, and PTB-ECG, demonstrate its ability to achieve state-of-the-art accuracy while significantly reducing resource usage and latency compared to expert-designed architectures. Furthermore, it surpasses GPU-dependent hardware-aware NAS methods based on reinforcement learning and evolutionary algorithms by drastically reducing search time. The code is publicly available at https://github.com/BidyutSaha/TinyTNAS.git