Twenty Billion Neurons: Building AI That Makes Sense of Video
Twenty Billion Neurons unveiled an AI system that automatically understands complex real-world behaviour in video. Trained on hundreds of thousands of dedicated video recordings, it is the first network of its kind to develop a nuanced understanding of the physical world and the actions we can perform in it.
Twenty Billion Neurons, who ‘build advanced machine learning systems that understand video’ announced a breakthrough in their technology at the RE•WORK Deep Learning Summit in Montreal. As humans, we process the world and solve intelligent tasks largely through common sense - something that can’t be programmed into a computer. At TwentyBN, the team are working to teach machines to understand the physical world through the development of ground-breaking technology that allows machines to perceive the world like humans.
At the summit in Montreal this morning Roland Memisevic, Chief Scientist, announced that earlier this week the team had a breakthrough in their lab:
“Our video systems are now able to accurately recognize highly complex human behaviours in video, which had been entirely unthinkable until now (think “imagenet for video”).”
TwentyBN aim to provide AI applications with better interface capabilities, and ultimately with common sense reasoning. To achieve this vision, TwentyBN have been working with enterprise customers on early prototypes to develop neural models that can make predictions about intuitive physics from videos. Over the past decade as deep learning has evolved, rather than linear progressions there have been a series of step-functions ‘with sudden unexpected outbreaks of capability, which fundamentally changed the capacity of what computers are able to do.’ Roland and his team made the prediction and took the risk that the next outbreak of capability would be related to video understanding: a hypothesis they have now proven.
TwentyBN created spatio-temporal video models, video infrastructure, as well as a data operation that allowed them to create hundreds of thousands of labeled videos. These videos showed the machines every day common-sense scenarios and situations that humans would be able to make sense of immediately. Many of these videos were designed to be extremely subtle and hard to distinguish, something that machines find challenging to learn from. This allowed TwentyBN to ‘successfully train the neural networks end-to-end on a wide range of action understanding tasks that neither hand-engineering nor neural networks had appeared anywhere near solving just a few months ago. At the Deep Learning Summit, Roland demonstrated how these recognition tasks now drive commercial value for TwentyBN as well as driving the long-term AI agenda which represents another long term bet on learning common sense world knowledge through video.
RE•WORK’s global AI Summits continually provide a platform for researchers to share their most cutting edge and impactful breakthroughs in their industry whilst learning from other pioneers in the space. At the Deep Learning Summit in Montreal this week, we welcomed the founders of AI, Yoshua Bengio, Yann LeCun and Geoffrey Hinton who each presented their current research as well as appearing on a panel together to discuss the landscape and progressions of deep learning.
The next RE•WORK events will take place in San Francisco on January 25 & 26. register now to guarantee your place at the summits and stay at the forefront of AI progressions: