Issue about Autoencoder Network Architecture

I'm having a problem with the autoencoder. I have both the encoder and decoder transfer functions for the network set to logsig. according to the formula that autoencoder runs on, the output should not be negative. However, this does happen, can you answer my query?

[jiajun]

Maybe you can provide detailed formulas for the encode and decode functions, as per the example file the calculations are not consistent with the actual run.

[Shreeya]

logsig activation function returns values between 0 and 1. Can you please elaborate why is the expected output a negative number, since the equation as shown in the image should return a positive number ?

[jiajun]

From a theoretical point of view, the activation function of both the encoder and the decoder is logsig, with no negative output values being possible. However, this does happen in my program (both input and output values are negative). So I would love to know why the decoder is still able to output negative values when using the logsig activation function

[Venu]

I think you have to review your network architecture, layer connectivity. Inspect the properties of each layer to ensure that the correct activation function is assigned. You can try checking Initialization of weights and biases, verify whether your input data is scaled (Normalization) properly to work well with logsig function.

Answer 1

[Ayush Aniket]

Hi jiajun,

As per my understanding, the output from your autoencoder is negative even after setting the encoder and decoder transfer functions to 'logsig'. The reason for this occurrence is scaling of data. Since you mentioned that your input data is negative, to apply 'logsig' the 'trainAutoencoder' function scales the data to the range of the decoder activation function. This is a default 'true' parameter. Now, if the input data has been scaled, the output data is also scaled back to the initial range by the 'predict', 'encode' and 'decode' methods.

Please refer to the following documentation section link to read more about the 'ScaleData' input argument and the process of scaling:

https://www.mathworks.com/help/deeplearning/ref/trainautoencoder.html#buxdjrt-ScaleData

Hope it helps.