Hardware Issues
- I'm sure these issues are relevant to SpiNNaker and I'm pretty
sure they're relevant to BrainScales (particularly if you want
to do agents).
- Runtime I/O: the more spikes you can pump onto the boards and the more
you can get from them, the better. It also helps if they're regular.
We could also build better mechanisms on either side of this link to
reduce the need for the communication, and to support reasonable
behaviour when things are irregular. For instance, we could
retain the last picture in the input neurons on the board.
- Loadtime retention: it takes a long time to load reasonable sized
topologies. It would be good if they or parts of them could be
retained to reduce new load times.
- Saving plastic synapses: if the agent runs for a long time (an hour)
and learns that whole time, the learning is saved in plastic synapses.
It would be good to retain that learning.
- Runtime saving of plastic synapses (backing up in case of crash).
- I don't think saving the current state of neurons is necessary (monkeys
fall out of trees).
- Generating new neurons and synapses at run time could help. Similarly,
removing
neurons and synapses would be useful.