Online Continual Learning for Loihi 2
Speakers | @Andreas Wild (Intel) @Danielle Rager (Intel) @Elvin Hajizada (Intel) @Joe Hays (US Naval Research Lab) |
|---|---|
Description | Online Continual Learning for Loihi 2 (Intel) 10:30 PT / 13:30 ET / 19:30 CET Understand how local, online learning as supported by Intel Loihi 2 works, what kinds of applications might take advantage of it, and how to use online learning rules in Lava.  INRC Members and Affiliates: Check out Danielle Rager’s Loihi 2 Learning Deep Dive
A Case for Continual Learning (US Naval Research Lab) 11:15 PT / 14:15 ET / 20:15 CET Tune into this INRC special guest talk by NRL to understand WHY and HOW to use continual learning, with connections to robotic and space applications. This effort leads into research on differentiable plasticity, a breakthrough feature enabling the combination of fast GPU-based model training with flexible Loihi-based online learning. To support differentiable plasticity, NRL researchers will discuss their work to close the gaps between plasticity simulations in Lava and their on-chip counterparts.
Continual Learning Prototype Classifier for Image Classification (Intel) 12:00 PT / 15:00 ET / 21:00 CET Learning continually in the wild requires detecting novel patterns and learning them fast. As we may have only one or few instances of the new class, we need to learn it fast and ideally in a one-shot manner. In addition, we want our system to be able to perform open-set recognition, i.e. rather than performing recognition purely on the known classes (close-set recognition), we also want to recognize instances from unknown classes. In this regard, Continually Learning Prototypes (CLP) is the Loihi 2 successor of the neuromorphic prototype-based continual learning algorithm we have developed for Loihi 1. It is built to learn in the above-described challenging setting. In this talk, we will first introduce CLP and discuss how it is developed in Loihi 2 - aware manner. Next, we will demonstrate first Lava - Loihi 2 implementation of CLP can learn on real image data. |