hi fellow neurohackers ;)
...maybe you could see that coming, I'm working on my diploma on "ALife
agents' Behavior with HTM/CLA - NuPIC";
>From my schedule, I'm nearing what I(more like we) could have done
programming-wise, and I'm writing my text now. I am posting the outline
here, would be glad if you could give it a thought and let me know:
-some comments/suggestions on the structure/ordering?
-some topics I definitely must not miss out
-code, benchmarks,pictures,demos, videos are always welcome!
-anything else I'm missing, doin' wrong?
Please try to keep it to the Requirements which I'm bound to focus on.
This help is really appretiated!
Thanks a ton,
Forever yours, breznak ;)
PS: forgive the scarcity and cryptic sentences, it's just my notes to "kick
in the thought.."
===Architecture of Autonomous Agent Based on Cortical Learning Algorithms===
Instructions & requirements:
*) Study the fundamental principles of Cortical Learning Algorithms (CLA)
inspired by mammalian brain.
*) Modify these algorithms to be able to produce also behaviour, aside of
learning.
*) Implement CLA (or augment a current implementation) with ability to
produce behaviour. This
implementation should support Robotic Operating System (ROS) communication
and should be as domain
independent as possible.
*) Compare the efficiency of resulting learning and behaviour of agent
controlled by this modification of
CLA with another today used learning and decision making techniques.
========================
Abstract
Intro
Principles of mind & brains:
-intro
-approaches
--neurolevel - LIF, Hodg-Huxley, BlueBrain initiative, ..,
POV-synapses/neurons/reqions/knowledge
--psychological - memories, emotions, un/supervised learning
--computional & hw - assumptions, domains, simplifications, HW (memristor),
processing power exp..
--philosophical - what are memories, can machines think?/learn?/sense?, how
do we learn, perceive?
--future prognosis - AI domains conquered, being tackeled, still miles
away; IBM/Watson, (big)data-mining, memristors,
Theory of HTM/CLA
(resources mostly Whitepaper, OnIntelligence, ML, videos, Fergal's blog)
-SDR
-CLA region
--column,cell,synapse
--SP
--TP
-HTM
--links
Producing behavior:
(resources mostly Vitku, Kadlec, psychology)
-action learning - sensomotory behav - SDR-pattern-matching, compare with
planning (LISP)
-memory - short term, long term, compare with other AI approaches to
achieve that (LTST-Mem), HTM sequences, ?Q "how are memories stored in
brain?", neuron fields, deepnets
-emotions! - goal, drive, implementations (low level hardcoded),
-forgetting - automatic in CLA region, missing values, outliners,
-attention - anomaly detection! - "staring at crippled ppl",
-abstraction
-high level concepts
-language - CEPT, current SotA
Implemention of CLA - NuPIC
(resources ML, community)
-my former experiments (htm java)
-why i ended up with NuPIC:
--comunity
--history
--implementations - other htm/cla-like projects
--projects! - hackatlons, all the cool stuff ppl do!
--ROS! - impl, connection to other research
--domain indipendence! - encoders, parameters (swarming), areas of what can
be done (see projects), what is problematic/incomplete currently
--things left TODO -hierarchy,vision domain, strip down impl,
--tools - swarming, cerebro, serialization, benchmarking, VMs
Compare efficiency of CLA-based learning and behavior with another AI
technique used for that matter
(help, any practical benchmarks??)
-RL, NN, SOM, RNN, prerequisite&effect based programming(LISP),
-benchmarks - CLA resistance to noise (SDRs)
- sequences
- ability to generate behav in a natural sense
-where is CLA ruling at the time?
-where is it losing?
Conclusion
-what has been achieved
-what's left
-possible future research directions
Appendix A - Demos
"a (theoretical) proof-of-concept examples to sections above"
--
Marek Otahal :o)
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
