AI today and tomorrow is mostly about curve fitting, not intelligence

As debates around AI's value continue, the risk of an AI winter is real. We need to level set what is real and what is imagined so that the next press release you see describing some amazing breakthrough is properly contextualized.

Unquestionably, the latest spike of interest in AI technology using machine learning and the neuron-inspired deep learning is behind incredible advancements in many software categories.

Achievements such as language translation, image and scene recognition and conversational UIs that were once the stuff of sci-fi dreams are now a reality.

However, there’s a growing frustration that the software isn’t really “intelligent”, artificial or otherwise. Even as software using AI-labeled techniques continues to yield tremendous improvements in most software categories, both academics and skeptical observers have observed that such algorithms fall far short of what can be reasonably considered intelligent.

Defining intelligence

Debates over the meaning of intellect date back to the Greek philosophers who were primarily concerned with what differentiated mankind from the animal kingdom, not machines. Aristotle famously stated that the mark of an educated mind is the ability to entertain a thought without accepting it. Little did he conceive that his maxim might equally be used to distinguish between biological intelligence and the artificial variety.

Our machines are still incapable of independently coming up with a thought or hypothesis, testing it against others and accepting or rejecting its validity based on reasoning and experimentation, i.e. following the core principles of the scientific method.

While some forms of AI, like adversarial networks, might play opposing goals against each other to reach an optimal result, few would call such algorithmic reasoning ‘intelligent.’

Rather, it’s a more effective means of achieving a defined goal by having two neural networks work at cross-purposes using a common data set, to achieve a more accurate model than either one alone.

The critical intellectual distinction between humans and machines is the ability to define goals and reason towards achieving them. It is also the ability to distinguish between cause and effect, understanding that despite two events often happen in close proximity, doesn’t mean that one caused the other. The correlation-causation fallacy, often seen in public discourse, is one of the most common logical errors.

AI - learning or describing?

An emerging debate among AI researchers is whether current machine and deep learning techniques amount a fundamentally new form of algorithmic reasoning or are merely an extension of longstanding mathematical techniques like descriptive statistics and curve fitting.

In the latter camp is the influential professor of computer science and Turing award winner Judea Pearl who’s new book on the science of cause and effect has ignited a discussion about the future of AI and whether deep learning can lead to anything approaching general human intelligence.

Pearl had some trenchant observations in a recent interview in which he covered both the ideas in his book and views on the state of AI software, including how the inability of current AI to perform causal reasoning is a severe shortcoming. His assessment of deep learning is both unsparing and enlightening (emphasis added),

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.

In essence, despite their cerebral inspiration, deep learning algorithms amount to another, albeit more powerful data analysis tool that is particularly adept at handling vast amounts of unstructured data.

Nevertheless, deep learning is an exceptionally versatile and powerful curve fitting technique that can identify previously hidden patterns, extrapolate trends and predict results across a broad spectrum of problems. One risk with curve fitting approaches that are too good at representing a given data set is overfitting, in which the algorithm fails to recognize normal fluctuations in data and ends up being whipsawed by noise.

Pearl acknowledges that the success of deep learning has surprised even experts in the field, but he worries that it could cause researchers to get stuck in a conceptual rut and jeopardize progress towards general forms of AI (emphasis added).

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?

AI researchers are aligning into factions defined by their view of deep learning with its most staunch advocates unreceptive to any criticism. Pearl characterizes the environment this way  (emphasis added),

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.

Pearl contends that until algorithms and the machines controlled by them can reason about cause and effect, or at least conceptualize the difference, their utility and versatility will never approach that of humans.

He says that it will be impossible to have meaningful dialog with robots until they can simulate human intuition, which requires the ability to understand cause and effect along with alternative actions and outcomes that it might have taken. In short, we're back to Aristotle.

While he might be in the minority, Pearl isn’t alone in recognizing the need for AI to think (sic) beyond deep learning before it can build machines that think like people.

A paper by MIT researchers argues that creating human-like learning and thinking machines requires that they be capable of building causal models of the world that can understand and explain their environment, not just solve problems using pattern recognition.

The paper also contends that such systems must be grounded in both the physical (physics) and social (psychology) sciences to have any capability for intuitive reasoning about the world that would enable machines to “rapidly acquire and generalize knowledge to new tasks and situations.” Much like Pearl, the authors conclude with an exhortation to AI researchers (emphasis added),

Rather than just building systems that recognize handwritten characters and play Frostbite or Go as the end result of an asymptotic process, we suggest that deep learning and other computational paradigms should aim to tackle these tasks using as little training data as people need, and also to evaluate models on a range of human-like generalizations beyond the one task the model was trained on.

Such artificial general intelligence (AGI) is the stuff of dystopian novels and warnings from tech luminaries like Elon Musk and Bill Gates. The worst case scenarios are unlikely to be achieved anytime soon, or perhaps ever, if as Pearl fears, researchers keep refining existing techniques and don’t expand their conceptual horizons.

My take

Although I frequently use the term AI, it’s always with reluctance and out of convention and not conviction.

Much like IoT and cloud, epithets that are equally abused and imprecise, AI is a widely understood, convenient shorthand for a set of techniques that are increasingly powerful, yet fundamentally distinct from human intelligence and rationality.

While I regularly chronicle the impressive applications of today’s ML and deep learning software, I’ve recognized that it is a stretch to call it ‘intelligent’. Perhaps ‘adaptive’ and ‘self-optimizing’ are better terms. Even these come with caveats since the models require extensive human tuning of parameters and structure, as I’ve discussed here and here.

The technology industry writ large needs to have a collective epiphany regarding the state of AI technology.

Yes, celebrate its successes, but acknowledge that far more fundamental research is needed before we’ll have software that can be legitimately called intelligent.

Much work remains to teach machines causality, although Pearl has already provided the necessary system of mathematics using Bayesian networks to describe causal relationships in what he named Do-Calculus.

In the meantime, users of various machine learning tools must be aware of their limitations and of not making categorical inferences from results from a particular problem and data set.