So, first of all, thank you so much for joining me to here today a little bit about, Neurosymbolic AI. And just to introduce myself, I'm a staff engineer.I'm Deirdre, a staff engineer at Unlikely AI based in London, where I'm working on, building neurosymbolic AI, which is what I'm going to speak to you about today. And during this talk, I want to make the case that neurosymbolic AI techniques provide a way to create trustworthy and scalable AI. But to get begin with, I wanna explain why that's important and identify some gaps that exist in current AI solutions. So let's imagine a scenario today that's actually quite realistic. So you've been having a stressful time at work. You need a holiday. You're considering where you want to go, and you're looking for some inspiration. You turn to your voice assistant, and you ask for some ideas. You have a little conversation.

You give some criteria. You want to see some sun, some weekend in the next month, somewhere within budget. And the AI very helpfully suggests a couple of places, and Mallorca comes to mind. It jumps out to you. You've never been. You've heard there's amazing beaches, great weather, ticks all of the boxes, and you ask your AI assistant to very helpfully, give you some hotel recommendations, show you some reviews, show you some pictures. You see all of this beautiful imagery presented to you, and you pick a place that suits your criteria and you book it. Or rather, your AI assistant books it for you. And you just pay the bill, confirm, and you get ready for your holiday, and you're really excited. A couple of weeks later, you arrive at the airport. You take out your flight. You're in Mallorca. You arrive to the sun. You're really excited.

You go outside the airport to try and find a taxi, and you speak to the taxi driver. You give them the address of your hotel, but they don't recognize it. And you double check yourself, and it doesn't appear on the map. You're really confused, and so you turn to your voice assistant and you say, what's going on? You have this interaction with your assistant where, the assistant rechecks the itinerary and realizes that it's booked a hotel in Menorca rather than Majorca. And in fact, all of the pictures and reviews that are presented to you during your interaction were actually fabricated. They were entirely hallucinated and constructed by the LLM that you're interacting with.

And you can see by the tone of voice here, it's mimicking a really helpful travel operator, and even offers to sort, paying for a new hotel for you, so you won't be out of pocket, which that's really confusing because this is your AI assistant, and it doesn't have any money of itself.

As you ask, how will you pay the cost of the difference? And the AI assistant agrees with you again in a very helpful tone, and says it doesn't have the ability to make payments, which it didn't realize when it made that offer in the first place. It predicted that this was the kind of response that would fit for a travel operator or helpful travel assistant rather than actually acting within its own guardrails, which doesn't really help you in this situation, without a hotel booking. So this is just a travel booking gone wrong, and we can all imagine this happening. So maybe it's not that big a deal. You can probably find another hotel. You're not in that much of risk. It's not something that's going to really, impact your life. But what if this was health care? What if it was a legal case?

There is some research to suggest that chatbots have at least 40% of their responses conflicting with scientific consensus when it comes to health care. And there has been already cases where citations have been used that were fabricated or hallucinated by an LLM in a court case, I think, against Walmart. In this talk, I'm going to make the case that neurosymbolic AI can be a solution. And in high stakes use cases, how can we develop AI systems we can trust? And so what I'm going to discuss today is, a brief overview of what is AI just so that we can ground this in some context of what I mean by AI in this case. I'm going to go through what is symbolic AI and what is neural AI, and then combine them what is neurosymbolic AI and give some examples. So what is AI?

According to IBM, AI is a technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity, and autonomy. So it's a lot. AI is expected to do a lot. It's expected to be fully human in terms of intelligence. But if we break it down, the way that I like to look at it for this use case, is to consider two features, perception and cognition. And an intelligence system must combine both of these capabilities in order to act with human level intelligence, and it's similar to how humans work, combining perception and cognition. What is perception? So perception in humans means taking sensory input from the environment and mapping them to symbols. So if you see, the shape of a cat, you immediately think this is a cat. You don't have to think about it.

You're just recognizing, the characteristics of a cat, and your brain is making that mapping automatically. And in machines, this is, doing pattern recognition by processing raw data. And for anyone who's familiar with Daniel Kahneman's, thinking fast and slow, this is analogous to system one type thinking. So it's fast, it's instinctive, and you can really quickly process large amounts of data. But on the other side, you don't know how you got to that conclusion, and you cannot control or explain, how you got to that conclusion. When you see a cat, you just know it's a cat. You can't you can post talk kind of make up some characteristics that describe why it's a cat, but, in the moment, it just happens.

