He was not very talented at writing, but still became a best-selling author. He is an academic who founded two startups – one acquired by Uber, another scoring just $ 15 million to make it easier to build smarter robots. He has a humanistic background, yet he became one of the most prominent and controversial figures in AI.
If you know Gary Marcus, then you probably know that it is very difficult to summarize someone like him. If you do not, here’s your chance to change it. Gary Marcus is a scientist, best-selling author and entrepreneur. Marcus is well known in AI circles, mainly for his critique of – and subsequent debates about – a number of topics, including the nature of intelligence, what’s wrong with deep learning, and whether four lines of code are acceptable as a way to infuse knowledge when sowing algorithms.
Although Marcus is sometimes seen as “almost a professional critic of organizations like DeepMind and OpenAI”, he is much more than that.
As a precursor to Marcus coming main note on the future of AI in knowledge connections, ZDNet caught Marcus on a wide range of topics. We publish the first part of the discussion today – come back to the second part next week.
From cognitive psychology to AI
In February 2020, Marcus published a 60-page paper entitled “The Next Decade in AI: Four Steps to Robust Artificial IntelligenceIn a way, this is Marcus’ response to his own critics who go beyond criticism and make concrete suggestions.
Unfortunately, to quote Marcus, the world has much bigger problems to deal with right now, so the paper has not been discussed as it might have been in a pre-COVID world. We agree, but we think it’s time to change that. We discussed everything from his background to AI debates and from his latest paper to knowledge graphs.
Marcus is a cognitive psychologist by training. This may seem strange to some people: how can someone with a background in the humanities be considered one of the top minds in AI? To us, it did not seem so strange. And it made even more sense after Marcus had expanded the subject.
Marcus comes to AI from a perspective to try to understand the human mind. As a child and teenager, Marcus programmed computers – but quickly became dissatisfied with the latest technology in the 1980s.
Marcus realized that humans were much smarter than any of the software he could write. He skipped over to the last few years of high school, based on a translator he wrote who worked from Latin to English, which he said was one of his first serious AI projects. But then another realization hit home:
“I could do a semester ‘worth of Latin by using a lot of tricks, but it wasn’t really very deep, and there was nothing else out there that was deep. This eventually led me to study human language acquisition and human cognitive development. “.
Marcus teamed up Steven Pinker, who was his mentor as a PhD student. They worked on how people acquire simple language parts like the past tense of English. Marcus spent a lot of time comparing neural networks that were popular then with what human children did. These neural networks went into obscurity, and then they re-emerged in 2012.
When they got together, Marcus realized that they all had the same problems that he had criticized in some of his early technical work. Marcus has spent much of the last decade trying to look at what we know about how children learn about the world, languages, and so on, and what it can tell us about what we might need to do to make progress. in AI.
Interdisciplinarity, debates and doing it wrong
As an interdisciplinary cognitive scientist, Marcus has tried to gather what we know from many areas to answer some really tough questions: How does the mind work? How does it develop? How did it develop over time?
This led to him also being a writer when he found out that people in different fields did not speak each other’s languages. The only way to bring different fields together was to make people understand each other. This was his motivation for becoming a writer. I have began writing for The New Yorker, and has written five books, including The Algebraic Mind, Clues, The Birth of the Mind, and New York Times best selling Guitar Zero.
At some point, AI came back into vogue, and Marcus felt that “everyone did everything wrong”. This led him to be an entrepreneur as he wanted to take a different approach to machine learning. Marcus founded Geometric intelligence, his first company, in 2014. Uber acquired it relatively early in the company’s history, and Marcus helped start it Uber AI Labs.
Marcus launched a new company called Robust AI in 2019. The goal is “to build a smarter generation of robots that you can rely on to trade on your own”. Instead of just being serviced and working with the assembly lines, Robust AI wants to build robots that work in a wide range of environments – home, retail, elderly care, construction and so on. Robust AI raised just $ 15 million, so apparently progress is underway.
Marcus believes that his activities inform each other in a constructive way. But he also believes that AI culture right now is unfortunate, citing his debates with people in the deep learning camp, such as Yoshua Bengio, Geoff Hinton and Yan LeCun – who recently won the Turing Prize:
You have this set of people who worked in obscurity. They are now in power. And I feel like rather than learning from what it’s like to be ignored, they disregard many other people. And because I’m not shy and I have some kind of courage..not the kind of courage that our health professionals have, but I’ve come out on a different kind of front line and said:
“This is what we have to do. It’s not popular right now, but that’s why the things that are popular are not working. And that led to a lot of people getting annoyed with me. But I think you “I learned from my father to stand up for what I believe, and that’s what I do.”
