Neurosymbolic Nerding Out

The idea of thinking machines has always captivated me. In science fiction, studying robotics, taking AI courses in undergrad. There’s just something about the concept of intelligence and creating intelligence. As much as I love using AI techniques to solve problems, it’s really only when I’m connected to this idea of machine intelligence that my brain lights up.

I love the LLM research and applications that we’re hearing about every day, but when you dig into the nature of LLMs it starts to feel little sterile in terms of modeling full intelligence. Next token prediction is elegant, powerful, effective, and… Involves no reasoning at all. We are not going to get to intelligence on the back of the architectures behind today’s LLMs no matter how far they scale (whether that matters in a practical sense is a different question).

Clearly, this is a very basic observation. Because many of the leading AI researchers who got us to where we moved on into this direction years ago. So I started reading. And reading.

Brain fireworks. I’m in love.

Neurosymbolic AI

Depending on who you ask, neurosymbolic AI is decades old or an emerging field of research. Either way, it has a lot of momentum. The goal of the field is to bridge two approaches in AI to build systems that cover a broader spectrum of intelligence than any current systems can.

GOFAI

Good Old-Fashioned Artificial Intelligence, which reflects a lot of early approaches to AI (especially in language) is symbolic. It relies on explicit representations and rules-based systems to create intelligent behaviour. We get a lot of research in logic, ontologies, search algorithms and the like from GOFAI.

The work tended to be labour-intensive, needed a lot of hand-crafted rules, and was really an attempt to encode human knowledge and reasoning into computers. It has some advantages. The systems tend to be precise and interpretable. You can create reproducible results that you can understand, and that’s extremely valuable. After all, a lot of the criticisms of current deep learning systems is that they’re opaque.

The problem is, symbolic AI falls down the way central planning for economies falls down. It’s brittle, unforeseen edge cases are everywhere, generalization is difficult, and these systems don’t handle incomplete or ambiguous data well.?

Statistical AI

Breakthroughs in deep learning are all over the news and have brought us the GPTs of the world. These systems are statistical models of the data they are trained on. Deep neural networks optimize loss functions; statistical measures of how well a network’s output match the true data. These systems are amazing and powerful. Many can be trained without supervision, they identify salient features of data on their own, and the results they produce…. Well. Just check the news. But the failure modes of these models are just as newsworthy. They take nothing away from the power of deep learning or its utility. But they’re not quite the models of intelligence that many people, like myself, really want to see.

System 1 and System 2?

The psychologist Daniel Kahenman popularized two kinds of thinking: System 1 and System 2. Yan LeCun and others have been very clear. If deep neural networks are doing anything like reasoning at all, it’s strictly System 1. That’s not an intuitive judgement, either, the research bears this out. Deep neural networks are not doing symbolic reasoning (a little louder for the breathless folks in the back).?

System 1: Fast, Automatic Reasoning

1. Automatic: Processes information automatically without conscious control.
2. Fast: Decisions and assessments are made quickly.
3. Heuristic-based: Often relies on rules of thumb, mental shortcuts, or stereotypes.
4. Emotion-driven: Emotions and gut feelings can heavily influence decisions.
5. Effortless: Requires little or no cognitive effort.
6. Examples: Recognizing faces, reading emotions, basic arithmetic like 2+2, and simple motor skills are often driven by System 1.

System 2: Slow, Deliberative Reasoning

1. Conscious: Requires active, deliberate cognitive effort.
2. Slow: Takes time to process information and make decisions.
3. Logical: Tends to rely on formal methods of reasoning, following rules or structured approaches.
4. Resource-Intensive: Consumes more cognitive resources and energy.
5. Controlled: Can follow rules, compare options, weigh pros and cons, and reason logically.
6. Examples: Solving complex math problems, making high-stakes decisions, and evaluating complicated situations typically involve System 2 reasoning.

Somehow, humans do both. Though, if you watch the same YouTube channels my kids do, you might wonder if humans do any System 2 at all. But let’s be generous.

System 1 + 2?

Statistical models of intelligence like deep neural networks roughly equate to System 1 thinking. Symbolic models of intelligence roughly equate to System 2 thinking. The field of neurosymbolic AI is looking for ways to create intelligent systems with System 1 and System 2 capabilities. Call it the holy grail of machine intelligence, just one more next step in machine intelligence, call it J.A.R.V.I.S. Whatever it is, the promise is huge. Uniting the strengths of statistical and symbolic AI in a fully integrated way really could give us systems that learn on their own, truly reason, solve a wider range of problems, and do so in a way that is more interpretable and explainable to us. Or so the theory goes.

It's a great theory. Many of the biggest names in AI research seem to think so. And the more I read the more interested I get. The possibilities are exciting. But even more, it’s an intellectually exciting goal.

Taxonomy of Neurosymbolic AI Architectures

Henry Kautz created a taxonomy of neurosymbolic architectures to categorize different approaches to developing the systems. Artur D’Avila Garcez and Luís C. Lamb a great overview of the categories in Neurosymbolic AI: The 3rd Wave (great paper, worth a read):

Type 1: Standard Deep Learning

• Input and output can be made of symbols (there’s some argument about whether this is Neurosymbolic AI).

Type 2: Loosely-Coupled Hybrid Systems

• Neural network integrated with a symbolic problem solver.

Type 3: Task-Specific Hybrid Systems

• Neural network focusing on one task interacts via input and output with a symbolic system with complementary capabilities (i.e. deepProbLog)

Type 4: Tightly-Coupled Localist Systems

• Neural network with symbolic knowledge baked into the training set (i.e. Logical Neural Networks).

Type 5: Tightly-Coupled Distributed Systems

• Symbolic logic rules captured as embeddings and used as a constraint on the neural network (i.e. Logic Tensor Networks, Tensor Product Representations).

Type 6: Fully Integrated Systems

• True symbolic reasoning inside a neural network.

Who to Follow

Neurosymbolic AI is a rapidly growing field with many key researchers contributing to its development. Here are some of the key researchers in the field:

• IBM Research : IBM is a leader in the research and development of neurosymbolic AI technologies.
• Gary Marcus: Gary Marcus is a well-known figure in the field of neurosymbolic AI. He has argued for the necessity of hybrid architectures that combine learning and symbol manipulation for robust intelligence.
• Artur S. d'Avila Garcez and Luis C. Lamb: They have been researching in this area for at least the past twenty years, dating from their 2002 book on neurosymbolic learning systems.
• Melanie Mitchell: Melanie Mitchell is a Professor at the Santa Fe Institute and a leading researcher in the field of neurosymbolic AI. Her research focuses on conceptual abstraction and analogy-making in artificial intelligence systems. She has authored or edited six books and numerous scholarly papers in the fields of artificial intelligence, cognitive science, and complex systems. I've heard Dr. Mitchell on a podcast or two as well, she's phenomenal.
• Ryan Riegel, Alexander Gray, Francois Luus, Naweed Khan, Ndivhuwo Makondo, Ismail Yunus Akhalwaya, Haifeng Qian, Ronald Fagin, Francisco Barahona, Udit Sharma, Shajith Ikbal, Hima Karanam, Sumit Neelam, Ankita Likhyani, Santosh Srivastava: These researchers have contributed to the field with their work on Logical Neural Networks and other topics.
• Pascal Hitzler and Md Kamruzzaman Sarker: They have written a book titled "Neuro-Symbolic Artificial Intelligence: The State of the Art".
• Henry Kautz: He has given lectures on the history of AI, including the rise of neurosymbolic AI.
• Francesca Rossi: She has given invited talks on the topic of neurosymbolic AI.