And on the other side, you have cognition. And cognition in humans means using background knowledge for activities like abstract thinking, making analogies, reasoning, and planning. And this is very similar to how it works in machines. An example of this would be if you consider what your commute to work would be. You know what transport options are available. You know what the weather is going to be like. So, you know, if it rains, you need to bring an umbrella or avoid being outside. So you use sources of information to come up with some sort of plan that fits some sort of optimization criteria, and you're very active and aware of this process going on in your head. And so this is, in the thinking fast and slow world, similar to system two two type thinking.

So it's slower, it's more deliberate, and you use it for tasks like reasoning and planning. And the important thing here is it can be controlled and explained. So in the past, in the past, before the current, like, LLM dominated based AI, symbolic AI dominated the space. And in some ways, it's so pervasive that we actually don't even consider it to be AI anymore. So if you look at the diagram here, you have this kind of flowchart decision tree type system, where you have a rule and, depending on the evaluation of the rule at each stage, you follow a different path. And so this is the kind of thing that you get in Symbolic AI. So they often rely on expert systems or knowledge bases where people have understood some rules or domains and have codified it.

And in this way, where it's strong is providing transparent, explainable, controllable reasoning. So you can see at every stage, if, if a rule is evaluated, what happens next, and you can see what rules are evaluated. On the downsides, it does lack self learning capabilities. So in order to construct this kind of tree, it's very costly. It requires a human to understand a huge domain and to be able to codify it. So they need to understand both what they're codifying and how to codify it, which is two systems. It's very difficult to scale partially for this reason. And then if changes come about, it's also very difficult to implement those changes because it's so costly to edit the graph, and it's also difficult to represent nuance.

But this is very analogous to cognition that we just discussed. Nowadays, what we consider AI is really LLMs, and they're neural network based. And so I'm sure a lot of people are familiar with this kind of diagram here on the right where you have a neural network, where you have all of these hidden layers that have connections that are learned, during self training. And LLMs are excellent at predicting based on the learned distribution. So like the chat that I showed a few minutes ago around the travel assistant, it can mimic a tone. It can predict what a travel operator is likely to say if they're a helpful travel assistant. It's really excellent at that. But on the other hand, it's not perfect. There are some errors with hallucinations. It's also not explainable, interpretable, or transparent. You have no idea exactly why an LLM says a particular output to a particular input.

And in fact, you can't even repeat that necessarily because it's a non deterministic system. And for this reason, it's not really usable in high trust use cases where you have real world consequences for reasoning and decision making. So it's really strong at perception, which is taking patterns from data, and instinctively mapping them into some, inference, but it performs really poorly at cognition where you do need that, checks and balances. And so how do these compare? On the neural side, the core principle is learning statistical patterns from data, and it's really good at vision, language, and prediction tasks. That's where it's strong. On the symbolic side, it's encoding explicit knowledge using rules and symbols, and it excels at planning, maths, logic, and explanation. Where neural AI really wins out is around pattern recognition, generalization, learning, adaptability, and robustness.

So it can learn from really large amounts of data really quickly and make some inferences. It's very flexible. You see a lot of emergent behaviors, and translation of, like, abilities in one domain to another. It's very adaptable. You just, give it new tasks, and it often performs very well. And it's very tolerant to noise because it is basically using statistics, and it that can help to remove the noise, which are all things that Symbolic AI is quite poor at. It's as we saw before, it's quite difficult to generalize. It doesn't self learn, and it's quite inflexible to changes. On the symbolic side, where symbolic is really strong is about being interpretable. There's really strong abstract reasoning.

The data needs are actually really low with symbolic AI other than the fact that you need a human expert or usually a human expert in the loop to understand, what rules need to be codified. But, otherwise, you need very little data unlike neuro a symbol or neural AI, which needs an awful lot of data. It's very data hungry. And then symbolic AI is also compositional and transferrable, so you can, combine, I guess, like, the decision graphs that we saw before. And it's very similar to, like, programming where you can, compose, and you can build modular systems that can be composed. And the natural conclusion from the tables above is that these approaches have complementary features. So where neural is strong, symbolic is weak and vice versa. And combining both, it also aligns well with how humans combine perception, which is neural, with cognition, which is more symbolic. And so that's always a great plus.

