Once you have a gate t, the layer's output must mix the transform h and the untouched input x. The highway paper writes this as a convex blend, which is what lets the network route signal straight through when that helps.
Implement highway_combine(x, h, t), which returns h * t + x * (1 - t).
t is in [0, 1].The gate weights the transform and 1 - t weights the carry path.
No results yet
Run your code to check it against the examples, or submit to run every test.