*|MC:SUBJECT|*

View this email in your browser

Dear friends,

Thank you for supporting us through a year like no other. This month we presented a record 31 publications at the premier venue for AI research, the Conference on Neural Information Processing Systems (NeurIPS). In November, we convened experts from academia and industry for a three-part event, What's Next in AI. If you missed it, take a look. The series has drawn more than 300,000 views so far.

Congratulations to MIT's Li-wei Lehman and colleagues for winning a best paper award at the American Medical Informatics Association (AMIA)’s annual symposium. Their paper weighs the costs and benefits of applying deep reinforcement learning in hospital intensive-care units. Looking ahead to 2021, we are excited to launch the ObjectNet Challenge, a competition aimed at developing more robust machine vision models.

We wish you and your family a joyful holiday season, and we look forward to seeing you in the new year!

Aude Oliva, MIT director of the MIT-IBM Watson AI Lab
David Cox, IBM director of the MIT-IBM Watson AI Lab

Research

A Vision Model that Sees More Like Humans

Adding a layer to a neural network that mimics the brain’s visual processing system can stop computer vision models from making common mistakes. "[Our paper] suggests that there is still a lot that AI can learn from neuroscience, and vice versa,” IBM's David Cox told MIT News. IBM Research explores the potential for deflecting adversarial attacks in this blog post.

Shrinking Language Models

Modern language models are not only massive but computational expensive. A new approach could make them leaner and more efficient. In new work, MIT and IBM researchers apply the Lottery Ticket Hypothesis to language models with the promise of lowering computing costs and bringing state-of-the-art AI applications to smartphones.

Vibrating Proteins May Help Infect Cells

Vibrations of the coronavirus spike protein allow it to invade human cells the way jiggling a key in a sticky lock can help open a door, says a new study. “If it's static, it just either fits or it doesn't fit,” MIT’s Markus Buehler told MIT News. But the protein spikes are not static; “they’re vibrating and continuously changing their shape slightly, and that's important.”

Bringing deep learning to IoT devices

A new AI system developed by MIT's Song Han and colleagues could improve the functionality and security of household appliances connected to the Internet of Things (IoT). “Our end goal is to enable efficient, tiny AI with less computational resources, less human resources, and less data,” Han told MIT News. Read additional coverage in Wired and Stacey on IoT.

In the media

Preparing for the coming deluge of AI-generated fakes

The fake biography, Barack Obama Book, reached Amazon’s Top 100 list in November before it was abruptly pulled from the site. How well can a machine detect another machine’s writing? Slate gave GLTR, a tool co-developed by IBM's Hendrik Strobelt, a test run to find out.

AI and quantum information science will remain a priority under Biden

Joe Biden has signaled AI and quantum computing will remain a focus for R&D funding, a move supported by IBM's Dario Gil. “U.S. policy must be a catalyst and a champion of this movement to ensure American leadership, security and prosperity,” he told the Wall Street Journal.

In the virtual world

Watch: What's next in AI?

A three-part series: AI We Can Trust; AI We Can Scale; AI We Can Reason With

AI offers a competitive advantage, but only a fraction of companies are using it to its full potential. In this three-part series, we convened scientists and business leaders to explain how to overcome three key barriers to implementing AI successfully — trust, scalability, and reasoning.

AI to fight discriminatory lending

People make biased decisions, and so can algorithms trained on data that reflect historic biases. But just as AI can perpetuate discrimination, it can also combat it. With Wells Fargo, IBM researchers have developed an algorithm that detects when subgroups have been charged unfair borrowing rates.

Join the ObjectNet Challenge!

With MIT's Center for Brains, Minds and Machines, the Lab this month launched a competition to create the next generation of robust object recognition models. Winners will be announced in June at the 2021 Conference on Computer Vision and Pattern Recognition.

You are receiving this email because you signed up to receive our newsletter or are affiliated with the Lab.

Our mailing address:
MIT-IBM Watson AI Lab, 75 Binney Street
Cambridge, Mass., 02142

Want to change how you receive these emails?
You can update your preferences or unsubscribe from this list.