Hi,
I understand that you want to know which autoencoder training method to use for non-image autoencoders.
The first line (autoenc1 = trainAutoencoder(train_data1, hiddenSize1, 'MaxEpochs', 400, 'L2WeightRegularization', 0.004, 'SparsityRegularization', 4, 'SparsityProportion', 0.15, 'ScaleData', false)) creates an autoencoder with a specified 'hiddenSize1', which represents the number of neurons in the hidden layer. It also uses L2 weight regularization and sparsity regularization to control the learning process.
The second line (autoenc1 = trainAutoencoder(train_data1, 10000, 'MaxEpochs', 400, 'DecoderTransferFunction', 'purelin')) creates an autoencoder with 10,000 neurons in the hidden layer, and it uses 'purelin' as the transfer function for the decoder. This configuration doesn't include regularization terms like L2 or sparsity regularization.
Both the above methods can be used for non-image autoencoders, if you want to encourage sparsity in the hidden layer representation and regularize the weights, the first option with sparsity and L2 regularization can be useful. If you don't require sparsity regularization and prefer a larger hidden layer, the second option with 10,000 neurons might be suitable. The 'purelin' transfer function is a linear function, which can be useful for certain types of data.
To learn more about “trainAutoencoder” usage and syntax, you may refer to the MathWorks documentation link below: -
I hope this helps!
