The week in AI, ML and bots #3

The week in AI, ML and bots #3

(cover image from Wired)

  • OpenAI has open their Requests for Research 2.0: Inspiring research by exposing seven original problems (plus two additional "warm-ups") and calling for ideas, papers and solutions. A great way to develop your AI skills by working on meaningful and well-defined cases.
  • Speedy neural networks for smart auto-cropping of images: How Twitter has dramatically improved the way it automatically crops pictures using saliency detection and making its implementation 10x faster as a vanilla implementation.
  • Hearing AI: Getting started with Deep Learning for audio on Azure: A deep-dive article explaining how Microsoft data scientists have trained a deep learning model to classify sounds. The entire experiment is available on GitHub and can be easily explored on Azure with a Data Science Virtual Machine.
  • Comparing Machine Learning as a Service: Amazon, Microsoft Azure, Google Cloud AI: Azure ML Studio stands out as the only solution offering a comprehensive graphical interface, catering for both newcomers and experienced data scientists. It also provides one of the largest collections of ML algorithms.
  • Deep Learning est mort. Vive Differentiable Programming!: A post from Facebook's director of AI research Yann LeCun, arguing that the next step in machine learning is the rising trend of "dynamic networks" that change dynamically based on the data they receive.
  • Artificial intelligence is for optimization — human intelligence is for innovation: There seems to be a solid consensus within the community that AI is not intended to replace humans, but to augment them; in other words, it is not about doing what we already do, but helping us to do it better. This article covers how AI can be extremely good at optimizing a specific task but does not understand our world as a whole, which will always be required to innovate.
  • Deep misconceptions about deep learning: Deep learning (or deep neural networks) has seen impressive improvements over the past years and is largely behind the current AI hype, but still remains a technique that's difficult to apply effectively. This article describes deep learning with a systematic approach and shows the path to building a successful model. Favorite quote: Data is greater than model design, and model design is greater than parameter optimisation.
  • 7 use cases for data science and predictive analytics: One of the most powerful applications of machine learning is the ability to find patterns in historical data and use these patterns to generate predictions. But for many businesses, it's still challenging to transpose this ability to their own use-cases and envision how ML can be leveraged for them. These seven examples describe practical ML applications in a range of different industries.
  • 3 companies using AI to make the world better: Another showcase of successful AI business applications.