The Power of Neurosymbolic AI

We are experiencing unusual levels of activity in AI that directly impacts enterprise decision makers, so I am diving a bit deeper into the science in this month’s edition of the EAI newsletter. The topic this month is on neurosymbolic AI (or EAI for the enterprise), so let’s first summarize the topic.?

In simple form, neurosymbolic EAI is the combination of neural networks and symbolic representation applied within and across the enterprise, preferably in a unified system to achieve optimal performance.?

Neural Networks

Artificial neural networks (ANNs) are inspired by the brains found in humans or animals. The current large language models (LLMs) we are seeing in chatbots like ChatGPT employ deep learning algorithms and autoregressive transformers. In very brief form, deep learning is a powerful technique employed in neural networks that runs many layers of machine learning, and autoregressive transformers are employed to generate human-like text (aka, ‘generative’ AI).

What we are seeing in chatbots is a sophisticated form of natural language processing (NLP) run on a vast corpus of text scraped from the web and presumably other mediums, combined with enormous electric and computing power, specialized chips, and algorithms to mimic human responses in an extremely inefficient manner, for just about any topic one can think of. If it has been written about on the web or in digital form available to the chatbot companies, has been scraped and is part of their data store, then the chatbot will return a reproduction of content based on probabilities. As we’ve seen, some LLMs can be very impressive, especially when the data stored on the topic is high quality, but can also be very inaccurate and even dangerous (see?Europol’s report?on criminal uses of ChatGPT).?

While LLMs do have similarities to how humans think, and they can certainly ‘remember’ far more than any human due to the vast data corpus, chatbots are not able to reason. They can provide intelligence we were not aware of previously so they meet my definition of AI, and can appear to be super intelligent, but don’t actually understand anything.? Melanie Mitchell and David Krakauer recently wrote a paper on the?debate about understanding in AI?that is quite interesting. Melanie is a professor focused on AI and complexity, and David is a professor and President of the Santa Fe Institute , which is where I first met them in 2009.?

In their paper, the authors?suggest that "those who attribute understanding or consciousness to LLMs are victims of the Eliza effect, named after the 1960s chatbot created by Joseph Weizenbaum that fooled people into believing it understood them”. They continue, “More generally, the Eliza effect refers to our human tendency to attribute understanding and agency to machines with even the faintest hint of humanlike language or behavior”.?

"Those who attribute understanding or consciousness to LLMs are victims of the Eliza effect, named after the 1960s chatbot created by Joseph Weizenbaum that fooled people into believing it understood them." — Melanie Mitchell and David Krakauer

I think this is an accurate description of what is occurring at a very large scale due to the large numbers of people who have access to chatbots on the web. Although the technology has improved primarily due to scale, it isn't new. What’s new is that it was unleashed in the wild for public consumption (much too prematurely in my view).

As I’ve mentioned in previous editions of this newsletter (see archives), one of the problems with chatbots like ChatGPT for enterprise use, is legal liability. It’s worth repeating due to the unusual level of risk involved and considerable investment many companies are making. Since these LLMs are scraping millions of copyrighted files and reproducing in full content, at least two types of liability exist. I’m not an attorney and not providing legal advice here, but I am familiar with these issues and I highly recommend enterprise decision makers consult with specialized counsel before adopting these models.?

One risk is the potential copyright infringement. Apart from the serious general copyright issue of reprocessing the knowledge base of humanity without permission or compensation, some chatbot content has proven to be nearly identical to published works by experts. The second liability risk is with false information. I think both issues are potentially fatal for some of the LLM chatbots available today due to the manner in which they collect and reprocess data. The degree of risk for enterprise customers would depend on whether they are integrating the entire chatbot services into their products or simply employing the LLMs to run on their own data. Since LLMs require very large corpus to generate human-like responses, and few if any enterprises have sufficient internal data to generate full content reproduction, LLMs will be limited in the enterprise, but still useful.?

