Neuromorphic computing takes inspiration from how brains compute: many simple units, local memory, communication by brief spikes, and work done only when something happens. The aim is very low power and fast reaction on sensory tasks, by departing from the separation of memory and processor used in conventional computers.
Key concepts
- Spiking neurons: units that integrate input over time and emit a spike when a threshold is crossed.
- Event-driven computation: activity occurs only when there are spikes, so idle parts use almost no energy.
- Synaptic plasticity: connection strengths change with activity, as in spike-timing-dependent plasticity.
- In-memory computing: storing weights where they are used, so data does not move back and forth.
- Event sensors: cameras and microphones that output changes instead of full frames.
Vocabulary
- Spike train: the sequence of times at which a neuron fires.
- Leaky integrate-and-fire: the simplest common neuron model, a charge that leaks away and resets after firing.
- Von Neumann bottleneck: the cost of moving data between separate memory and processor.
- Rate and temporal coding: information in how often neurons fire, or in exactly when.
What to study, in order
- Basic neuroscience of the neuron, synapse and action potential.
- Simulate a leaky integrate-and-fire neuron and plot its response to different inputs.
- Build a small spiking network and train it with a local learning rule.
- Compare it with a conventional neural network on energy and accuracy for a simple task.
- Read about event-based sensing and how analog and digital neuromorphic chips differ.
Related tags
#computational-neuroscience #deep-learning #computer-architecture #analog #edge-computing