WebGL MNIST learning improved

We have improved the previous MNIST implementation in order to make this implementation faster than the Python/Numpy one. We still use our WGLMatrix library to process all linear algebra computations.

The full learning may take a few tens of minutes.

Controls :

The whole dataset may be too large for you graphic card memory. If it overflows, the WebGL context is killed. You need to refresh this webpage and to try again with a reduced version of the dataset.

START WITH THE WHOLE DATASET (50000|10000) START WITH THE HALF DATASET (25000|5000) START WITH THE QUARTER DATASET (12500|2500)

Network hyperparameters :

layers : activation function : connectivity :

Learning parameters :

number of samples for training : number of samples for testing : number of epochs : minibatch size : eta :

Monitoring :

Status :
Logging console :

References :

• Chapter 1 (Using neural nets to recognize handwritten digits) of Michael Nielsen's online book, Neural networks and deep learning : It details the neural network used,
• Direct download link to the MNIST neural network written with Python/Numpy on which this example is based (MIT license).
• We use JSON converted MNIST data from ConvnetJS repository (MIT license)