However, LLMs have serious problems when it comes to governance unless limited to high quality data. As I?posted here on LinkedIn?on 3/25/2023, “The model is flawed” (LLMs), Yann LeCun Tweeted on 3/26/2023, “Auto-Regressive LLMs are doomed”, and even Geoffrey Hinton admitted in an?interview on CBS Saturday Morning?(a strong proponent of LLMs), guardrails are “very difficult” and “don’t work very well” for LLMs. Hence the need for rules-based governance to run neural networks for the type of work required of most enterprises, including but not limited to regulatory requirements and compliance. JPMorgan Chase & Co.?is?one of several companies that prohibits internal use of ChatGPT?due to regulatory concerns.?

The risks involved with LLMs such as ChatGPT are many, highly complex, and to some extent unknowable until rigorously tested. In addition, the incentives to release research prematurely before safety measures can be built and tested are exceptionally large, and then places great pressure on others to follow, which is why I decided to?sign the open letter to pause giant AI experiments. Indeed, I would go further than the letter and recommend pausing the existing experiments until safety measures are successfully tested and the Supreme Court of the U.S. has time to make a determination on the?copyright infringement question?as well?as liability protection under Section 230. In the interim, neural networks and deep learning can continue to be applied for a great many other important applications.

Symbolic AI?

The use of symbols for communication dates back to prehistoric times, and the vision of robots?dates back thousands of years, so in hindsight it is unsurprising that the early pioneers in AI would focus on symbolic reasoning.?

Symbols have proven to be an effective way to communicate. Indeed, letters of different languages are simply a form of symbols arranged in various ways to communicate effectively and accurately. Some symbols represent vast quantities of data by today’s standards, including entire philosophies, belief systems, or nations, conveying meaning.?

Symbolic AI was the dominant method in AI in the 1950s – 1990s. In the 1970s symbolic reasoning was applied to rules-based decision making for expert systems, which were widely employed in business during the 1980s to mimic human experts for?if-then?rules, providing precision and efficiency with some automation. Expert systems are a good example of knowledge-based systems, or simply knowledge systems (KS), which is a sub-discipline of AI. My lab in the 1990s was one of few that specialized in KS at the time. In the 2000s expert systems were resurrected and employed by major vendors to improve and automate business logic in enterprise software.?

In 1994, Tim Berners-Lee unveiled his idea for the Semantic Web at the First International WWW Conference, which resulted in the formation of the W3C.?I first became involved with the W3C and related Semantic Web efforts in the late 1990s while operating Global Web Interactive Network (GWIN), which was a learning network for thought leaders and sort of the grandfather to LinkedIn (the early LinkedIn was quite similar in some respects).?

Formal semantics is the study of meaning with roots in logic, the philosophy of language, and linguistics (Cambridge), so the term semantic was technically appropriate, but it was a challenging term for the public from a communications perspective.?The vision for the Semantic Web consisted of metadata, using descriptive standardized languages such as RDF or OWL, which could be read and translated by machines, and hyperlinks for understanding relationships and to provide more accurate searches.?Although the Semantic Web did not scale as rapidly as hoped for consumers, it did influence many large web sites, including search engines, and the technology was adopted widely by enterprises (the semantic enterprise).?

A significant portion of the relatively small group of people who worked on the Semantic Web and W3C standards became well-known professors, founders of graph database companies, and senior scientists in leading companies. Semantic languages, descriptive metadata, and graphs have become critical for providing precision accuracy, productivity, and visualization in EAI. The current state-of-the-art for symbolic AI, expert systems, and the semantic enterprise are a hybrid of symbolic representation and structured languages with neural networks, or neurosymbolic AI.???

Neurosymbolic AI?

Neurosymbolic AI is the appropriate combination and integration of neural networks and symbolic AI for the specific intent of an AI system. As with any other types of system, the more refined and optimized, the better. Those functions and tasks requiring precision data and accuracy should be run with symbolic AI and/or semantic data structure when possible, and those functions that require more generalized AI should be run on neural networks with algorithms like deep learning.?

