On May 27, the first Global Machine Intelligence Summit (GMIS) hosted by Synced, a world leading AI information and service platform, invited experts from across the world to gather in Beijing to present their insights into the AI industry. Spanning two days, the conference was jam packed with events, including 32 presentations by 47 expert presenters, five breakout sessions, four panel discussions and one human vs. machine competition.
Feiyue Wang, Director of The State Key Laboratory of Management and Control for Complex Systems under the Chinese Academy of Sciences Institute of Automation, made the opening keynote. In his speech, he said, “My personal view is that in a few years, 90 percent of our jobs will be provided by AI, just like the majority of jobs today are provided by machines. We know AI will give us a better future.”
Special guests for the first day of the conference featured many world renowned AI experts such as“The Father of LSTM”- Jürgen Schmidhuber, Chief AI Officer at Citadel – Deng Li, Assistant Director at Tencent AI Lab – Dong Yu, Director of Intel AIPG Data Sciences Department – Yinyin Liu, and CTO of GE Transportation Digital Solutions – Wesly Mukai. Together, through keynotes and panel discussions, they provided the audience with their insights and perspectives on the future of AI.
In the afternoon session, Gary Marcus the Professor of Psychology at New York University, CEO and Founder of Geometric Intelligence, and author of multiple popular science books presented his keynote, entitled “The Road to Artificial General Intelligence”. Currently, the tide of AI is rising quickly and aggressively, but the ultimate goal of AI is to participate in the exploration and application of multiple domains. Marcus presented views and insights which some would consider different from the mainstream.
Whether in his published works, commentaries, or public speeches, Marcus always presents himself as a critic of deep learning. He believes that to achieve true artificial intelligence, deep learning is far from sufficient. The industry needs to explore more research areas, such as cognitive sciences, especially developmental psychology and developmental cognitive neuroscience. He even founded a startup, Geometric Intelligence, with the goal of helping AI systems learn with less training data. His venture, so far, has been successful. Last year, Geometric Intelligence announced the release of an algorithm that would decrease the amount of training data machine learning required, while increasing its speed. In December, Uber acquired his company for an undisclosed amount and used his core team to build its own AI research lab (Uber AI Labs). Marcus joined as Director of the labs, only to leave after four months.
Right after starting his keynote, Marcus presented a tweet by Andrew Ng “If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future”. Marcus believes Andew Ng is only partially correct, “AI is only good at some of the things a person can do almost instantly”.
At this point, Marcus began talking about his view on big data. People can easily perform data analysis, but for machines the task is more than just simple classification. In fact, a lot of simple human tasks are very difficult. For example, a person can easily describe te a picture, but a machine can misinterpret the same content.
Marcus believes that perception is just a small part in realizing artificial general intelligence. Although it is important, human intelligence has many more elements, such as common sense, reasoning, and planning.
Marcus also expressed some of his concerns for AI. Not the Terminator Skynet type of apocalyptic concern, but rather that AI may cause a stall in progress. Currently, we are already employing AI for some tasks such as search optimization, automated monitoring, speech and image recognition. But we have no way of building a machine that can cure cancer, understand how the brain works, or even safely drive a car.
We’ve spent the last 60 years researching AI, and have gathered a huge amount of data, so why haven’t we found the path to artificial general intelligence? Marcus provided his answer in three points. First, even if the technologies are available, due to difficulties in debugging, incremental modifications, and testing, it’s not easy to apply these technologies to real world applications. Here, Marcus recommended everyone to read a paper on machine learning entitled “Machine Learning：The High-Interest Credit Card of Technical Debt.” Secondly, data does not equal knowledge, which means machines require a large amount of training before they can understand the input information. Lastly, the standard bias in this domain is to start everything from scratch. But looking at the history of human evolution, we see how wrong that idea is. Humans evolved for millions of years before we achieved our level of intelligence, even mountain goats must make multiple attempts to master their survival skills, it all boils down to the evolutionary idea of survival of the fittest.
Marcus then mentioned how he thinks this domain should continue moving forward. “Do we need to improve deep learning? Of course this will help, but it won’t solve everything.” People should solve complex problems in developing AI by taking a multidisciplinary approach.
At the end of his speech, Marcus used his daughter as an example. He asked her to put her artwork in her mother’s dresser, and although his daughter is less than three years old, she realized the intent behind this. But AI, on the other-hand, only knows what we tell it to do It cannot understand the purpose of a task.
Marcus asked Siri the same question, but Siri didn’t know how to answer. AI systems still have a lot to learn, reasoning skills being just one of many. When machines finally learn these skills, true artificial general intelligence will finally be within reach.
Localization: Xiang Chen | Editor: Nicholas Richards