Felix' Workshop

Neuron Bench

A neuron is a function that multiplies each input by a weight, adds them up with a bias, and puts the total through one bend. That is the whole thing. Everything a network can do comes from stacking that.

This is a bench rather than a course. Pick a set, build a network over it, train it, and then take it apart: every control stays live, so you can add and remove neurons, change the bend, break it on purpose, put one observation through by hand, and click any unit to see what it is doing.

One line per weight. The thicker it is the larger the weight; blue is positive, red is negative.
0
0
Training loss against held-out loss. When the two separate, the network has started memorising rows instead of learning the pattern.

A loss is only a score against the same question, so a record is kept per set together with the three things that decide what the number means: whether the inputs were normalised, how much was held back, and the seed that chose which rows those were. Change any of them and the bench starts a new record rather than pretending the old one still applies. Nothing leaves this browser — there is no scoreboard behind this page, and clearing the site's data clears the book.

Put a single row in and watch it go through. Each hidden value is what that unit answered for this row alone, and the last line is what the network made of them. Click any of them to pin that unit above.

CSV, one row per example. Every column but the last is an input and must be a number; the last column is the answer. A numeric last column with many distinct values is treated as a regression, anything else as classes. Nothing is uploaded — the file is read in this tab.