Our first-generation neuromorphic algorithm for signal learning and identification under noise has been ported to the Intel Loihi chip (full paper here), and has been featured in an Intel Newsroom article.  As an Intel INRC member lab, we will continue to work closely with Intel to optimize this algorithm for use in embedded devices.  Look for our upcoming paper in early 2019.  

Neuromorphic algorithms are spiking neural network (SNN) designs based on features and attributes drawn from or inspired by the neural circuits of the brain.  In contrast to traditional deep learning algorithms, neuromorphic/SNN designs are based on strongly localized computational elements (cells, synapses), the colocalization of memory and compute, and strong priors embedded in the initial network topology and learning rules  (“the architecture is part of the algorithm”).  Accordingly, a given SNN is not as generic as a backpropagation-based deep network, but can achieve asymptotic levels of learning much more rapidly (with fewer training examples) than a deep network.  Our design, in particular, is capable of genuine online learning (the learning of new signals does not disrupt prior learning) and lifelong learning (this process can, in principle, continue indefinitely).  

Despite the association with SNNs, neuromorphic networks are not always “spiking”, though spike-based communication does enjoy critical advantages in power expenditure and robustness to environmental conditions — both important features for field-deployable systems.  Accordingly, spike-based computation makes the best use of Loihi resources, though the chip provides alternative infrastructure for computations not yet implemented as SNN algorithms.