Autoencoders can be used to generate a fixed-length feature representation of the input data. Now consider the case that your input data (e. g. videos) has unequal length. So you would train autoencoders for your individual data points in order to engineer the features of your input data.

Is that a "valid" approach in terms of stochastic properties? Because I am wondering whether this reduction from e. g. 2 minutes material to a fixed number of features compared to a video of 30 seconds is valid to generate a reasonable model. What would be alternatives to this approach?


It's valid in the sense that you can do it. In fact the most basic sequence to sequence translation / autoencoder models do the same thing -- encoding a variable length sequence in a fixed size latent space. It's not too different from setting a video encoder to produce a fixed size output. You are correct that too much variation might produce weird results. This is usually mitigated by attention mechanisms.

Alternatively, with a sequence to (latent) sequence to sequence model, you could produce a variable length latent space.

| cite | improve this answer | |

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.