HBP NRP or Oneshot character recognition project

Hello all,

I am a third year student doing my Bachelor in Computer Science and following Numenta since almost a year.
Currently I am working on my degree project and I have troubles to decide which direction I should focus on. My supervisor can not help me much since he is not familiar with the topics, so I thought maybe some of you can give me guidance.

I decided to work close with HTM-theory but my focus wasn’t completely set so I am still exploring by reading a lot of papers and doing online courses to acquire background knowledge.
However I need to decide now on a direction to stay in the given time constraints.

Mainly the choice is between two paths (listed with some of the pros/cons):

(1) As I attended to an HBP conference last week and got introduced to their Neurorobotics platform, I could try to implement HTM as Spiking Neural Network and experiment on their platform.

  • (+) They already provide a platform with ready experiments
  • (+) It would definitely be useful to have such a bridge from HTM and HBP as there exist almost none yet.
  • (+) I could build on top of the Numenta Paper to port on the Heidelberg BrainScales Project as it was implemented in PyNN
  • (-) Coming from Computer Science I do have relatively little knowledge about electrical circuits and neuroscience, huge learning curve would have to be made and e.g. Neuron Models would need to be completely based on the Numenta Paper
  • (-) HBPs NRP does not provide very useful analytical tools for HTM yet (e.g. you can not see connections between neurons only spiking behaviour)
  • (-) It adds complexity introducing SNNs to an architecture that is already nicely reduced to binary operations.

(2) Trying to use HTM combined with RL for One-shot character recognition on the omniglot dataset (Paper “Building machines that think and learn like humans” - Numenta wants to comment on it now) and evaluate with the stated criteria.

  • (+) An architecture combining HTM with RL could be very useful
  • (+) Already many discussions and resources on the forum. Including the master thesis for an autonomous agent.
  • (+) Simplified complexity in contrast to SNNs
  • (+) Less Neuroscience and Electronics needed.
  • (-) Experimental setup has to be designed by me. The results might not be comparable if the experimental setup is different (e.g. designed for an autonomous agent).
  • (-) If I want to continue in that field in the future it would be useful to learn more from the neuroscience/electronics (?)
  • (-) Evaluation with the stated criteria might be very hard, as they are relatively general.

Additionally I planned to study either “Autonomous Systems”(CS heavy) or “Neural Systems”(Neuroscience heavy) after the summer for Master studies. Which fits well to the decision…

I would be very thankful for some viewpoints from the experienced community here. I’m sure some of you know this kind of difficult situations.

Kind regards


Sounds like a great, well thought out plan!

I know the Heidelberg group is quite interested to continue the HTM work, so they would be very interested if you want to take on (1). You could continue with PyNN and not fully implement it on their hardware, which avoids the electronics. A robust continuous time implementation of HTM using spiking networks would be quite cool.

(2) would be very interesting as well. IMO it will require more hard-core understanding of both RL and HTM, and perhaps the neuroscience of reinforcement learning / dopamine / etc. You would be more on your own here I think, but might learn quite a bit.