Interview Of The Week

Interview Of The Week: Barak Berkowitz, Neuro-Symbolic AI Expert

Barak Berkowitz, a veteran technology entrepreneur, has established many businesses during his 30+ years in the tech sector.  He is currently Chair of the Board of Resilience Labs, an online behavioral health company, a board member and investor in UK-based Unlikely AI, which specializes in neuro-symbolic AI, and a consultant to venture capitalists and large companies.

Most recently, Berkowitz was the Director of Operation and Strategy at the MIT Media Lab.  Before the Media Lab, Barak was CEO of Evi, the virtual personal assistant acquired by Amazon, and is now the intelligence in Amazon Alexa.  He was a founder or senior executive in numerous start-ups, including Wolfram Alpha, Six Apart, OmniSky, The Go Network, and Logitech.  Before Logitech, he spent over nine years at Apple USA and Apple Japan, leading consumer marketing programs; before Apple, Barak was the founding manager of  Macy’s Computer Stores.

Berkowitz recently spoke to The Innovator about why neuro-symbolic AI should be on the radar of corporate executives.

Q: What is neuro-symbolic AI?

BB: Neuro-symbolic AI relies on language and reasoning, rather than statistical predictions, to represent knowledge and uses rule-based systems and logical inference to derive conclusions. The concept of symbolic systems dates back to the early days of AI and gained prominence with the development of expert systems, knowledge-based reasoning, and early symbolic language processing techniques.

To understand neuro-symbolic AI, it is necessary first to understand how it relates to the LLMs [large language models] and Generative AI that dominate the market today.  LLMs have certain characteristics. One is a lack of transparency. They are a black box. Nobody, including the people who create them, knows why an LLM model gives the exact answer it gives or why it makes a particular decision. We often talk about LLMs hallucinating. The reality is that everything an LLM does is hallucination, and while it can reflect what is happening in the real world, it often does not, so the challenge is that there is no way to make this type of system safe, reliable, and auditable.

Neuro-symbolic AI takes the best qualities of LLMs and other generative models—the ability to be creative and generate interesting language and images, and posit unique approaches to problems—and combines these with more reliable and trustable symbolic AI, which can check the output against actual records to affirm that it is correct, show the reasoning, and allow that reasoning to be examined. It is the best of both worlds: leveraging LLMs’ creative and generative capability and the symbolic approach for fact-checking and error-correcting.

Q: If neuro symbolic AI is more accurate, why are we using LLMs?

BB: The breakthrough with LLMs makes it possible to create better neuro-symbolic AI models. A decade ago, symbolic AI’s challenge was that everything had to be hand-built and the ontology filled out by experts. It was a slow, arduous process. Without all the knowledge of the world, symbolic systems would be unable to answer questions. This is one of the problems with a system like Amazon Alexa, { a virtual assistant technology.] It sometimes misunderstands a question that is being asked or uses awkward phrasing when giving an answer. LLMs can accelerate the feeding of the world’s knowledge into neuro-symbolic models and make them extremely accurate. Combining with LLMs gives symbolic AI the ability to understand human input better and be far more fluent. Combining the technologies makes perfect sense; a lot of the challenges of neuro-symbolic models can be overcome with LLMS, and LLM’s lack of audibility can be overcome by neuro-symbolic AI.

Q: Are we going to see combined systems on the market?

BB: Combining the two is a quickly growing field. Unlikely AI is one of the companies working on this. (I am a board member.) The founder, William Tunstall-Pedoe,  has a unique background in neuro-symbolic AI.  William is best known for his key role in creating Alexa after Amazon acquired his previous voice assistant startup, Evi Technologies, in 2012; several groups are taking different approaches. Almost everyone in the field agrees that LLMs alone will not solve AI’s current limitations. We will end up with some compound system.

Q: If corporations have already rolled out AI networks, can they superimpose compound systems, or will they have to rebuild from scratch?

BB: They can probably superimpose them. Most approaches to neuro-symbolic AI are agnostic to the LLMs that are used. But so far, not that many companies have rolled out AI networks, or if they have, they have had to roll them back because they found them to be inaccurate. Going forward, I believe that neuro-symbolic AI will be used for applications that have high value, such as financial transactions and medicine, factual question answering  where it is important to have an auditable trial. In Europe, the AI Act will require this. But even in countries that don’t have specific regulations, business rules – and the public markets – will require it. Every business will have to use auditable AI that is guaranteed to be as accurate as current knowledge allows, so LLMs will have to have the kind of error-correcting capability that neuro-symbolic AI provides. While you can’t buy off-the-shelf compound systems today, the market is moving extremely fast. These systems are coming very soon.

Q: Will compound systems still require humans in the loop?

BB: You will always need some level of human feedback in the loop, but you won’t need humans to check each answer. Compound systems using neuro-symbolic AI will be able to provide companies with confidence in automated answers.

Q: What advice do you have for corporate leaders?

BB: Leaders need to understand both LLMs and neuro-symbolic AI. Search for vendors that can provide training and customize symbolic AI for your business. Be cautious about adopting LLMs too fast because that can lead to errors and open your company to liability and significant problems, but also be wary about rolling out too slowly because if you are not applying AI to your business in a significant way, you won’t remain competitive. There are certain areas of business, like customer service and call centers, where you can build up your AI knowledge, but start now to develop a plan to adopt a compound system that includes neuro-symbolic AI’s error correction and can guarantee the value and accuracy of your data. The next thing we will see are multi-modal AI systems that will hear, see, smell, and touch. These advances will be on the market in the next few years, and it will be critical to have auditable systems on top of those systems as well.

This article is content that would normally only be available to subscribers. Sign up for a four-week free trial to see what you have been missing.

To access more of The Innovator’s Interview of the Week articles click here.

About the author

Jennifer L. Schenker

Jennifer L. Schenker, an award-winning journalist, has been covering the global tech industry from Europe since 1985, working full-time, at various points in her career for the Wall Street Journal Europe, Time Magazine, International Herald Tribune, Red Herring and BusinessWeek. She is currently the editor-in-chief of The Innovator, an English-language global publication about the digital transformation of business. Jennifer was voted one of the 50 most inspiring women in technology in Europe in 2015 and 2016 and was named by Forbes Magazine in 2018 as one of the 30 women leaders disrupting tech in France. She has been a World Economic Forum Tech Pioneers judge for 20 years. She lives in Paris and has dual U.S. and French citizenship.