search

wildtreetech / advanced-comp-2017

💻 Material for a course on applied machine-learning for scientists. Taught at EPFL in spring 2017
23 stars 13 forks source link

Reservoir computing #9

Open betatim opened 7 years ago

betatim commented 7 years ago

http://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1004967 via @sharkovsky

This thread is to collect links and discussions related to the ideas in the paper.

First thoughts: contextual bandits, reinforcement learning (in particular A3C), and the various "learning to learn" approaches.

betatim commented 7 years ago

A good intro to "Learning to learn": https://hackernoon.com/learning-policies-for-learning-policies-meta-reinforcement-learning-rl%C2%B2-in-tensorflow-b15b592a2ddf

sharkovsky commented 7 years ago

This seems like a very thorough review of reservoir computing methods:

http://www.sciencedirect.com/science/article/pii/S1574013709000173

betatim commented 7 years ago

Non paywall-ed link http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.470.843&rep=rep1&type=pdf

betatim commented 7 years ago

how important is the reservoir computing aspect to you? There has been a lot of work on sequence-to-sequence models (with recurrent NNs) mostly using LSTMs and GRUs. This is the first time I've seen reservoir computing as a term.

For example learning to go from a sequence of digits with an operator in between to a sequence of digits that is the answer (sth like http://betatim.github.io/posts/algebra-from-scratch/) or time series data, etc. But I don't know how much of this is based on the engineering point of view: if it get's the job done we don't care how it works vs trying to learn about neuro science.

sharkovsky commented 7 years ago

Hi,

I'm actually a PhD in neuroscience (computational neuroscience to be more precise). The reservoir computing aspect is pretty important to me, because it would allow me to relate it to my everyday research. Also, I wouldn't have to learn much more about neuroscience because, even though I know very little, it is enough to have a decent idea of what they're talking about.

Obviously, I have to keep in mind also that this project can't be a pain for you, so if you think it would be too much work I'd be willing to switch to an "artificial" model such as sequence-to-sequence using GRUs or similar.

However, from the point of view of my research, the reservoir computing aspect is a pretty interesting topic that I'd like to explore!

betatim commented 7 years ago

Loopy neural networks: http://cs231n.stanford.edu/reports/2016/pdfs/110_Report.pdf a course project form Stanford that might be relevant.