Last Week in AI
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: Where We See Shapes AI Sees Textures
Image analysis is one of the hottest areas of artificial intelligence research. In recent years, AI image classification algorithms have become famous both for its progress as well as its mistakes. We all have seen the news of AI models misclassifying images of dark-skinned humans or tricky objects. While vision systems powered by AI have been able to outperform humans on some image recognition tasks under fixed conditions, they also fail miserably with the introductions of the simplest distortions. Now a team of German AI researchers has an idea why.
In a recent study presented at the International Conference on Learning Representations in May, a team of researchers from the University of Tübingen in Germany highlighted the sharp contrast between how humans and machines “think,”. While humans are clearly more biased towards shapes when analyzing images, the current generation of AI techniques focuses more on textures which also introduces a lot of confusions. If this result is proven to be correct, it can help to improve the accuracy of image analysis systems to surpass humans under all sorts of conditions.
Now let’s take a look at the core developments in AI research and technology this week:
AI Research
AI researchers from education powerhouse Udacity, published a paper proposing a method to generate videos lessons based on audio narrations.
>Read more in this coverage from VentureBeat
In a shocking discovery, AI researchers from the University of Tübingen, Germany Proved that image recognition systems are typically bias towards textures, not shapes, which is the root caused of many misclassifications.
>Read more in this coverage from Quantas Magazine
In a somewhat surprising paper, AI researcher from the University of Tartu, Estonia proposed a test that shows that deep reinforcement learning agents can learn from the perspectives of other agents.
>See the complete research paper here
Cool AI Tech Releases
Facebook open sourced an implementation of a deep learning recommendation model that combines concepts of collaborative filtering and predictive analytics.
>Read more in this blog post from the Facebook AI Research team
AI researchers from the Massachusetts Institute of Technology open sourced Gen, a new probabilistic programming language for causal inference.
>Read more in this coverage from MIT News
TensorFlow Lite has been ported to the Arduino IOT operating system which means that soon we will see deep learning models running on Arduino micro-controllers.
>Read more in this blog post from Hackster.io
AI in the Real World
Self-driving startup Tier IV raised $100 million for an open source platform for autonomous vehicles.
>Read more in this coverage from Crunchbase News
IBM and MIT collaborated on launching GANPaint, a system that uses AI to help designers add, modify or remove objects to images without loosing the core context.
>Read more in this coverage from MIT News
Researchers from Harvard University use AI to try and figure out whether the authorship of some of the most disputed works in The Beatles’ back catalog can be attributed to John Lennon or Paul McCartney.