A recent example of neurosymbolic AI was the combination of Wolfram Alpha (symbolic expert system for mathematics) and ChatGPT (LLM neural network) through a plugin.

Another interesting example of a neurosymbolic system is the?AI Feynman Algorithm, developed by researchers at MIT. The goal was to improve on the?symbolic regression problem?by combining neural network fitting with a suite of physics-inspired techniques, applying to 100 equations from the Feynman Lectures on Physics. They were able to improve the state-of-the-art success rate from 15% to 90%.?

As an illustration of the potential power of neurosymbolic AI, the team found that the use of neural networks to solve equations require more data points?(between 10 x 100 and 10 x 1,000,000). Expressions requiring neural networks are more complex and require much larger datasets, so the ability to apply more efficient methods for equal or greater accuracy can save significant financial and computing resources while also reducing harmful environmental impact. This speaks to the problem of focusing excessively on methods that require vast scale, which has dominated AI research, rather than exploring the most efficient methods.

I found the AI Feynman Algorithm to be an elegant solution to a difficult problem, and one that can likely be applied to many tasks across most industries. It is conceptually quite similar to our approach in our Synthetic Genius Machine (SGM) I disclosed in a?recorded talk?at the ExperienceITNM Conference in New Mexico seven months earlier, which ironically included a staff member in the front row from the Feynman Center for Innovation at Los Alamos National Lab ( Christopher (Chris) Meyers ).?

The important takeaway for enterprise decision makers is that neurosymbolic programs can be applied to solve many problems that can’t be solved by either symbolic or neural networks alone, or in some cases can be solved in a much more efficient manner.?

KYield’s Approach

For the sake of this discussion, we can separate our R&D into two phases. Although significant overlap exists with neural networks, the first phase between 1996-2008 was primarily focused on expert systems, semantic languages, learning networks, and NLP. This work resulted in our KOS (EAI OS), which is our flagship system that is technically viable today and scalable to any size of organization. The KOS provides governance over the entire system with embedded security, prevention of crises, personalized learning, and enhanced productivity, tailored to the needs of each entity through?DANA?(our Digital Assistant).

The second phase of our R&D between 2009-2022 was mostly focused on newer methods of AI, including rapidly evolving deep learning and neurosymbolic AI, which manifest in our SGM. We focused our research on several areas, including accelerating discovery across disciplines, stronger security, and significantly improved efficiency. Although the two systems can be run together or independently, our plan is to integrate the technology from our SGM into the KOS as soon as it can be rigorously tested. It's plausible we could begin to integrate some of the SGM technology into the KOS later this year.?

Conclusion

Neurosymbolic AI offers great potential for enterprise applications, some of which are already in the process of being realized. As has been the case with AI all along, competitive systems require a significant investment in infrastructure, services, and people. However, when applied well business leaders will find neurosymbolic AI can provide impressive advantages in both revenue potential as well as cost savings, especially when adopting refined systems.

Additional recommended resources

Those who want to drill down into neurosymbolic programming, Yisong Yue at Caltech provides a nice lecture here.

1. KYield EAI Principles (article and video interviews)
2. Goldman Sachs report:?Generative AI to affect 300 million jobs?in the U.S. and EU.
3. Recent interview with Geoffrey Hinton?on CBS Saturday Morning (video - 42.30)
4. Eric Schmidt interview?on the consequences of the AI revolution (Amanpour and Company video - 18.00)
5. Risk of ‘industrial capture’ looms?over AI revolution (FT)
6. Data Deluge: Businesses Struggle With TMI (WSJ)
7. AI’s Powers of Political Persuasion?(HAI at Stanford)
8. Rana Foroohar ?at the FT on the?reckless manner?in which some companies have released AI models?
9. Unpredictable emergent properties of LLMs?(Quanta Magazine)
10. Beware the Napster Precedent?(The Economist on LLMs)
11. The Risk of of Pasting Confidential Company Data Into ChatGPT?(Security Affairs)