ericjang / RNN-dynamics

Code and report for APMA136 Final Project

RNN-dynamics

Code for my APMA136 Final Project, based on "Opening the Black Box : Low-Dimensional Dynamics in High-Dimensional Recurrent Neural Networks".

Codes for simple2d, threebit_flipflop, motor_control tasks are found in respective subdirectories. All code written by Eric Jang, under the MIT license. Most of the code is implemented in MATLAB, although there is a bit of python here and there.

Trained Submanifold Attractor

3-bit Flip Flop

4-bit Flip-Flop

4-bit Flip-Flop, Logistic Nonlinearity

4-bit Flip-Flop, pre-trained on 3-bit FF

Formation of Phase Space from Training

3-bit FF

4-bit FF

Formal, academic writeup and informal blog posts are coming soon.