🤖 AI Diary #5

Topics include machine learning for health & creativity, AI annual reports, new language modeling technique, summaries of topics such as deep reinforcement learning and AI developer tools,…

On People…

Google AI announces improvements to their deep learning models used for diabetic retinopathy, one of the fastest growing causes of vision loss. The improvements include efforts to improve explainability and applicability in clinical settings — Link

The AI NOW 2018 report has been released and it covers topics such as accountability, fairness, bias, ethical principles, interdisciplinarity, and policy-making, including recommendations for how to improve future AI systems — Link

A recent study conducted by Dr. Fei Fei Li and team proposes a machine learning model that tracks your face and voice features and is able to predict the severity of depression — Link

SpaceSheets is a tool that offers a spreadsheet interface to explore the latent space which represents properties and relationships of images in a generic vector space representation typically trained by generative models like GANs — Link

The AI Index 2018 annual report has been released and includes information and statistics about areas such as AI course enrollments, published papers, jobs, patents, among others — Link

The Google AI team is investing efforts to reduce bias in their language translation tools by incorporating gender-specific translations for gender-neutral words such as “doctor”, which is more critical in languages like Turkish — Link

James Zou et al. (2018) released a tutorial and guide on how to apply deep learning for genomics via a Google Colab notebook — Link

Hear from Irene Chen, an MIT student, how she is investigating the problem of bias in machine learning systems. Preliminary results indicate that one of the big problems is how the data itself is collected, and she tests this hypothesis via different experiments that make use of different datasets to obtain the same level of predictive accuracy — Link

Read about how Geoffrey Hinton, one of the pioneers of deep learning, wants to improve AI systems my making them more data efficient and what he thinks about the progress of AI, in an elusive interview with Wired — Link

On Research…

Amazon researchers have proposed a new language modeling technique that adapts to different conversational contexts given an algorithm that analyses a particular mathematical representation of a grammar’s rules that calculates the probability that a grammar will produce any given string of words. This training strategy is important since acquiring a large body of training data to support a new capability may not be feasible in the real world. Thus, formal grammars are used to generate sample sentences for the language model instead — Link

Stateoftheart.ai is a platform for crowdsourcing state of the art results in different areas of AI research such as reinforcement learning and natural language processing — Link

Read about Johnson Thomas’s experience in building a deep learning model to diagnose malaria from cellphone captured microscopic images. The author uses tools such as fast.ai and TuriCreate to train and compare models — Link

Zimbra et al. (2018) recently published an article reviewing the state of the art models for performing sentiment analysis on Twitter. This is a great review for those who are interested in working in the field as it covers a few of the challenges present in SA, and it even provides a benchmark evaluation — Link

CariGANs is a method proposed to perform photo-to-caricature translation by modeling geometric exaggeration and appearance stylization using GANs — Link

Alex Graves and Marc Aurelio present at NeurIPS 2018 on the important topic of transitioning to fully unsupervised deep learning techniques, including proper definitions of the problem, recipes, and other open research problems in this area of research — Link

On Education…

Spinning Up is a course on Reinforcement Learning for beginners made publicly available by Open AI — Link

Goku Mohandas has released a set of notebooks (called Practical AI) that teach how to program machine learning models using a practical approach via PyTorch — Link

Lilian Weng releases a detailed article explaining the concept of Meta Learning and how it can be used to build models that learn to adapt quickly to new environments given just a few examples — Link

Jay Alammar provides a neatly illustrated guide to BERT, ELMo and other transfer learning techniques used to solve a variety of NLP tasks — Link

Learn about how reinforcement learning is used in NLP in the new NLP Highlight episode with Matt Gardner featuring Hal Daumé III — Link

On Resources and Developers…

The Facebook AI team has released PyText, a set of NLP libraries and pre-built models built on top of PyTorch that provide functionalities to build NLP applications that scale and are efficient at inference time. The library is mostly aimed at developers rather than researchers — Link

This popular twitter thread by DyamicWebPaige provides a comprehensive summary of the TensorFlow ecosystem — Link

TensorSpace.js is a tool built with TensorFlow.js to design and visualize how a model learns, trains, structures information, and predict results using pre-trained models that are rendered via 3D visualizations in the browser — Link

Gab.ai corpus is a large-scale dataset for studying hate speech and toxicity on social media platforms — Link

A new large-scale dataset and task for visual common sense reasoning has been made publicly available, with the goal to enable cognition-level understanding in AI systems — Link

TensorFlow releases a beginner’s guide on how to apply probabilistic programming to real-world problems. It basically serves as a guide on how to leverage TensorFlow Probability, a library built for scientists, statisticians, and ML researchers, to encode domain knowledge to understand data and make predictions — Link

Explore how language is seen and analyzed by models such as BERT and ELMo in this nice visualization tool where you can query for specific tokens and find similar contexts to them — Link

On Industry…

CoinRun is a training environment developed by OpenAI which provides a metric for an agent’s ability to generalize across new environments, which is a challenging task in deep reinforcement learning — Link

FAIR share new open source tools for fastMRI with the aim to increase the efficiency and speed of MRI scans (up to 10x times faster) — Link

Worthy Mentions

• NYU announces a course called “Neural Aesthetics” for learning how to teach different artistic capabilities to neural networks — Link
• The Machine Learning Tokyo group has open sourced a series of GAN models implemented in both Keras and PyTorch — Link
• DGL is a library to build graph neural networks including Graph Convolutional Networks — Link
• Complete slides for the EMNLP 2018 keynotes have been made publicly available — Link
• Anna Huang builds the Music Transformer, a music generation tool built which is able to generate harmonious, minute-long pieces based on a newly proposed self-attention mechanism — Link
• A very nice introductory linguistics course for those wanting to jump into the field of NLP — Link
• Francois Fleuret has updated his deep learning course material to PyTorch 1.0 — Link
• A syllabus for clinical linguistics — Link