Future Work
- CABot3 is currently in early stages of developement. It will include
a
- neuropsychologically motivated parser with good PP attachment,
vision recognising five objects, and using figure/ground separation
(probably by texture), bigger CS game, labelling, learning
goal-actions in a different CS game, and cognitive mapping.
- We hope it will also include, arbitrary Maes nets, learning goal
fullfilment with two or more actions, PP attachment via use, and
compound visual objects.
- CA Mark 3. This work is inspired by Hebb's Cell Assembly hypothesis,
the neural representation of concepts are reverberating neural
circuits. Our work currently depends mostly on binary CAs, either
on or off. However, CAs should be more flexible. We would like
to implement this theoretical CA as a better basis for neural
concepts.
- CABot4 to 6: are the milestones for the nascent follow on grant.
- We propose a conversational agent, that uses better CAs, to learn
the semantics of a more complex virtual environment. This will
also involve better vision, and better learning.
- Neural Cognitive Architecture. A parallel goal to the CABot agents
is to build an architecture for modelling neural cognitive behaviour.
- Currently, the neurons work, but there is no unifying framework for
building cognitive agents. Reusable subsystems will need to evolve.
- Once the architecture is there, it should be possible to
start learning a range of domains. The system may then
pass the Turing test.
