Introduction to Neuromorphic Engineerings

Graduate course at ETH, ETH, 2023

This lecture introduced the core principles of neuromorphic engineering, focusing on the design of real-time, adaptive, and energy-efficient systems that emulate the brain. It explored how sparsity, locality, and event-based computation underpin biological efficiency and how these are translated into hardware. The talk presented neuromorphic sensors, analog and digital implementations, and in-memory computing as enabling technologies. Special emphasis was placed on training and architecture strategies for event-driven systems, including the Mosaic framework and delay-based SNNs. Finally, it highlighted how physical dynamics and enriched computational units can enhance neuromorphic processing.

Find the lecture’s slides at this link

Key points from the lecture

  1. Vision and Motivation
    • Goal: Build intelligent, adaptive machines that close the sensory-motor loop in real time.
    • Key challenge: Edge computing under severe constraints (power, memory, latency).
    • Strategy: Learn from the brain’s mechanisms—sparsity, locality, dynamics, and adaptation.
  2. Sparsity and Event-Driven Paradigm
    • Sparsity in time and space: Only respond to meaningful events.
    • Neuromorphic sensors: Vision (DVS), audio (cochlea), touch, olfaction, etc.
    • Event-driven processing: SNNs that compute only when spikes occur, reducing power and computation.
  3. Neuromorphic Hardware Implementations
    • Digital vs Analog: Digital offers precision; analog exploits physical properties for ultra-efficiency.
    • Local computation: MAC operations using RRAM crossbars for in-memory computation.
    • Event-based In-Memory Computing: Mosaic architecture with small-world connectivity and local communication.
  4. Training and Optimization
    • Hardware-aware training: Penalizing non-local connections, RRAM-aware quantization (using STE).
    • Delay-based SNNs: Using spike timing as a computational variable.
    • Online learning strategies: Local, sparse updates to preserve memory lifespan and reduce energy.
  5. Physical Dynamics & Rich Computation
    • Temporal processing: Matching time scales between hardware and signal dynamics.
    • Analog substrates: Using RC time constants, volatile memory, and dendritic delays.
    • Enriched neuron models: Moving beyond ReLU to multi-compartment spiking neurons with temporal logic.