CPC scores each candidate future with a bilinear function of the context and the future latent. The matrix W is learned, and it lets the score depend on the relation between the two vectors, not only on their dot product.
Implement bilinear_score(c, W, z), returning c^T W z.
W need not be square.Multiply the context by the matrix, then dot with the latent.
No results yet
Run your code to check it against the examples, or submit to run every test.