Part 1 of a two-part interview
Brain-computer interfaces enabling paralysed patients to turn thoughts into actions. Brain-inspired computers that far exceed human ability in certain tasks. Brain-like, low energy-consuming chip, with high-computing power. Factory workers using brain-controlled exoskeletons for manufacturing. Precise, personalised deep brain stimulation helping patients suffering from neurological ailments to lead normal lives.
All of the above are already real or expected to become real in the not-too-distant future. Brain technology has the potential to transform our world and the way we work and live.
OpenGov had the opportunity to have a wide-ranging discussion on brain technology and its foundations with Dr. Oren Shriki, Principal Investigator at the Department of Brain & Cognitive Sciences, Ben-Gurion University of the Negev, Israel. He heads the Computational Psychiatry Lab.
We wanted to start with the basics, the science behind the technology. So, the first question we asked was, “At the moment, how well do we understand how the brain works? How much do we need to understand it to develop applicable brain technology?”
Dr. Shriki replied that though we have substantial knowledge about how the brain works, the mysteries are far greater than what is known. Dr. Shriki said, “If you see the amount of knowledge, at all levels, from molecules and cells to the full human brain, there is a lot. But we are far from complete understanding on the very basic questions. Nevertheless, in terms of brain technology, it’s kind of surprising, but what we see is that we can make much progress, we can go quite a long way, even without understanding how the brain works.”
Many medications for the brain were discovered through serendipity. They work but no one knows how they work. Dr. Shriki mentioned the technology of deep brain stimulation, which has seen significant success, as an example. Small devices are implanted very deep inside the brain at a very precise location, at the resolution of millimetres. They help patients with Parkinson’s disease by stimulating their brain. When it works in say a patient with a tremor in his hand, once the device is activated, the tremor is completely gone.
Dr. Shriki said, “This is an already existing technology. There are various theories but no one really understands how it works. The advantage is that some way we can grow without understanding the brain. Nevertheless, I believe that breakthroughs in understanding how the brain works will be translated to important brain technologies. Currently, without understanding exactly what happens in attention deficit disorder, in schizophrenia, in depression, in other disorders we cannot really know how to treat them. Because we cannot understand the mechanism. We just have some treatments that there were discovered by serendipity, like Ritalin, like anti-psychotic medications for schizophrenia.”
But do these treatments discovered through serendipity work equally well across all patients, at all times? There is a lot of variation. For instance, we think of epilepsy as one disease, but there are many, very different, subtypes of epilepsy. When doctors try to find a medication for an epileptic patient, they manage to find a medication that works in the first time for only 50% of the patients.
Dr. Shriki said that sometimes doctors try many times until they find something that works. “Psychiatrists try many different types of medications. But if you push them they will tell you that they don’t have very good principles on how to choose the medication for a particular subject,” Dr. Shriki said.
Overview of research at the Computational Psychiatry Laboratory (CPL)
The research in CPL involves both experimental work with human subjects and theoretical work. The research is aimed at developing new insights into psychiatric and neurological phenomena.
Researchers typically go to mental health centres and measure the brain activity of patients using electroencephalogram (EEG) equipment. They are also asked to perform some cognitive experiments.
Computational models are also developed which accompany the experimental work and drive it. There are computer simulations of neuronal networks, which are used for modeling what might go wrong in certain brain disorders.
The final line of research in the lab is what is called neural feedback and brain-computer interfaces. “So, we not only measure brain activity but we also reflect it back to the subject in real-time and see if we can help the subject learn how to control his brain activity and also use it in certain cases to control other devices like robots”, Dr. Shriki explained.
Two approaches to generating computational models
The bottom up approach starts with mathematical equations that describe the neurons, how they work and interact with one another. Based on these equations, computer programs are written that simulate the activity of the neurons.
“In the bottom up models people guess the pattern of connectivity. They choose random connectivity, or they plug in some rules that modify the connections among the neurons in a way which resembles what we know about the brain. Then they let those models evolve and see what patterns of activity they see, how the network model behaves and whether it resembles certain phenomena that you see in patients”, Dr. Shriki elaborated.
For instance, in a model of epilepsy, you can prepare a network model, with two types of neurons like in the brain, the ones that excite each other and neurons that inhibit other neurons. If the excitation and inhibition are balanced, you will see normal brain activity. If you increase the excitation compared to the inhibition, the model can start presenting epileptic activity. Then you can compare it to epileptic activity in real brains.
In the top down approach, researchers try to guess the computational principles that the brain uses, like optimal information representation or trying to obtain rewards. Starting with networks that mimic the brain, a computational principle is plugged in. This principle dictates the learning of the neurons, and how the neurons change the connections among themselves. The experimental data can then be used to provide feedback and deduce the computational principles the brain uses.
