Three Turing Award winners debated at the Heidelberg Forum whether deep learning can achieve human-level reasoning

85-year-old Turing Award winner Raj Reddy participated in the recently held ninth Heidelberg Laureate Forum. He sighed sincerely: "I have been working in the field of artificial intelligence for nearly 60 years, but I never thought that this technology would be practical in my lifetime."

10 years ago, in 2012, deep learning made a breakthrough. At that time, an innovative algorithm for image classification based on multi-layer neural networks suddenly proved to be much better than all previous algorithms. This breakthrough enables the application of deep learning in areas such as speech and image recognition, automatic translation and transcription, and robotics.

As deep learning is embedded into more and more daily applications, more and more examples of possible errors have surfaced: artificial intelligence systems will discriminate and formulate stereotypes Impressions, making elusive decisions, and requiring large amounts of data and sometimes large amounts of energy. ​

In this context, the 9th Heidelberg Laureates Forum organized a seminar on the application and impact of deep learning for about 200 young researchers from more than 50 countries. group discussion. Discussions included Turing Award winners Yoshua Bengio, Yann LeCun, and Raj Reddy, 2011 ACM Computing Award winner Sanjeev Arora, and researchers Shannon Vallor, Been Kim, Dina Machuve, and Shakir Mohamed.

Meta chief AI scientist Yann LeCun was the most optimistic of the panelists: “There are a lot of people claiming that deep learning can’t do this or that, and most of those claims have been resolved after several years of work. Proved wrong. Over the past five years, deep learning has been able to do things that none of us could have imagined, and progress is accelerating."

LeCun For example, Meta The company's Facebook now automatically detects 96% of hate speech, up from 40% about four years ago. He attributes this improvement to deep learning. "We are bombarded with tons of information every day, and it's only getting worse. We're going to need more automated systems that allow us to sift through this information."

Shannon Vallor, a professor at the University of Edinburgh in the UK, objects to LeCun's view that technology is just moving forward, it seems to have a will of its own, and society just needs to adapt. “That’s exactly why we get into some of the problems we have. Technology can take many forked paths, and people decide which fork is best. Deep learning systems are built and built by humans based on their own values, incentives, and power structures. deployments are outright artifacts, and therefore we are fully responsible for them."

One of the criticisms of deep learning is that while it is good at pattern recognition, it is not currently suitable for Logical reasoning, while old-fashioned symbolic AI is suitable. However, both Bengio and LeCun see no reason why deep learning systems can’t be used to reason. As Bengio observed, "Humans also use some kind of neural network in their brains, and I believe there are ways to achieve human-like reasoning through deep learning architectures."

However, Bengio added that he doesn't think simply scaling up today's neural networks will be enough. "I believe we can draw more inspiration from biology and human intelligence to bridge the current gap between artificial intelligence and human intelligence."

Theoretical Computer at Princeton University Scientist Sanjeev Arora added that it's not just deep learning that can't reason, but we can't reason with deep neural networks either. Arora said: "We need to understand more about what is going on inside the black box of deep learning systems, and that is what I am trying to do."

Raj Reddy is so far The longest-standing group member in the artificial intelligence community, he has been involved in the doctoral research of artificial intelligence pioneer John McCarthy since the 1960s. Reddy sees the glass as half full, not half empty. "An important application of deep learning is to help people at the bottom of the social pyramid. About 2 billion people in the world cannot read or write. Various language technologies are now good enough to use, such as speech recognition and translation. I work in this field For nearly 60 years, I never expected that this technology would be practical in my lifetime. In ten years, even illiterate people will be able to read any book, watch any movie, and have a conversation with anyone, anywhere in the world, in their native language.” ​

However, handling smaller niche languages ​​remains an unsolved problem for deep learning techniques because far less data is available. In Africa alone, there are 2,000 languages ​​spoken but no AI technology is available, says data science consultant Dina Machuve. It’s important to go into a community and see what works for that community, so when looking for deep learning applications for Africa, Machuve focused on image applications – “We have developed early detection of poultry diseases and crop diseases based on image recognition. Detection systems."

Unfortunately, in many ways, Africa remains the "missing continent" in deep learning research and deployment, adds Shakir Mohamed, a researcher at DeepMind. “We counted how many papers from Africans were submitted at NeurIPS, a well-known neural information processing conference, between 2006 and 2016, and the answer was: 0. The same is true for Latin America, maybe 1. I hope you all People, wherever you are, can take seriously the question of representation, who is doing the work, where it is being done, and how you share your experiences with others.”​

Been Kim, a research scientist at Google Brain, said she hopes people realize that deep learning is not a magic tool that can solve all social problems. In fact, she observed, "There may be non-AI solutions that are better suited to your problem than machine learning. You have to stop and question: Is this the right tool?"

When asked what the general public should know about artificial intelligence and its prospects, Mohamed said: "The future has not been decided yet. We can still create and shape the future, and that is what we should always remember."

The above is the detailed content of Three Turing Award winners debated at the Heidelberg Forum whether deep learning can achieve human-level reasoning. For more information, please follow other related articles on the PHP Chinese website!