And we can also see neurosymbolic systems evolving organically, even when it hasn't been identified in neurosymbolic AI. So an example I really like to think of is cogeneration, where you use an LLM like Copilot, or Cursor to generate some code. And then, obviously, we know with code, it gets compiled and run. And so you have some, like, IDE or some compiler which is checking that this code is syntactically correct, follows a certain rule, it works. And then you usually have a human in the loop as an additional check to make sure the intentions are being followed. And so how would neurosymbolic AI have helped in the travel assistant case that I started with?

So if the travel assistant was less neural and more symbolic, it potentially would have bound the chosen destination to a variable. So instead of making it possible to confuse Menorca with Mallorca, it would have allowed you to select I want to go to Mallorca, bind this to a variable. This variable could have been used with a more traditional symbolic search engine to search for hotels either on the Internet or on booking.com. And the LLM could then have evaluated these options that came from a symbolic system, and return them to you with links to the source and citing that. And then for booking the hotel, it could have sent the full unedited confirmation to the user, so you could double check that the destination is correct. And another example of a neurosymbolic system that isn't always necessarily tagged with neurosymbolic as the headline is AlphaGo and the alpha family of models.

So AlphaGo is a DeepMind project from a number of years ago where an AI system was trained to play the game Go, which is considered a really complex game where, unlike chess, you cannot generate all possible board states to win. And in 2016, Go very famously beat the world champion Lee Sedol in a historic match, winning four one. And it was unexpected by both AI experts in the industry, even the team who built AlphaGo, and also by experts in Go. It showed, a computer mastering intuitive strategic reasoning in a way that surprised everybody, and people didn't see it coming so quickly. And it combined neural and symbolic components. So there was neural network components in the form of a policy network and a value network, which predicted the probability distributions of of moves given a specific board state and the probability of winning from a particular board state without predicting to the end of the game.

And then the symbolic side was a Monte Carlo tree search, which is a symbolic algorithm for decision making, and it allows you to explore future move sequences by maintaining a tree of states and actions. And, basically, the AlphaGo used the neural components to optimize the symbolic search. And this has been developed further into a a number of family of alpha models by DeepMind, including AlphaFold, AlphaGeometry, and more recently, AlphaEvolve, which are all neurosymbolic systems that combine neural components with some more traditional search methods. And this brings us to unlike the AI and what I'm working on, which is decision intelligence. And for me, I'm particularly interested in the explanation side. So I think it's really powerful that with a neurosymbolic system, it makes it possible to say no and don't know. So, one really big issue with, the current state of AI is that, models can be sycophantic or they can agree too much with users.

And when we use symbolic AI in incorporated with neural AI, we know when to say no and when to say don't know. And so unlike the AI, we use neural components to provide a natural language interface. And then this is used to go into an understanding phase where natural language input from a user is translated into a symbolic a symbolic representation, which is optimized for reasoning. And this is done, in advance, really, to learn some task or understand some document. And then there's an answering phase where an a user can ask a question, again, in natural language, and it's evaluated according to the already understood symbolic representation. Find a yes, no, or don't know answer.

And because you're tracing through symbolic representation, the path to whichever answer you get is known, and it can be audited, controlled, and explained. And so more concretely, what we actually do is in a case like, let's say, understanding an insurance document. So you have an insurance policy, and you want to figure out if claims are covered or not. You can or what we do is take in source documents like a policy and some training scenarios that just help you to figure out, like, what is the purpose of this, document and what kind of rules are we going to be evaluating. We translate it into a symbolic representation, which is optimized for reasoning on those kinds of questions. And there you can fully review and audit this reasoning, and this this, I guess, decision tree, which helps to make the reasoning more guided and controllable.

And then at query time, natural language comes in, is evaluated according to the reasoning graph, know answer is returned, including steps, taken to achieve that answer. And so, to conclude, with this, I wanted to say that, the AI industry is here. And in industries where wrong decisions have real life consequences like insurance, health care, accounting, banking, illegal, it's really important, that we think about how to develop trust with users. And a huge gap exists with the existing neural AI because you have things like hallucinations, an inability to control the reasoning, and a lack of explainability. And neurosymbolic AI, I hope you'll agree, helps to overcome these hurdles because it allows AI to be a little bit more accurate and reliable. It allows you to control what's happening, and it's fully explainable.

And it's important for those kinds of industries, but it's also important for us as humans who are living in an AI first world that we need to be able to trust and understand the systems that we interact with. And we need them to work for our benefit, to make the world better and safer and not just provide convincing stories that sound great but are based on incorrect assumptions or information. And with that, I think we're