0xAB
Andrei Barbu
Andrei is a research scientist at MIT working on natural language processing, computer vision, and robotics, with a touch of neuroscience.
Below are some notes explaining a few recent threads in our research. To see a bigger picture overview click research above and of course Scholar has all of our papers.
Computer vision datasets are too easy!
We measured how hard computer vision datasets are for humans. 90% of images in current popular datasets are easy for us. Those are also the images that machines perform on. This could explain why machines perform well in benchmarks but poorly in the real world, we're not challenging them enough.
Brains, Minds, and Machines Summer School
The study of intelligence spans cognitive science, neuroscience, and computer science. If we're going to make serious progress on what intelligence is we need to train a generation of researchers that are competent in all three areas. That's what the BMM Summer School has been doing for almost 10 years now.
Social interactions are a blind spot for AI/ML
Almost everything we do is social. Yet there are almost no social benchmarks in AI research. And there are almost no mathematical models of what social interactions are. Our group set out to fix this and introduced Social MDPs. There's much more to do, but social computing could revolutionize both our relationship to machines and our understanding of ourselves.
ObjectNet — uncovering a wide gulf between humans and machines
Machines perform well on computer vision benchmarks, but they don't perform anywhere as well in the real world. Why? We show that machines are taking advantage of systematic weaknesses in our benchmarks.
Andrei is a research scientist at MIT working on natural language processing, computer vision, and robotics, with a touch of neuroscience.