Interest in biologically inspired learning and reinforcement learning #42
Replies: 3 comments
-
Here's the image to make things a little easier (you can paste an image directly into the textbox). I would be interested in RL, and I think the environment looks good!
What kind of learning algorithm are you referring to? Some kind of Hebbian rule? |
Beta Was this translation helpful? Give feedback.
-
I love this idea! Definitely go ahead and try it! |
Beta Was this translation helpful? Give feedback.
-
I will divide my comments in 2 sections gradient and Hebbian: Gradient: I think applying RL to gradient methods is straight forward. I have familiarity will RL especially proximal policy optimization (PPO). The RL task in the above image might sound simple, but it can lay the ground work for tougher RL tasks like conjunctive learning of what and where using a 3D-moving sound sources for a moving agent. The environments can be built in the unity game engine. Aside from sound localization, RL can work with surrogate gradients to train agents with different kinds of input, visual, acoustic, self-generated etc... Regarding delays, I think, it is possible not to use them with RNNs trained with backpropagation through time? Hebbian: I am a little biased towards Hebbian (I like the beauty of causality). I think Dan said that STDP for sound localization is well studied, but (I hope what I am saying is not dumb), as far as I saw, the methods used were supervised learning? Like the authors trained on a predetermined times for output classes. What I thinking of is, an input-robust unsupervised or self-supervised model. That can use arbitrary sound sources with the MSO frequency cap. If this is also done, we can learn conjunctive representations of what and where, where a middle neuron fires a code representing what is the source and where, then this code could can be decoded with separate neurons for location and object. |
Beta Was this translation helpful? Give feedback.
-
This might have some redundancy with my previous question.
But is there any interest to add to this work biologically inspired learning maybe trained with reinforcement learning?
I have attached the image of a simple sound localization environment: https://ibb.co/GcmCcmj
I think, an even easier environment can be made with a circle.
I think, compared to visual tasks the input space is smaller, which is proportional to the number of frequency channels? I mean the input space will also depend on the desired frequency resolution.
I also believe an advantage of adding a biological learning section is to compare it to surrogate gradient methods.
Beta Was this translation helpful? Give feedback.
All reactions