Sunday, May 20, 2018

Truly intelligent machines

This interview with Judea Pearl in Quanta Magazine apropos his book “The Book of Why: The New Science of Cause and Effect” has been making the rounds on social networks . Here are some thoughts:
“[…] as Pearl sees it, the field of AI got mired in probabilistic associations. These days, headlines tout the latest breakthroughs in machine learning and neural networks. We read about computers that can master ancient games and drive cars. Pearl is underwhelmed. As he sees it, the state of the art in artificial intelligence today is merely a souped-up version of what machines could already do a generation ago: find hidden regularities in a large set of data. “All the impressive achievements of deep learning amount to just curve fitting,” he said recently.”
It is true that the hype around ML/DS is at a peak. The stream of new posts on Twitter and LinkedIn on algorithms like Gradient Descent or their applications/pitfalls give me new things to read every day. I have downloaded a few free Machine Learning books a couple of weeks ago, and what is notorious is that they jump very quickly into math and statistics, from linear algebra to derivatives and calculus. A few days ago my partner was doing something with Deep Neural Networks, and I suddenly saw her get paper and pencil and start solving derivatives. “What are you doing?” “Oh, I have to solve this to implement back propagation”…
I had one class of Artificial Intelligence when I did my Computer Science degree way back then. We studied things like A* (the algorithm used to explore and rank alternatives in games like checkers), Expert Systems to help diagnose health problems, SNePS for knowledge representation and talked about Neural Networks briefly. There was no complex maths or statistics. Now, I’m not saying that “those were the good old days” - as the area was having one of its “AI Winters” and finding limited success, but stating that today’s conversation is dominated by a relatively narrow set of Narrow AI techniques, heavily based in statistics and focused on training models with very high volumes of data. These are proving successful and having a major impact in many areas, with more to come. But is this it? Will this be what leads us to AI, is this THE critical approach that will crack [General] AI, or will we bump into a “local maximum”?
“I can give you an example. All the machine-learning work that we see today is conducted in diagnostic mode — say, labelling objects as “cat” or “tiger.” They don’t care about intervention; they just want to recognize an object and to predict how it’s going to evolve in time.
I felt an apostate when I developed powerful tools for prediction and diagnosis knowing already that this is merely the tip of human intelligence. If we want machines to reason about interventions (“What if we ban cigarettes?”) and introspection (“What if I had finished high school?”), we must invoke causal models. Associations are not enough […]”
This seems to be a powerful argument. But where Judea Pearl gets really challenging is with the following:
“As much as I look into what’s being done with deep learning, I see they’re all stuck there on the level of associations. Curve fitting. That sounds like sacrilege, to say that all the impressive achievements of deep learning amount to just fitting a curve to data. From the point of view of the mathematical hierarchy, no matter how skillfully you manipulate the data and what you read into the data when you manipulate it, it’s still a curve-fitting exercise, albeit complex and nontrivial. […]
I’m very impressed, because we did not expect that so many problems could be solved by pure curve fitting. It turns out they can. But I’m asking about the future — what next? Can you have a robot scientist that would plan an experiment and find new answers to pending scientific questions?
The first part of this will no doubt create aversion in most people – myself included – who are passionate about where tech is going, but it’s hard to shake off the feeling – especially with the second part – that he probably has a point here. It may not matter for the people and companies, as the achievements are indeed impressive, and all across areas like healthcare, retail, autonomous driving, finance – or anywhere with data to process – there is a wealth of data to process and new automation/personalization to do. The impact in our lives will continue to happen. But what we’re doing at the moment is likely not enough.
And just one more quote from this interview.
“AI is currently split. First, there are those who are intoxicated by the success of machine learning and deep learning and neural nets. They don’t understand what I’m talking about. They want to continue to fit curves. But when you talk to people who have done any work in AI outside statistical learning, they get it immediately. I have read several papers written in the past two months about the limitations of machine learning.
[…] a serious soul-searching effort [is developing] that involves asking: Where are we going? What’s the next step?
Maybe the momentum, the exponential speed of development in the field, will make this question moot. Something will will come up – or many somethings - that will allow the tech companies to go on. Or maybe we just set a bar so high that, even though these data-based approaches can perform better than humans in more and more areas, we deem them insufficient because “self-driving cars using similar technology run into pedestrians and posts and we wonder whether they can ever be trustworthy.“ (WSJ, paywall, article on the same topic)

No comments:

Post a Comment