Energy-efficient activity-driven computing architectures for edge intelligence

We present an overview of different methods to increase the energy efficiency of Tiny Machine Learning networks running on resource-constrained edge neural network accelerators. Most commonly reported are accelerator designs that support networks with sparse weight matrices, that skip over zero activations, and that use low-bit precision parameters. We introduce a brain-inspired sparsity, i.e. temporal sparsity, that can further reduce computes needed during network inference. Accelerators that support this sparsity require storage of the network states. This storage is already needed in a recurrent neural network. Activity-driven sensors like the event camera sensor, when interfaced to a neural network, can further reduce the computes by triggering network updates only when necessary, especially for tasks involving temporal changes in the input. The different sparsity methods and the input activity-driven form of computing can pave the way to edge devices with the high energy-efficiency seen of brains when operating in uncontrolled natural environments. These devices are targeted at edge tasks for domains such as wearables, brain-machine interfaces, and Internet-of-Things (IoT) applications.