Deep Learning

By leveraging MTL solutions for negative transfer, we improve the training dynamics of variational autoencoders (VAEs), boosting their performance on heterogeneous datasets.

By using relative gradients we propose a new approach for exact training of neural networks where the log-determinant of the Jacobian appears in the loss function as it happens, e.g., in normalizing flows.

We test the Causal Feature Extractor in a series of different scenarios (text/image/audio) and show its great versability with minimal fine-tuning.

We tackle the problem of text normalization using a character-level encoder-decoder network which uses a novel CNN network designed to better use attention layers.

An example of using the in-built project page.