Bottom-up and Top-down Neural Processing Systems Design: Unveiling the Road Toward Neuromorphic Intelligence

Explore the future of embedded cognition and neuromorphic intelligence in this 55-minute keynote from the tinyML EMEA 2021 conference. Delve into bottom-up and top-down neural processing systems design approaches, examining their potential to address the limitations of conventional computing. Learn about digital time-multiplexed spiking neural network processing devices like ODIN and MorphIC, and discover mixed-signal analog-digital neuromorphic processors. Investigate the Direct Random Target Projection (DRTP) algorithm for low-cost on-chip learning, and understand how combining event-driven and frame-based processing can support adaptive edge computing. Compare the tradeoffs between bottom-up and top-down approaches, and explore their applications in edge computing scenarios.

Intro
Neuromorphic approaches
Methodology
Analog or digital
Bottomup approach
Local learning rules
Topdown approach
Morphic
Benchmark
Results
Q A
Silicon integration
Spoon
Accuracy energy tradeoff
Timebased computation
Spoon chip
Whats next
Promising avenues
References
Paper
Wrapup