Neural Network Errors

For my M.Sc. in Intelligent Systems, I had to present the practical issues of neural network training — the material drawn from Goodfellow, Bengio, and Courville's Deep Learning. I could have made slides. But the whole point of these failure modes is how they behave, and behavior is exactly what static slides can't show.

So the presentation became an interactive web page instead: the audience watches curves diverge and weights collapse live, and the artifact keeps working as a study reference long after the class ended.

The five failure modes every practitioner eventually meets: underfitting, overfitting, vanishing and exploding gradients, convergence difficulties, and local or spurious optima — paired with their remedies: L1/L2 regularization and dropout, parameter sharing and convolutional structure, early stopping, deeper architectures, and better optimizers.

Each concept has a manipulable visual: a polynomial-fit playground where you crank model complexity and watch it swing from underfit to overfit, live training-versus-validation error curves, weight distributions morphing under different regularization modes, a convolution filter-sharing demo, and a loss-landscape view that builds intuition for optimizer behavior, local optima, and saddle points.

The entire site is a single self-contained index.html — no build step, no framework, no dependencies. That was a presentation-day constraint turned feature: it opens from a file:// URL, survives a venue with no Wi-Fi, and will still render unchanged in a decade.

Presented for my Master's coursework and now live on its own subdomain, where it doubles as the revision reference I wish I'd had before building my undergraduate CNN project.