How AI Can Learn Better

At the upcoming BrainFair, you’re participating in a discussion forum called The Brain and the Computer – The Interplay Between AI and Neuroscience. You also conduct research at UZH and ETH Zurich on neural learning and intelligent systems. How intelligent are today’s AI systems, for example ChatGPT and other such programs?

Benjamin Grewe: Our current AI systems are pretty impressive. I use such systems daily: to brainstorm, summarize texts or lead me to new ideas and solutions. So in certain ways, they make me more effective and to some extent smarter. But AI isn’t capable of completing tasks completely on its own. I think we’ve now reached the first level of truly interesting artificial intelligence. And at the same time, AI systems are still a long way from being comparable to human intelligence.

What can we do that AI can’t?

Grewe: To some extent, AI systems already know more than we do – at least in terms of their breadth of knowledge. They have information from many different fields at their disposal. For example, they can know things about chemistry that we’ve never heard of. But as soon as it comes to truly understanding complex relationships, AI systems reach their limits. For example, they have trouble understanding concepts – such as objects like cars or people, actions or entire situations. Humans can connect such elements together and then use these connections to explain a situation or make predictions. That’s exactly where today’s AI systems often fail. And that’s why the phenomenon of hallucination exists. We have all likely experienced that an AI system gives an answer that seems linguistically correct: grammar, style, etc. are all fine. But when you take a closer look, you realize the connections don’t make sense. In my opinion, this problem stems from how these systems are trained.

And how is that?

Grewe: You have to realize that behind almost all of the current AI systems there’s a training algorithm. A task is defined as well as a function that measures how well that task is performed. This function serves as the target. Then the network, which consists of millions of parameters, is trained until the target metric is reached as effectively as possible. The problem is that depending on the task, the system can learn completely different things. For example, it can learn to recognize a dog or a cat in an image. But each time the task is performed, all parameters in the network are adjusted. If you applied this process to the human brain, it would mean that every time a new task is performed, every synapse in the brain learns. That doesn’t make any sense to me.

That doesn’t seem very efficient.

Grewe: I sometimes explain it using an example from engineering. A standard car engine has many individual components – like pistons, a carburetor and a fuel pump. Each of these components has a clearly defined function. Now imagine an engine in which each individual part performs every function. It would be extremely difficult to find out which part is responsible for what. And if someone wanted to replace or check a part, they would theoretically have to change the entire engine. That’s not how the brain works. For most tasks, we use only a small portion of our neural resources. For example, when we solve a math problem, we activate certain functions while other parts of the brain are not used. The opposite happens with today’s AI: the entire system is always used.

So does that mean artificial neural networks need to have a better division of labor?

Grewe: Exactly. We need networks with specialized subsystems. Specific parts should act as experts for certain functions. When an AI system receives a new task or is asked a question, it should first decide which parts of its network are relevant for that task or question. Only then should the information be processed there. And when the system learns something new in the process, improvements should be made just to the specific part used – not the entire network.

What does this mean for the future development of AI?

Grewe: A fundamentally new approach is needed – and that’s what we’re working on. We have to move away from the backpropagation algorithm that dominates today and is the method practically all large AI systems use. This algorithm has delivered impressive results, but I don’t believe it’s the right approach in the long run.

You conduct research at the interface of biology and neuroinformatics. What can we learn from the brains of mammals that applies to the development of artificial neural networks?

Grewe: An important question to ask is: if the current machine learning methods are suboptimal, what is the alternative? One possibility is to develop methods that echo how the brain works. That’s why our new algorithms are based on biological principles. These algorithms operate with simple, local learning rules rather than the global differentiation methods found in conventional deep learning. At the moment, we’re especially interested in the pyramidal cells located in the cerebral cortex.

Why is that?

Grewe: The cerebral cortex makes up about 80 percent of our brain. Many researchers attribute our high level of intelligence in part to the large size of this area. Pyramid cells are a special type of cell found in the cerebral cortex with an interesting ability: they can connect information from different levels. On the one hand, they receive inputs representing abstract concepts from higher brain areas. At the same time, they receive signals from sensory regions, for example visual information entering the brain through the eyes and thalamus. As a result, these pyramid cells can make connections between what we see and the abstract concepts we use to interpret our sensory perception.

So pyramid cells are kind of like interface neurons?

Grewe: Precisely. And these connections happen across multiple hierarchical levels. This is what makes it possible, for example, to explain complex situations using just a few abstract concepts. It probably also plays a key role in enabling people to understand new situations. Even if we have never seen the exact same situation before, we can make sense of it by piecing together familiar concepts. In this context, we also talk about corner cases, which are situations in which known elements appear in a new combination. Such situations can pose a problem for a self-driving vehicle: if the vehicle encounters a specific combination of events it has never seen before, its system may fail. If, however, you understand the underlying concepts, you make sense of a new situation by combining concepts you already know.

What’s the current status of the systems you have developed?

Grewe: We’re still in the early stages. At the moment, we’re testing our neural networks with relatively simple tasks. One example is what’s called the N-body problem, which looks at how multiple bodies interact with one another under the influence of gravity. Our systems can predict how these bodies move in space. Although conventional machine learning systems can do this as well, they often fail when the conditions are changed slightly – for example when a body is removed or forces are combined differently. Our systems, on the other hand, learn separate modules for different concepts such as gravity or individual bodies. This enables our systems to combine these modules flexibly and to model new situations more effectively.

Looking at long-term developments, do you think biological and artificial intelligence will move closer together?

Grewe: I think in five to ten years we will see intelligent systems that are capable of explaining complex situations as well as, or even better than, humans. At the same time, they will possess an extensive amount of knowledge. Initially, we will work with and guide these systems. But eventually, they could surpass us in certain areas.

The motto of this year’s BrainFair is 20 Years of Progress in Brain Research. In your opinion, what have been the most important developments in the past 20 years?

Grewe: I’ve been especially interested in studies exploring the parallels between deep learning and brain functions. Our own research findings, however, indicate that the brain probably does not utilize conventional deep learning.

What are the major unanswered questions in the neurosciences related to AI and the brain?

Grewe: The biggest question for me personally is the how the brain’s learning algorithm works at a conceptual level. I don’t need to understand every molecular detail, but I would like to understand the mechanism well enough to reproduce it on computer hardware. When we’re able to do that, we can develop systems that truly understand our world – without hallucinations and with a true understanding of concepts.

What would this mean for real-world applications?

Grewe: The AI revolution has shown how useful such systems are in many fields: from chatbots and self-driving vehicles to robotics and process automation. If we develop algorithms that are able to truly understand our world and correctly connect concepts, they could perform much better than now. At the moment, people don’t trust AI very much. More reliable systems could change that. In the future, we may be able to develop virtual assistants that work at the same level as competent human assistants or researchers.