I am focusing on the output of the third layer in face recognition
So, the hidden layer neurons are accumulating the parts of the face, as they are using different parameters and different subsets of the features, so it is expected that pixel intensities, as the activations of the neurons, would not be the same, and also that the appearance would not be exactly the same. That’s why the difference is visible in the two marked square boxes in the second attached image, I think. Am I wrong?
Waiting impatiently for response.
Hi @farhana_hossain,
That sounds like a pretty good intuition!
A minor thing here is that, depending on the network’s configuration, neurons don’t always only add up features from the previous layer. The weights can be negative, right? However, I don’t think this should invalidate your intution. It’s a pretty good one!
Cheers,
Raymond
1 Like