TinyML

In order for AI to fulfill its potential to revolutionize the world, it must grow beyond the confines of the server farm and enter, and interact with, the human environment. Thus far, efforts to democratize AI have predominantly focused on mobile inference, but in recent years, there have been major strides towards expanding the scope of ML to ultra-low-power devices. The field, known as “TinyML”, achieves ML inference under a milliWatt, and thereby breaks the traditional power barrier preventing widely distributed machine intelligence. By performing inference on-device, and near-sensor, TinyML enables greater responsiveness and privacy while avoiding the energy cost associated with wireless communication, which at this scale is far higher than that of compute. On the other hand, TinyML comes with a unique set of challenges which stem from tight constraints on the computational power, memory capacity, and energy budget of these devices.MLPerf Tiny BenchmarkTo overcome these challenges, TinyML systems must be fully optimized, from data to hardware, to fit the specific application, which is often an intractably time consuming process. Our research aims to decrease the barrier of developing TinyML applications by making the co-design of systems across layers of abstraction easier and standardizing generalizable workloads. We have developed benchmarks, datasets, models and tools to enhance the TinyML application design process and thereby enable a new class of intelligent systems.