We devised several novel solutions for the execution of complex Artificial Intelligence workloads on resource-constrained wearable devices. Our work includes edge-friendly optimizations of Machine Learning algorithms including real-time energy-aware feature selection.
Additionally, we adapt Deep Neural Networs for execution on the edge using techniques such as per-layer heterogeneous quantization (which leads to the use of low-bitwidth data representations on dfferent AI layers), and ensambling (which employs several small networks instead of larger monolithic ones).
E2CNN: ensembles of CNN | |
HDTorch: HyperDimensional computing library | |
Neural Network Quantization and Pruning |