Now let’s nonlinearize our feature extraction algorithm.

First, we can apply an activation function, e.g. logistic function, to the scores obtained through the weighted sum for nonlinear transformation.

Then, we can repeat the above process on these nonlinear outputs to further nonlinearize the outputs.

After obtaining all the feature extraction results, we begin the process of image reconstruction.

Since the features (PCs) we extracted are nonlinear, we will also use a similar approach during the reconstruction process to achieve a good reconstruction outcome.

Congratulations! You have unlocked the nonlinear feature extraction model, known as the AutoEncoder.

Yes, just like encrypting information, the left half of this model represents the encryption part, which we call the encoder part, while the right half represents the decoding process, known as the decoder.

In other words, if I want to secretly transmit an image to someone, I can encrypt the image through encoder for transmission. Only the person with the decoder can access the true content of the image.