NEAL: Neuromorphic Embodied Agents that Learn
- This work came at the end of the NEAL project, part of the HBP.
- Aside from a few agents and some exploration of work with BrainScales,
this came down to making extensible components for Nest and SpiNNaker.
- We made components for natural language parsing, natural language
generation, planning, vision, simple cognitive mapping, and
associative memory.
- Most of them were data driven, so you could specify (for instance)
the lexicon and grammar, and it would generate a neural topology
to parse (or generate) the text.
- There were some problems handshaking with the neurorobotics platform,
but in general they worked.
- Though vision, planning and cognitive mapping used some other
neural mechanisms, most of the work came through
binary cell assemblies.
- With these you could implement general finite state automata with
10 neurons per state.
- A slight modification enabled you to make sequences.
- This all fit into the PyNN, Nest, SpiNNaker system and it was
pretty easy to make a system that had both a Nest and a SpiNNaker
version.
