Papers / highlights:
- Approaches to eliminate bugs in ML models:
adversarial testing, robust learning, and formal verification - the ideas are cool, but these do not seem to really practical in real life. Back-propagating some error boundaries seems interesting;
NLP
Smart reply
Google’s smart reply works well … but it is kind of useless. It is interesting … that it uses EmbeddingBag as well (lol):
- How smart reply by Google works - looks just like seq2seq with post-processing and filtering. Most interesting part - canonical response set generation;
- Google then substitutes seq2seq network with a ngram based network;
Free sentiment annotation
- Cools idea - use emojis in tweets as annotation. As for CVAEs - this seems kind of academic, but cool as well. Just a classifier trained on a corpus of texts with emojis will be cool;
- Separate attention leyar per each emoji allows easier visualizations;
Graph networks
FAIR released a toolkit to learn embeddings from graphs.
Transformer from FAIR
The coolest thing - now they go on subword level.
1m breast cancer images
- New dataset;
- By the comments from people from industry this dataset is even properly assembled and useful;
Posts / articles
- MFCCs explained;
- Google now recognizes actions in videos;
- Turing award to LeCun and Hinton;
- Memorization in RNNs;
- Adversarial validation in practice (RU);
- Attacking Tesla autopilots;
- Brief history of security keys;
- Overview of how moderation is handled in large tech companies;
- Daimler buys autonomous trucking company;
- DCTs in image compression (JPEG);
- Why you do not need ML;
- Pseudo … technologies?;
- Debugging … mobile interfaces with CNNs;
- Typization in programming languages;
- Idiot’s guide to SVMs;