Future Work
- I really do think that the best way to get to Turing test passing AI
is to follow the human model.
- That is, use neurons, be an agent in an environment, perform
psychologically accurately, be domain general, and take a lot of
time to learn (years though simulated years are ok).
- The problem is that's a big ask.
- So, we're stuck working with relatively small systems and scaling up
slowly
- Maybe a big company will pick this up. The HBP seems to be dead now.
- Problems with lots of neurons:
- Fix up spike source model
- Use neuromorphic systems (SpiNNaker)
- Just figure out how to run Nest or a simulator on GPUs
- Use the NEST supercomputers from the HBP.
- Figure out how to manage this with a more reasonable neural model of
adaptation.
- Integrate with Cell Assemblies, and our earlier work with Hebbian
Learning for Categorisation (Huyck and Mitchell 2014).
- Make neuropsychological models of learning.
- Integrate the learning with an agent. (We've got some good spiking
agent stuff with the NEAL system.)
