Looking for informative applications of where deep learning is now, what possibilities are we currently in? Is there anything we could add to further improve what is already here? I would love to have discussions here. Thank you!
Recently, I worked on cross modal matching.
I believe it is pretty exciting area to work on. Here are two papers that you can study:
Preprint Disjoint Mapping Network for Cross-modal Matching of Voices and Faces
http://www.robots.ox.ac.uk/~vgg/research/CMBiometrics/
Instead of searching for new ML methods, I would search for problems that cannot be addressed with existing tools. Here are my top 3, but there is so much more we need good solutions for:
* I believe that soon enough we will see the ML field's return to small data, so how do you build models to solve problems where there is little to no data. Model cross-domain transference would be one way of doing it, the art of building synthetic training sets would be another ... and a lot more.
* The second area is building decision support tools for mission critical systems. For example in the field of medicine, we need transparent, explainable,... models, and the frameworks/models we build are everything but transparent... Check out :
https://www.frontiersin.org/articles/10.3389/frobt.2018.00120/full
or
Article Survey of Image Processing Techniques for Brain Pathology Di...
* Finally, ML field is doing depth first dive into narrow problem domains, and more and more it feels like we are moving away from building human grade intelligence. The topics that could be explored here are the fuzzy ML systems, concept slips, etc.
Deep Learning usually require large amount of training data and it is sometime difficult to collect. A scenario I faced is image quality assessment.
There are several public datasets which have 700 to 3000 images which have mean of human opinion as a response variable. A minimum of 10 humans are required to evaluate an image for quality in a controlled environment but usually higher number of evaluators are required. In order to get 10K or more images with human opinion seems a difficult and challenging task therefore a solution to this problem is as follows:
1. Generate large set of images from pristine images using simulated distortions.
2. Use synthetic scores such as using PSNR, SSIM or VIF and then obtaining weighted average.
3. Use generative deep models to learn the scoring on small dataset and then assign the scoring to larger dataset for training complex models.
The following link list the project I am working and using deep learning methods for synthetic scoring: https://www.researchgate.net/project/Non-Reference-Quality-Assesment-of-Digital-Images-and-Videos
- Badges
- Science topic