Message Boards Message Boards

Find derivative of a network with NetPortGradient?


There's a function NetPortGradient, but I can't figure out how it works.

To give a specific example (with poetic license of the notation), consider the image

x = ExampleData[{"TestImage","Flower"}]

The probability <flower|orchid>[x] in the ImageIdentify network is 0.182434

Then a Newton quotient can be defined:

y[[row, column, channel]] = Limit[h -> 0, (<flower|orchid>[ReplacePart[x,{row, column, channel} -> x[[row, column, channel]] + h]] - 0.182434) / h]

That is, the functional derivative of the identification probability with respect to each pixel.

How can this image be found?

POSTED BY: Collin Merenoff
1 year ago

Group Abstract Group Abstract