Language models do not know what they are talking about
Becoming personal is not the best way to make progress in science or technology. Marcus admitted to spending a lot of time in these debates, saying he was exposed to a lot of (verbal) abuse, especially between 2015 and 2018, as his Twitter feed testifies. The good side, he added, was that this brought attention around the subject.
“Even deep learning leaders recognize the hype and some of the technical limitations of generalization and extrapolation that I have pointed out for a number of years. AI has a lot to offer or could have a lot to offer if we did better. And I’m happy to see people start looking at a wider range of ideas. Hopefully that will lead us to a better place. “
The epitome of Marcus’ critique of deep learning focuses on language models such as GPT-2, Meena, and now GPT-3. The crux of it is that these models are all based on some sort of “brute force” approach. Although hailed as AI’s greatest achievement, Marcus remains critical and unimpressed. In his articles as well as in his “Next Decade in AI” paper, Marcus presents a convincing critique of language models:
“These things are approximations, but what they are approximations to is language use rather than language comprehension. So you can get statistics on how people have used language, and you can do some amazing things if you have a large enough database. And that is what they have gone and done.
“So you can predict, for example, what the next word is likely to be in a sentence, based on what words have happened in similar sentences over a very large database of gigabytes of data and often locally it is very good. The systems are very good at to predict category.
“So if you say I have three of these and two of these, how many have there in total? It will definitely give you a number. It will know that a number has to come next because people use numbers in certain “These systems have no real understanding of what they are talking about”.
Marcus has made benchmarks to demonstrate his point and shared his results. He uses simple questions to show the difference between understanding the general category of something to happen next and the details of what is really happening in the world:
“If I say, I put three trophies on one table, and then I put another on the table. How many do I have? You as a human can easily add three plus one. You build a mental model for how many things are there. And you can interpret that model.But these systems, if you say I put three trophies on the table and then another, how many are there?
“They say maybe seven or twelve. They know it’s supposed to be a number, but they do not actually understand that you’re talking about a set of objects that can be counted in a certain place, and they do not know how to do the count. Then came more models with even larger databases, and I had examples like, what’s your favorite band called? Avenged Sevenfold.
“Then they asked if the same system was the name of your least favorite band? And again it’s Avenged Sevenfold. Every human would realize that your favorite band and your least favorite band can not be the same unless you lie or try to be funny. These systems do not understand “.
The next decade in AI: Four steps towards robust AI
Marcus points out that this is a really deep shortcoming, and one that goes back to 1965. ELIZA, the first expert system, just matched keywords and talked to people about therapy. So there is not much progress, Marcus argues, certainly not exponential progress, as people like Ray Kurzweil claim, except in narrow fields such as playing chess.
We still do not know how to make a system for general purposes that e.g. Can understand conversations. The counter-argument to that is that we just need more data and larger models (hence also more computation). Marcus asks to be different and points out that AI models have grown and are using more and more data and computing, but the underlying problems remain.
Recently, Geoff Hinton, one of the ancestors of deep learning, claimed it deep learning will be able to do everything. Marcus believes that the only way to make progress is to put together building blocks that are already there, but no current AI system combines.
Building block number one: A connection to the world of classic AI. Marcus does not suggest getting rid of deep learning, but uses it along with some of the tools for classic AI. Classical AI is good at representing abstract knowledge, representing sentences or abstractions. The goal is to have hybrid systems that can use perceptual information.
Number two: We need to have rich ways of specifying knowledge, and we must have great knowledge. Our world is filled with many small pieces of knowledge. Deep learning systems mostly not. They are mostly just filled with connections between certain things. So we need a lot of knowledge.
Number three: We need to be able to argue over these things. Let’s say we have knowledge of physical objects and their position in the world – for example, a cup. The cup contains pencils. Then AI systems need to be able to realize that if we cut a hole in the bottom of the cup, the pencils could fall out. Humans do this kind of reasoning all the time, but current AI systems do not.
Number four: We need cognitive models – things inside our brains or inside computers that tell us about the relationship between the devices we see around us in the world. Marcus points to some systems that can do this some of the time, and why the inferences they can make are far more sophisticated than what deep learning alone does.
Interesting what Marcus suggests seems very close to how the latest technology is in real life. But we can only scratch the surface in trying to sum up such a rich conversation on such a deep topic. We are visiting again to expand with more topics and nuances and anchor the conversation to specific approaches and technologies next week.