AI summary by TwinMind
New computer will mimic human brain -- and I'm kinda scared
Sabine Hossenfelder
6m 37s
Neuromorphic Computing Overview
- New supercomputer "Deep South" being built at Western Sydney University
- Part of neuromorphic computing paradigm - computers modeled after human brains
- Different from AI - two separate research areas with different goals
- Both use neural networks but neuromorphic computing integrates hardware structure with algorithm
Traditional vs Neuromorphic Computing Differences
- Traditional AI: neural network structure represented in software, physical basis still transistors on microchips
- Human brain: network is physical thing, unified with algorithm
- Neuromorphic computers: built to resemble human brain in hardware structure, not just software algorithm
- Not built from biological tissue - connectivity and functionality reflected in processor structure
- Still looks like box with wires but works differently than phones
Current Neuromorphic Research Examples
- Basic research: throwing wires together to create networks
- Some success teaching wire networks to recognize handwritten numbers by adjusting junction conductivity
- IBM's True North processor (2014):
- 1 million cores designed to resemble neuron firing
- Maps virtual neurons onto physical ones
- IBM has updated/improved since but still far smaller than human brain
Deep South Computer Specifications
- Named after location and nod to Google's DeepMind and IBM's True North
- Uses field programmable gate arrays (FPGAs)
- Small circuits whose function can be electronically changed (unlike normal microprocessors)
- Will simulate neuron spiking - sudden threshold at which neurons pass information
- Built-in randomness in artificial neuron behavior (important but not well understood role in human brain)
Research Goals and Energy Efficiency
- Goal: understand how human brain operates on minimal power
- Human brain: barely needs 20 watts to operate
- Current AI training: consumes huge amounts of energy, expensive to run
- Mapping neural networks from software to hardware could dramatically decrease energy requirements
Processing Speed Comparisons
- Human brain neurons: few milliseconds to update state (chemical reactions)
- Modern computer processors: update billions of times per second
- Processors ~1 million times faster than brain neurons
- Need slow electronics to physically resemble human brain
- FPGAs chosen because they are slow
Project Timeline and Future Implications
- Completion planned by April (current year)
- Will be remotely accessible for research
- Expected long waiting list for access
- Potential future: device working like human brain but million times faster
- Performance: 228 trillion operations per second
