Last Week in AI

Source: Deep Learning on Medium


Every week, my team at Invector Labs publishes a newsletter to track the most recent developments in AI research and technology. You can find this week’s issue below. You can sign up for it below. Please do so, our guys worked really hard on this:

From the Editor: Teaching AI to Program Like a Kid

Can AI agents learn to code? The thought of AI programs creating programs is as scary as fascinating. The subject of representing computer programs as machine learning problems is an active area of research within the AI community. However, most efforts have focused on sophisticated programming languages. What if we could tech AI to learn how to program the same way children learn? Among the methodologies for teaching kids programming skills, Sketch reigns are one of the most popular options. The programming language created by researchers from the Massachusetts Institute of Technology(MIT) has been the first introduction to programming for millions of kids. As a result, Sketch has become an invaluable source of knowledge about how kids learn to program.

Recently, researchers from MIT decided to leverage Sketch to train AI programs. The result was SketchAdapt, a new AI system that was trained in thousands of AI programs and learned to create effective high level programs that outperformed some of the traditional versions. If AI programs can learn to program like kids do maybe we can use that knowledge to unlock the secrets of mastering more sophisticated programs.

Now let’s take a look at the core developments in AI research and technology this week:

AI Research

Microsoft AI researchers published a new paper and framework to train indoor AI agents using imitation learning.

>Read more in this blog post from Microsoft Research

IBM AI researchers published a paper that adds some optimizations to the famous Information Bottleneck Theory to explain how neural networks build knowledge.

>Read more in this blog post from IBM Research

Also from IBM came a new paper proposing a new method for optimizing transfer learning strategies.

>Read more in this blog post from IBM Research

Cool Tech Releases

Facebook open sourced AI Habitat, a simulation platform for training robots in 3D environments.

>Read more in this blog post from Facebook AI Research

Google is releasing some AutoML enhancements to its famous Transformer architecture as part of the Tensor2Tensor library.

>Read more in this blog post from Google AI research

Uber published a new video explaining Ludwig, their open source, no-code tool for training deep learning models.

>Read more in this blog post from the Uber engineering team

AI in the Real World

AI researchers from the Massachusetts Institute of Technology published a program that trains AI agents to write code.

>Read more in this article from MIT News

The American Civil Liberties Union published a comprehensive report highlighting how advancements in AI can be used to supercharge surveillance.

>Read more in this coverage from the New York Times

A new AI model pioneered at MIT is able to effectively generate images of human faces by analyzing speech segments.

>Read more in this article from Smithsonian