January Data Digest

Welcome to the first post of many. In our monthly digests, Datalore team will bring you the best and latest news from the field of machine learning, deep learning, and artificial intelligence. Read about latest frameworks and algorithms, get relevant books and discussions recomendations, and find out about research breakthroughs.

Tools, frameworks and datasets

  • Google announced the launch of public alpha for AutoML Vision - a framework that lets developers with a limited machine learning experience to train custom vision models
  • Facebook open sourced Detectron - deep learning platform for object detection tasks built on Caffe2. More than 70 pre-trained models are available to download from the model zoo
  • Standford ML group opened new large dataset of musculoskeletal radiographs, where each case is manually labeled by radiologists as either normal or abnormal. A model that uses 169-layer convolutional network achieved performance compared with those of best radiologists - although it performed worse on wrists, elbows, and shoulder.

New algorithms

  • Research groups from Microsoft Research Asia and Alibaba submitted algorithms which beat human performance in text comprehension task on SQuAD (Stanford Question Answering Dataset) by ExactMatch score. Take a look at the leaderboard
    Image Description

Books and articles

  • Discussion on the responsible AI implementation and six ethical principles from Microsoft that should guide the development and use of artificial intelligence in government, academia, or business.
  • Google UX community reflects on a way machine learning transforms human interaction with the technology

Research

  • Improved implementation of PixelCNNs, a recently proposed class of powerful generative models with tractable likelihood
  • "Active Neural Localization" is a new, trained with the reinforcement learning fully differentiable neural network, which combines traditional filtering-based localization methods with a policy model.
  • Investigation of the way various text classifiers work: do these models actually learn to generalize meaning of the sentences or simply reuse key lexicons?
  • LDA-based approach to the Word Sense Disambiguation problem where the whole document serves as the context for a word to be disambiguated
  • Demystification of the generative adversarial networks that use the Maximum Mean Discrepancy as a critic
  • Reducement of deep neural networks' sensitivity to adversarial perturbations by attaching the discriminator to one of the network's hidden layer and training it to filter the adversarial noise
  • Researchers from Montreal Institute for Learning Algorithms present MILABOT, a deep reinforcement learning chatbot that integrates natural generation and retrieval models and relies on deep reinforcement learning - and can make small talk through both text and speech
Show Comments