Making a library-less MLP

I’ve made a Multi-Layer Perceptron (MLP) in C++ as a learning project. The goal, initially, was to make a replica of the MLP described in 3Blue1Brown’s lectures, with no reliance on ML libraries.

What I’ve made achieves that. It trains from the MNIST database for 25 epochs/until a loss threshold is exceeded and then predicts some numbers from the test set correctly.

It’s Not that Good

As of right now, it is quite bad. Size is kept small because I’m running it on a laptop, all of the matrix multiplications are implemented with for loops and nothing is parallelised.

The goal is now changing from making an MLP without libraries to making a good MLP (hopefully) without libraries.

Sources

• Link to the GitHub repo
• 3Blue1Brown’s lecture series
• Eigen