The AI Breakthrough Will Require Researchers Burying Their Hatchets

The next AI breakthrough might require ending a longtime rivalry.

For years, artificial intelligence researchers have generally taken one of two approaches when creating problem-solving algorithms: symbolism, or rule-based AI, which is centered on manually encoding concepts, rules, and logic into computer software; and connectionism, which is based on artificial neural networks and digital representations of the brain that develop their behavior organically by comparing many examples over time.

Until recently, symbolic AI was more popular, and neural networks were shunned by many researchers and companies. But in 2012, computer scientists from the University of Toronto had a breakthrough by using deep learning—AI algorithms based on neural networks—to win ImageNet, a famous annual computer vision competition.

Since then, deep learning and neural networks have triggered a revolution in the AI industry and helped to solve problems that were previously thought to be beyond the capabilities of computers. Earlier this year, the pioneers of neural networks were honored with the Turing award, the equivalent of the Nobel prize for computer science.

As neural networks rose to popularity, symbolic AI fell from grace and was pushed to the margins of research. But now, seven years into the deep-learning revolution, we've seen that deep learning is not a perfect solution and has distinct weaknesses that limit its applications.

One group of researchers at MIT and IBM believe the next breakthrough in AI might come from putting an end to the rivalry between symbolic AI and neural networks.

In a paper presented at the International Conference on Learning Representations (ICLR) earlier this month, these researchers presented a concept called Neuro-Symbolic Concept Learner, which brings symbolic AI and neural networks together. This hybrid approach can create AI that is more flexible than the traditional models and can solve problems that neither symbolic AI nor neural networks can solve on their own.

Relying on Labeled Data Is Limiting

"Deep learning is tremendously powerful, and it's not going away. But it has limitations today," says David Cox, Director at the MIT-IBM Watson AI Lab. "One of them is that it depends on huge amounts of data. You need to have alarming amounts of data to train one of these systems, data that must be carefully annotated."

At their core, neural networks are complex mathematical functions composed of thousands of variables. During the "training" phase, the network ingests numerous labeled examples and tunes its variables based on the common patterns it finds among each class of examples. Afterward, when you run a new piece of data through the network, it can classify the data based on its statistical similarity to examples the network has previously seen. Neural networks are especially efficient at tasks such as image classification, voice recognition, and natural language processing, areas where rule-based AI has historically struggled.

But this reliance on data poses a serious difficulty. "For many of the problems we have, we just don't have the amount of data necessary to train deep-learning algorithms as they stand today," says Cox.

As a rule of thumb, the more quality training data you have, the more precise your neural network will become. In many cases, you'll need millions of examples for adequate training.

In fact, the concept of neural networks is almost as old as AI itself. It's mainly because of the increased availability of large amounts of annotated data, and computer resources that can rapidly process this data, that the technique has become practical in the past years.

Explainability is another problem of deep learning. It's hard to investigate and audit the decision made by neural networks, because they are extremely complex and have their own way of developing behavior. This poses a challenge to their application in areas where mistakes can have critical or fatal consequences or where the law requires the adopters of AI systems to provide explanations of automated decisions.

Combining Symbolic AI and Neural Networks

A difficult challenge for AI is the task of visual question-answering (VQA), in which you show the AI an image and ask it questions about the relation between the different elements present. It's difficult because VQA involves elements of image recognition, natural-language processing, and logical reasoning—tasks that are best handled by symbolic AI and neural networks working in tandem.

The MIT and IBM researchers used the Neuro-Symbolic Concept Learner (NSCL) to solve VQA problems. The NSCL uses neural networks to process the image in the VQA problem and then to transform it into a tabular representation of the objects it contains. Next, it uses another neural network to parse the question and transform it into a symbolic AI program that can run on the table of information produced in the previous step.

"One of the interesting things with combining symbolic AI with neural networks—creating hybrid neuro-symbolic systems—is you can let each system do what it's good at. The neural networks can take care of the messiness and correlations of the real world, and help convert those into symbols that a rule-based system can use to operate much more efficiently," Cox said.

Benefits of the Hybrid Approach

The researchers tested the NSCL on CLEVR, a dataset of images of rendered objects used in VQA problems. Previous attempts to solve CLEVR problems with neural network–only approaches yielded impressive results, but they required a lot of training examples, and the developed models performed poorly on edge cases (settings for which the networks have not been trained).

Recommended by Our Editors

4 Reasons Not to Fear Deep Learning (Yet)

What Is Artificial Intelligence (AI)?

AI and Machine Learning Exploit, Deepfakes, Now Harder to Detect

NSCL proved to be able to reach 99.8 percent accuracy on CLEVR with a fraction of the data, because instead of brute-forcing its way through millions of examples, it develops a conceptual representation of the domain, which makes it much easier to tackle scenarios it hasn't seen before. This is important because, in many domains, there isn't enough quality annotated data to train neural networks to solve problems.

Also, to some degree, NSCL solves the explainability problem of neural networks. In traditional neural network models, the AI is provided with the problem and it outputs the result, but there's no clue to how it's solving the problem, so correcting errors is difficult. In contrast, the hybrid system produces a rule-based program that provides step-by-step descriptions of its functions.

"Here, you get to see the program, and you get to step through it and see what it did. If it got the wrong answer, you can see why it got the wrong answer and where it went astray. If it got the right answer, you can verify if it did so for the right reasons. You can understand and audit what came out," says Cox.

The Roadmap to Creating True AI

The AI industry is constantly evolving, and there still isn't consensus on which approach is the best. Disputes break out regularly.

In a recent blog post, deep-learning expert Rich Sutton lays out reasons for sticking to deep learning and discarding methods that try to manually transform human knowledge into computer code.

"[The] only thing that matters, in the long run, is the leveraging of computation," Sutton says. "The human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation." (Sutton's argument was—predictably—disputed by other scientists, including by robotics pioneer Rodney Brooks.)

But as the work of the IBM and MIT scientists shows, the way forward might not be to pit the two approaches against each other but to combine them to make something greater than the sum of its parts.

Artificial Intelligence Develops Its Own Language