2019 DS/ML digest 11

2019 DS/ML digest 11

Posted by snakers41 on May 27, 2019

Buy me a coffeeBuy me a coffee

Blog posts

  • NER for people from Wikipedia tags;
  • Snapchat has a face swap CycleGAN-like feature?;
    Ben Evans on Google keynote;
  • Blogging advice from fast.ai;
  • SWA now featured in torch contrib;
  • A tutorial on normalizing flows;
  • Google makes voice-to-voice translation w/o intermediate text representation - spectrogram to spectrogram + vocoder;
  • Chinese Face++ raises US$750m;
  • Pre-training BERT on TPUs:
    • Pre-training a BERT-Base model on a TPUv2 will take about 54 hours;
  • GitHub sponsors = patreon?;

Interesting papers

  • New paper from Ian GoodFellow on reducing the amount of annotation required for DL:
    • Use supervised labels as well as semi-supervised predicted labels;
    • Sharpen the resulting predicted distributions;
    • Use mix-up;
    • Have 2 losses - for supervised and semi-supervised parts of the dataset;
  • Parallel Neural Text-to-Speech. The paper does not recognize all the progress made by IAF vocoders in 2018, but the idea is cool - distill a recurrent spectrogram network into a non recurrent one;
  • New attention blocks for CV?
  • Mobile network search = lottery ticket?;

Internet / statistics

  • Huawei is one of smartphone leaders in Europe;
  • Effectiveness of basic methods for account protection;
... adding a recovery phone number to your Google Account can block up to 100% of automated bots, 99% of bulk phishing attacks, and 66% of targeted attacks that occurred during our investigation.... an SMS code sent to a recovery phone number helped block 100% of automated bots, 96% of bulk phishing attacks, and 76% of targeted attacks.On-device prompts, a more secure replacement for SMS, helped prevent 100% of automated bots, 99% of bulk phishing attacks and 90% of targeted attacks.

Cool things

  • A competition to map what people see into their MRI signals using CNNs. Small set (~200 images * 15 people), but very cool!;
  • A paper that aims to visualize the internals of a trained BERT network?;