What are some (non image) datasets where deep networks are required for good performance?

Question

It is well known that deep networks tend to outperform shallow networks and other classical machine learning techniques such as boosting on learning tasks involving images. I believe this is because extracting useful high-level features from low-level pixel intensities requires deep models.

I'm curious to know if there are any non-image data sets (e.g., tabular datasets that are typically found in UCI repository) where deep networks are required for good performance and where shallow networks, boosting tend to perform poorly.

Answer 1

Welcome to the community!

In text-related tasks, DL have shown amazing breakthrough e.g. in NLU for reading comprehension or question answering. If you have heard about Google's BERT, that was one of big steps which, to some extent, pushed the boundaries of SOTA in NLP/NLU last year and currently, a heavy wave is going on doing research on different NLP/NLU tasks mostly based on either the idea coming from BERT or directly from its pretrained models.

The quora QA dataset, SQuAD dataset, etc. are great starting points to play with bert and its successors and see how great the idea was.

Hope it helps!