ABSTRACT
In 2018, Intel published its Loihi neuromorphic research chip and launched the Intel Neuromorphic Research Community (INRC). As an exploratory AI research chip inspired by the principles of neural computation in nature, Loihi explores a very different design regime compared to the Von Neumann and matrix-acceleration architectures in wide use today. Loihi supports sparse and irregular communication between its units packetized in the form of event-driven “spikes”. It integrates self-modifying processes allowing it to autonomously adapt and learn in response to discrete events and an evolving statistical environment. It is implemented using a novel asynchronous design methodology that allows it to fully exploit activation sparsity.
Members of the INRC are now collaboratively advancing the state-of-the-art in neuromorphic algorithms, software, and applications for Loihi. Exciting recent results to date suggest that neuromorphic chips that combine fine-grain, spike-based parallel neural network architecture and an asynchronous design implementation can provide compelling gains in computing efficiency, scalable to orders of magnitude, for the right kinds of adaptive dynamic problems. These include many exciting applications at the forefront of AI that call for highly efficient real-time interaction with real-world environments.