A toy viz of fully connected neural networks
Features
- Neurons are sorted by their similarity
Limitations
- It works for networks that intend to solve a scalar input, scalar output (i.e.
$f: \Re \rightarrow \Re$ ) regression/classification problem. - It works for small networks, i.e. ~4 layers, up to ~50 neurons per layer, due to svg render limitations.
Usage
Open index.html for a demo.
Detail
dump.py
is a demo script that train a network, record and dump intermediate data (e.g. intermediate training weights/biases, layer activations, loss values, optimized layouts) into a js file (with var data = [...list of intermediate states...]).
index.html
then loads data.js and contains scripts for visualizing all activation levels at each training stage.