Features n
Hidden m
< n
Sparsity S
Activation
ReLU
Linear
Training
↻
The network
Test input: click to toggle. Updates live as W trains.
Computation
Hidden space
Shape
Training data
W, WᵀW and b
Toy Models of Superposition Demo
Yan (Stella) Si · BU CDS PhD
Based on
Elhage et al., 2022