Hidden layer

A neural network turns an input vector into an output vector in stages. The stages in between, which neither take the input nor give the answer, are the hidden layers. “Hidden” only means you don’t see their numbers from outside.

A neuron

Each unit in a hidden layer, often called a neuron, computes one number from the whole previous layer $x$: a weighted sum plus a bias, passed through a nonlinearity $\sigma$:

$$h_i = \sigma\!\left(\sum_j W_{ij}\, x_j + b_i\right)$$

The weights $W_{ij}$ and bias $b_i$ are learned. A layer of 512 neurons gives 512 such numbers, its activations, which the next layer reads.

What a neuron means

Nobody assigns a neuron a job. Training adjusts the weights so the network’s output improves, and the neurons end up with whatever roles help. Some respond to one clear thing; many respond to mixtures. Finding out which is the subject of interpretability.

In this project

The student distill9-196.6M has two hidden layers of 512 ELU neurons between its 465 inputs and 48 joint commands. In the torque and neuron post each neuron of the second layer is drawn as a row of the raster, lit when its z-score against quiet standing is over 3. Standing, none is lit; at the peak of a technique 179–232 are.