Prof. Lukoye Atwoli explores the latest breakthroughs in neuropsychiatry. Discover the cutting-edge research and insights that are shaping the future of mental health.
“Because the distinction between the two, in my view, is becoming more and more artificial, and the distance between the two is getting smaller and smaller. Psychiatry is getting more and more involved in brain mechanisms, and brain structures that cause behavioral problems, and cause other issues in thinking and emotions. Neurologists are also discovering, on the other hand, that they cannot cut out what we call psychiatric issues in their practice. So with time, my expectation is, and especially after what we will talk about today, that we will use the same mechanisms, the same structures, and the same spaces to understand problems that today we call neurological problems and those that we call psychiatric problems.
So based on this view of neuropsychiatry, the mind is thought to be a product of brain activity, and the brain is like a black box that receives input from the environment. Based on those inputs, it creates and develops an image of the world in which we live. Now, if it creates a good, adaptive image of the world in which we live, then we thrive. If it creates an image that is not in keeping with what is around us, then we struggle a little bit, and therefore mental illness comes from that understanding. This forms the template around which we build artificial intelligence.
For a long time, more than 60 years, scientists have been trying to understand how the brain works, this black box, in order to make us adaptive and thrive. In order to do this, some decided to start developing algorithms that are derived from statistics, mathematics, and data science to try and understand if you give certain input into this black box, what kinds of outputs do you expect to get. So, they modeled this on how they thought and continue to think the brain works.
In 1943, these two fellows developed an algorithm that they figured mimics how a neuron operates because a neuron gets input, and a neuron gives output. Then they have multiple layers of neurons that ultimately give even more complex outputs. So this algorithm that they developed, they called it a perceptron, and we’ll be speaking about this a little bit more today.
But just in general, if, for nothing else, we don’t remember anything else from today, to think about what AI is. Why is it called artificial intelligence? Because it mimics human brain function in two main ways. One is that those neural networks built to solve complex problems are modelled on actual neurons and how they function, using the perceptrons and so on. The second thing is that AI is designed to mimic human intelligence so that it does things that one might think an intelligent human being would do, things in that way.
Formally speaking, artificial intelligence is a field of computing to begin with. It started with computing, as I have just shown you, with programs or algorithms that are developed with the ability to learn, and I should put ‘learn’ in quotes because I’m going to explain what that means, and reason, again in quotes, like humans. So the ultimate goal is to develop algorithms that, when given input, the same way humans get input from our sensory system, gives output that is similar to that of an intelligent human. A subset of AI, known as machine learning, is about developing algorithms that can learn without you going at every step and giving it input. So you give one input, it looks at the data, and it learns from the structures in that data without additional human input. Deep learning is a further subset that involves algorithms like what I just talked about, artificial neural networks. I’ll go into that a little bit more, and they learn from vast amounts of data. So what we call Big Data now.
This is a perceptron on the right. It looks very complex with those symbols as wiggles and x’s and so on, but the ‘x’ represents the input, whatever the input is. Usually, it’s data of some kind. Then the ‘w’ right next to the ‘x’ are weights that are given to the different inputs based on importance. You can decide what importance you want to assign to each weight, and you might give one weight more importance than another and so on and so forth. So there are different relative weights. Then they come to the transfer function that sums up those inputs multiplied by the weights and then takes it into the activation function, which is the decision point. Do we fire, or do we just stay in our current state? Zero or one, that is the language that computers understand. If it has reached a certain threshold, it fires. If it hasn’t reached that threshold, it doesn’t fire, and that input doesn’t result in anything. That is with one artificial neuron, and it mimics the actual human neuron here, which receives input, again, the ‘x’ on the left, and then gives output, the ‘y’ on the other end. In between, there are certain operations, including weighting of those inputs by importance.
That’s just a summary of how the neuron inspires artificial neurons. So when we talk about artificial neural networks, we look at it down here. In the beginning, I have simplified it by saying there’s input, then there are these weights and so on, and then there’s output. But then in artificial neural networks, the middle layers, which are called the hidden layers, can be many. There can be as many as 100 different connections, hundreds and thousands of different connections, depending on the neural network that you’re building and the inputs and operations that you would like to happen. As a result, what happens in the hidden layers is not often apparent to the people giving input and getting output, but it is still a series of operations and calculations and algorithms doing their thing until you get an output. The theory is that does it really matter what exactly is happening in the hidden layers if the output is accurate? So that is what the computer scientists say. It doesn’t really matter as long as the output is consistently accurate given inputs. During the training of the algorithm, you don’t want to concern yourself too much with what exactly is happening inside those hidden layers.
In general, that is a summary of what we do as artificial intelligence. What is AI? AI, as I said, uses programs that use huge volumes of data to classify things based on similarities and differences, to predict future outcomes based on previous data, and to recommend interventions based on previous outcomes. That’s really a summary of what AI does. It is designed to mimic human thinking, human speech, human appearance, human movements, and so on. It is able to perform calculations with vast amounts of data at incredible speeds, and therefore, it’s very good at number crunching and data processing, things that we would struggle with. Finally, it helps to deliver on very human needs, so the solutions that AI is designed to serve are very, very human, for human problems that we encounter, but that would take us a lot of time and resources if we were just to say that let humans do it themselves.
What is it not? Because I think it is important for us to also clarify what AI is not, and it will surprise you to note that, in my view, AI is not intelligent in the classical sense that you would think. Having given you that long description of what AI is, I think you would agree with me that it is not intelligent. It is a series of mathematical calculations which have been designed by a very…”