What if traffic lights turned red or green at the optimal time? What if a robot could clean up after your kids? What if the city could monitor railroad tracks in real time, preventing collisions between people and trains?
These scenarios and similar ones are not too far in the future, according to experts at Tel Aviv University, where scientists and other researchers are working on several machine learning (ML) and artificial intelligence (AI) projects.
Prof. Amir Globerson said, “Soon it won’t be just self-driving cars or talking devices. It will be machines performing intelligent tasks that help us on a day-to-day level and improve safety and security.”
AI involves developing machines that can perform tasks that are characteristic of human intelligence. ML is the ability for these machines to learn without being explicitly programmed. And deep learning is a subfield of ML specifically concerned with algorithms inspired by the structure and function of the brain called artificial neural networks.
According to Globerson, AI/ML has become one of the hottest fields in global technology today and will play a major role in most, if not all, of the new technologies that emerge in the next 10 years.
Globerson, a researcher at TAU’s Blavatnik School of Computer Science, is heading the university’s new Yandex Initiative for Machine Learning, which is augmenting the number of AI/ML course offerings at TAU and supporting research in the field. Yandex is the largest technology company in Russia and one of the largest Internet search engines worldwide.
According to Globerson, Yandex’s founder Arkady Volozh believes so much in growing the AI/ML field out of Israel that he spearheaded Yandex’s funding of the program.
In his role, Globerson is tackling theoretical deep learning questions, such as why certain algorithms work, and how algorithms should be designed to obtain better performance.
Specifically, he is trying to determine what kinds of algorithms would be needed to help machines understand natural speech, as well as language and visual clues, which would narrow the gap between the tasks that humans and machines could do.
“What if a robot could look at an image and understand the relationship between its different parts?” Globerson asks. “Let’s say this robot could analyze the content of an image or scene. You would ask, ‘Where are my keys?’ and the robot could identify your keys, where it last saw them, and help facilitate your finding them.”
He continued, “I think we are closer to the dream of building machines that can behave like humans and maybe even outperform humans on intelligence-related tasks.”
These machines could also keep humans safer.
IN ONE EXAMPLE, TAU recently signed a cooperation agreement for joint research in “Smart Cities and Digital Living” with Stanford University, which draws heavily on big data and machine learning. The five-year grant, funded by the Koret Foundation of San Francisco, is bringing together leading scholars and scientists from TAU with their counterparts at Stanford to advance multidisciplinary, basic and applied research in data science that enhances the quality of life, safety and efficiency of cities, while supporting communications across people and organizations.
Raja Giryes, a faculty member in the School of Electrical Engineering, explained that today, traffic lights change from red to green based on preset timings.
“But what if you could understand how traffic flows from one direction to another at different hours of the day in an automated way? Then you could change the duration of the traffic lights in real time to reduce traffic and improve safety,” Giryes explained.
Similarly, he said, while traffic monitoring is currently limited to what police or traffic cameras can monitor (speeding or running a red light, for example), these methods cannot account for other violations or factors such as mobile phone usage or wild driving that might contribute more to accidents but are harder to detect automatically.
“Machine learning can help us understand what violations are being perpetrated, and where people are violating rules, and even predict these occurrences in advance, so that monitors can catch them in the act, or so conditions can be changed to preempt such incidents,” Giryes said.
IN RELATED RESEARCH, Giryes, TAU researcher Ido Yovel and Prof. David Mendlovic are using deep learning mathematical tools for optical signal processing. Together, they are working to develop an efficient light field camera based on compressive imaging capability. They believe a combination of light field photography and compressive imaging will provide the capability to capture a three-dimensional image with a much more compact and cheaper camera than anything that exists today.
Giryes said such cameras would improve the ability to capture photographs in rain or fog, such as when monitoring traffic, as explained above.
Further, Giryes said that a hybrid design of the optical system and the digital computations have led to new cameras with multiple applications. These include all-in-focus imaging, which is restoring to focus objects that would not be in focus in a photograph taken by a traditional camera. Another example is a depth camera that estimates the depth of a scene and the various objects within it with only one sensor.
In both instances, the cameras combine optics and AI to achieve their goals.
Researcher Lihi Shiloh is combining AI methods with optical fiber sensors. Such sensors can perform high spatial resolution measurements at high scan rates and very long distances. These sensors produce massive amounts of complex optical data and by using AI she is developing new algorithms to analyze and extract interesting events.
“In the real world, this would mean that you could perform railway monitoring – for example, detect if a train is on the track at the same time a person is walking on it and avoid a collision by allowing you to act in real time,” Shiloh said.
At TAU, a 100-meter-long fiber is buried in the campus. Lihi refers to it as the “optical garden,” where she carries out experiments to see if the sensor is capturing various seismic events. She then analyzes and classifies them. She said the research looks promising.
According to Giryes, “The real impact of AI/ML on our lives is still ahead.” In addition, he said that TAU is poised to lead the field in Israel, both because of its proximity to the high-tech arena, which allows for cross-fertilization, and because of the university’s focus on running multidisciplinary AI/ML research.
“TAU is at the forefront because it covers a spectrum of sides of the field – between theoretical understanding and application, as well as taking a cross-disciplinary approach,” said Globerson, noting that in general, deep learning is becoming a field that pulls together math, science and electrical engineering, among other fields, and only those universities that are cross-disciplinary and flexible will thrive.
Said Yovel, “When we work together, we can overcome challenges we each had in our own subjects before now. It is a really exciting time to be in research.”This article was written in cooperation with Tel Aviv University.