I want to build HELM neural network that consists of autoencoder (AE) and one class classification (OC).

HELM with AE and OC have following shape:
enter image description here

That is, hidden layer output of AE is input of OC.

Training of HELM consists of training AE and OC separately. In order to train each neural network in HELM, first there are generated random weights and biases between input and hidden layers, and then based of them and activation function (for example sigmoid), only weights between hidden and output layers are trained. But then what's the point in training weights between hidden and output layers in AE, since only output of its hidden layer is provided as input into one-class classifier? What is point in use AE in HELM if weights between input and hidden layers of AE are basicaly random?

Following paper (page 6): https://arxiv.org/pdf/1810.05550.pdf

confirms that output of hidden layer of AE is input for OC, but also on the contrary in Algorithms 2 and 3 (pages 6, 7), there is shown that input of OC is AE input vector multiplied by matrix of weights between hidden and output layer, what sounds weird for me.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.