Dr. Shriki said that the disadvantage of these models is that they are somewhat removed from biology. But they are interesting because they can arrive at phenomena like hallucinations, starting with a simple computational principle. The bottom up models show activity that looks like the brain, but the insights that they provide into the underlying computational principles are limited.
Dr. Shriki’s work tends more towards the top-down use of computational principles.
Deep learning – an application of the top-down approach
Deep learning networks are based on the top-down approach described above. They mimic the brain but they are still very simple compared to the brain.
Dr. Shriki explained, “I show the network some inputs and I show it the desired outputs. For example, I show it some pictures and some description of what’s in the picture. Then I want the network to learn and to be able to generalise. So that if I show it new pictures it has never seen before, it would know how to describe them.”
How a computational model can ‘hallucinate’
Dr. Shriki and his team have developed computational models for neuronal networks that evolved to optimise the representation of information, which means that the connections among the neurons change with time in such a way that the neurons learn to represent the information that they are exposed to. These serve as models for healthy brain activity and also give insights into certain disorders.
Dr. Shriki said, “The most important phenomenon that I discovered is that these networks tend to operate near the border between normal processing of inputs and hallucinations. This is kind of strange because you can say what do you mean by hallucinations in a computer or in a mathematical model.”
Usually, the network receives some input, and there’s output in the form of patterns of activity in the neurons. Under certain conditions, you can get patterns of activity in the neurons even without an input. The neurons just talk to one another and maintain their own persistent activity without any external input. This is what Dr. Shriki described as ‘hallucination’. Where you think that you see or hear something but it is not really out there.
When networks try to optimise information processing, they tend to work on the border. “Our theory is that all of us, even healthy people are not very far from this border, not very far from having hallucinations,” Dr. Shriki said, “The thing is that under certain conditions, some processes that take place in the brain may push you beyond the border and you will start having hallucinations. We model for instance the phenomenon of tinnitus, debilitating ringing in the ear. Also, patients with schizophrenia hear voices, they suffer from verbal hallucinations. Our model cannot really present complex verbal hallucinations but we can try to understand the basic mechanisms that drive hallucinations.”
Dr. Shriki’s models can also display the phenomenon of synaesthesia, the condition in which stimulation in one sensory modality elicits conscious response in another sensory modality. This is an unusual phenomenon prevalent in about 1-3% of the population. In most cases, it is associated with colours. People hear sounds and they see vivid colours that are not really out there. This model could explain under which conditions synaesthesia can evolve in the brain.
Regarding treatment applications, Dr. Shriki said, “The unique thing we do is to develop measures of brain activity that recognise the state of the patient. Like we take several minutes of complex brain activity data and we just give you a single number which tell us if your brain is too excited or inhibited. Schizophrenia, autism, attention disorder are also all associated with some imbalance of excitation and inhibition.”
The only way for a psychiatrist to know if a treatment for schizophrenia actually works is to look at the behaviour of the patient and reviewing them day after day. But it’s not objective.
Dr. Shriki’s work aims to provide objective measures of, for instance the connectivity in the brain and the underlying dynamics, and checking whether the medication actually normalises the brain towards healthy behaviour. This will help to make the process of selecting the medication more efficient.
Analysis of experimental data
An important aspect of the work in the Computational Psychiatry Lab, is that the analysis of the acquired data is driven by the theoretical modelling work.
In the 80s, some scientists realised that there are similarities between the brain and physical systems. In physical systems people care about the collective behaviour of many simple elements that interact with one another. There are fields like statistical physics, which are interested in the collective behaviour and in concepts like temperature, pressure, which cannot really be associated with a single atom. Then people realised that there is a similarity between neural networks and those physical systems. Even in the mathematical descriptions they are somewhat similar and ideas can be derived from physics and applied to neural systems.
So, Dr. Shriki, who has a strong background in physics, especially in the area of critical phenomena, uses ideas inspired from physics for analysing the data, specifically physics of systems that operate close to a border. He explained, “For example, you can take a magnetic system, a piece of iron that could be either magnetic or not magnetic. When you cool it down to a certain unique temperature, it would be exactly on the border between magnetic and non-magnetic. Under these conditions, the system has unique properties, and these properties are universal, namely, they can be found in other systems that are critical. Thus, we take tools that are similar to what people use for studying these critical phenomena in physics, and we apply them to activity in the brain.”
In the next part of the interview, we look at recent developments in invasive and non-invasive brain technologies and the scientific challenges, which have to be surmounted for faster development of better technologies. We conclude with a discussion on the ethical aspects of brain technology.