This year I was at EUSIPCO to present the paper that wraps up our work on upsampling artifacts. In the following website you have a summary of our work, including a video “for dummies” with the main ideas:
Five ideas I found interesting:
- From the encoder features one normally reconstructs the target signal for source separation. For targeted source separation, it is interesting to also enforce the output of the encoder to be disentangled source embeddings with a triplet loss (Eisenberg et al.).
- If we can synthesise each source with a generative model, with “smart” sampling strategies one can perform source separation. Think of this approach as synthesis-based separation (Villasana et al.).
- For some problems, the number of output classes of our neural network can be huge. For example, our output could be hundreds of thousands of classes like in music recommendation (Borges et al.)
- Projected Believe Networks as a novel generative classifier for acoustic events (Baggenstoss et al.).
- How to use auto-encoders for anomaly detection? This slide from by Harada et al. is very illustrative.