ReckOn: A 28nm Sub-mm2 Task-Agnostic Spiking Recurrent Neural Network Processor Enabling On-Chip Learning over Second-Long Timescales

TL;DR

Autonomous edge devices need on-chip adaptation to changing users, environments, and tasks, but prior memory-constrained learning devices handled mainly static or instantaneous data. ReckOn uses a spiking RNN with eligibility-propagation-based online learning, sparsity, and address-event processing to train over seconds across navigation, gesture recognition, and keyword spotting. It achieves task-agnostic end-to-end learning with sub-150µW training power and sub-mm² area, including 0.45mm² core area and 138kB SRAM.

• Problem

  Changing users, environments, and tasks limit inference-only edge-device robustness, while prior end-to-end on-chip learning was restricted mainly to static data or instantaneous decisions without temporal memory.

• Method

  ReckOn is a spiking RNN processor using local eligibility traces, sparse processing, and address-event representations for online learning over seconds with millisecond temporal resolution.

• Results

  ReckOn demonstrates end-to-end on-chip learning on gesture recognition, keyword spotting, and navigation, achieving 87.3%-, 90.7%-, and 96.4%-accuracy respectively, with a 0.45mm² core and 138kB SRAM.

• Takeaways & Limitations

  ReckOn supports task-agnostic learning over thousands of timesteps within constrained edge-device power and area budgets, enabling chip repurposing and sensor fusion.