Requests for Research

Open RFRs 10

Continual Few-shot reinforcement learningHippocampal replay for continual RL Preprint Implementation
Left/Right brain in an RL agentTwo hemispheres
Left/Right brain, the hippocampus and episodic learning in AI/MLTwo hemispheres
Left/Right brain, human motor control and implications for roboticsTwo hemispheres Preprint Implementation
Active VisionActive perception
Self-organizing neural network hierarchies Part 2Self-organising architectures
Using Reinforcement Learning to discover attentional strategiesLearning without backpropagation through time (RSM)Masters thesis
Long-range navigation without perfect information using local reinforcement learning rulesPredictive models and navigation
3D agent test suitePlatform createdPlatform (PyLIS)
Decision-MakingExecutive controlImplementation createdImplementation

Past RFRs 4

Improving episodic memory with disentangled representationsPosted 2020.01Episodic memory and continual few-shot learningMasters thesis
Left and Right Neural Networks, inspired by our bilateral brainsPosted 2018.09Two hemispheres Preprint
Self-Organising ArchitecturesPosted 2018.09Self-organising architectures Conference paper
Improving Memory with a Hippocampal ModelPosted 2018.06Episodic memory and continual few-shot learningReport incomplete

Taking one on

The RFRs here are jointly compiled by Cerenaut and the Whole-Brain Architecture Initiative (WBAI). Our shared objective is to foster research into general intelligence. They are selected to encourage research that we believe will help to realise human-like Artificial General Intelligence (AGI) via the whole brain architecture approach, which assumes that the path to AGI can be shortened by mimicking the macro-architecture of the brain.

The RFRs are substantial projects and we expect are best suited to individuals or groups with the following knowledge and skills:

Ability to program (e.g. simulations, technical models)
At least basic machine learning (ML) experience
Basic knowledge on the architecture of the brain (neuroscience)

But you don’t have to worry. For learning ML, there are ample materials on the Web. To get some background in neuroscience, you can start with CCNBook.

More importantly, you won’t work alone. You are encouraged to contact us before you start working on the requests. Then you’ll be part of our community where you can get support including resources (such as software and data) and advice. We can also help to link you with like-minded researchers so that you can attack these problems as a team.

If you start work on one of our RFRs, please provide a link back to our page(s) when you publish your results.

Contact

Public forum (for discussion, team-building and support): please use reddit, r/BioAGI.

Other inquiries: please send your message to rfr [at] wba-initiative.org.

FAQs

If I work on the RFR, who owns the intellectual property (IP)?: You (and your team) own your work and full rights to your IP. We publish these RFRs to support and guide research into areas we find interesting. If you wish, we can help to promote your work.
Am I free to publish my work?: Of course, anywhere you like. We can provide advice on write-ups and suitable journals or conferences.
Which programming language should I use?: Since most machine learning frameworks nowadays use python, we recommend python. But if you would like to use another language, please consult us.
Which machine learning framework could I use for these RFRs?: If you don’t have a preference, we are encouraging standardization around TensorFlow or PyTorch. Otherwise, please consult us. We accept submissions in all frameworks. Besides, we also offer a ROS-like modularization framework to encapsulate multiple frameworks.
Can I get a (cash) reward by doing one of the RFRs?: No, but if your research is good, we’ll tell people about it. And you may want to apply for an annual WBAI Incentive award after publishing your research.
Should I make my code open-source?: It is up to you, but if you would like it to be recognised in the research community, making it open-source would help, and that’s what we expect.