What's a little more computing power can do

Neural networks have provided researchers with a powerful tool for looking ahead and making predictions. But one drawback is their insatiable need for data and computing power ("computation") to process all this information. At MIT, it is estimated that the demand for calculation is five times higher than what the Institute can offer. The industry intervened. An IBM $ 11.6 million supercomputer recently launched online this fall, and last year both IBM and Google provided cloud credits to MIT Quest for Intelligence for on-campus distribution. Four projects made possible by IBM and Google donations in the cloud are described below.

Smaller, faster and smarter neural networks

To recognize a cat on an image, an in-depth learning model may need to see millions of photos before its artificial neurons "learn" to identify a cat. The process requires a lot of calculation and has a high environmental cost, as shown by new research to measure the carbon footprint of artificial intelligence (AI).

But there can be a more efficient way. A new study from MIT shows that models only require a fraction of the size. "When you have a large network, there is a small network that could have done it all," says Jonathan Frankle, a graduate student in MIT's Department of Electrical and Computer Engineering (EECS).

With co-author of the study and Professor Michael Carbin, Frank Frank estimates that a neural network could be a tenth of the number of connections if the appropriate subnet was found at the start . Normally, neural networks are adjusted after the training process, and irrelevant connections are removed. Why not train the little model first, Frankle wondered?

Experiencing a two-neuron network with his laptop, Frankle got encouraging results and switched to larger sets of image data, such as MNIST and CIFAR-10, using GPUs wherever he could. Finally, via IBM Cloud, it has got enough computing power to form a true ResNet model. "All I had done before was toy experiments," he says. "I was finally able to run dozens of different parameters to be able to make the statements in our document."

Frankle intervened in Facebook's offices, where he worked during the summer to explore the ideas outlined in his article on the lottery ticket hypothesis, one of two candidates selected for the Best Paper Award at the International Conference on Representations of Learning this year. According to Frankle, the potential applications of the work go beyond image classification, and include reinforcement learning models and natural language processing. Researchers at Facebook AI Research, Princeton University and Uber have already published additional studies.

"What I like about neural networks, is that we have not yet laid the groundwork," says Frankle, who recently dropped the study of cryptography and the technological policy for the benefit of AI. "We really do not understand how he learns, where he is good and where he fails. It's physics a thousand years before Newton.

Distinctive fact of the false news

Networking platforms such as Facebook and Twitter make it easier than ever to search for quality information. But too often, real information is drowned by misleading or downright false information published online. The confusion surrounding a recent video of US House Speaker Nancy Pelosi, falsified to make her sound drunk, is only the latest example of the threat of misinformation and false information for democracy. .

"You can just put everything on the Internet now, and some people will believe it," says Moin Nadeem, senior MIT officer at EECS.

If technology has helped create the problem, it can also help to solve it. That's why Nadeem chose a superUROP project focused on creating an automated system to combat false and misleading information. Working in the laboratory of James Glass, a researcher at MIT's Computer Science and Artificial Intelligence Laboratory and supervised by Mitra Mohtarami, Nadeem contributed to the formation of a linguistic model to verify facts in searching in Wikipedia and in three types of news sources classified by journalists as high quality, mixed quality or low quality.

To verify an allegation, the model measures how well the sources match, with the highest agreement scores indicating that the claim is likely to be true. A high score of disagreement for an assertion such as "ISIS infiltrates the United States" is a strong indicator of false news. One of the drawbacks of this method, he says, is that the model does not identify independent truth, but describes what most people think is the truth.

With the help of Google Cloud Platform, Nadeem has conducted experiments and created an interactive website allowing users to instantly assess the accuracy of a request. He and his co-authors presented their findings at the North American Association of Computational Linguistics (NAACL) conference in June and continue to develop their work.

"Before, we said that seeing is believing," says Nadeem in this video about his work. "But we are entering a world where it is not true. If people can not trust their eyes and ears, the question of knowing we trust? "

Visualize global warming

From the rise of the seas to the multiplication of droughts, the effects of climate change are already being felt. In a few decades, the world will become a warmer, drier and more unpredictable place. Brandon Leshchinskiy, a graduate student in Aeronautics and Astronautics at MIT (AeroAstro), experiments with generative conflicting networks, or GANs, to imagine what Earth will look like at the time.

GANs produce hyper-realistic images by opposing one network of neurons to another. The first network learns the underlying structure of a set of images and tries to replicate them, while the second decides which images seem implausible and asks the first network to try again.

Inspired by the researchers who used the GANs to visualize sea-level rise projections from street-level imagery, Leshchinskiy wanted to see if satellite images could also customize climate projections. With his advisor, Professor AeroAstro Dava Newman, Leshchinskiy is currently using free IBM Cloud credits to form two GANs on images of the US East Coast with their corresponding elevation points. The aim is to visualize how sea level rise projections for 2050 will redraw the coastline. If the project works, Leshinskiy hopes to use other data sets from NASA to imagine future ocean acidification and changes in phytoplankton abundance.

"We have passed the point of mitigation," he says. "Visualizing what the world will look like in three decades can help us adapt to climate change."

Identify athletes from a few gestures

Some movements on the field or on the court are enough for a computer vision model to identify individual athletes. This is the result of a preliminary study conducted by a team led by Katherine Gallagher, researcher at MIT Quest for Intelligence.

The team trained computer vision models on video recordings of tennis and football matches and basketball and found that models could recognize individual players in a few images from key points in their body, thus providing a rough overview of their skeleton.

The team used a Google Cloud API to process video data and compared the performance of their models to models formed on the Google Cloud platform. "This pose information is so distinctive that our models can identify players with almost as good accuracy as models with much more information, such as hair color and clothing," she says. .

Their findings are relevant to the automated identification of players in sports analysis systems, and they could form a basis for further research on player fatigue deduction to anticipate when players need to be replaced. Automated pose detection could also help athletes improve their technique by allowing them to isolate specific movements associated with the golfer's expert driving or the tennis player's winning swing.