Memory which is one of the fundamental functions of the brain is still largely a mystery to neuroscience researchers. Computers over the time have become more and more powerful, therefore researchers, over the years have used them to detect patterns and learn about key processes in the brain. Well, now researchers have discovered ways in which they can use AI techniques to boost memory.
In a recent paper in Nature, Michael J. Kahana and fellow researchers have investigated methods that improve memory. They used a closed-loop system to monitor and decode neural activity from brain recordings. After a series of experiments, they found that they could apply targeted stimulation to lateral temporal cortex (a part of the brain responsible for organizing sensory input, auditory perception, language and speech production) and rescue periods of poor memory encoding.
The subjects enrolled for the study were epilepsy patients, who initially performed sessions of free recall which allowed researchers to collect data to build a multivariate classifier. The researchers fit a logistic regression classifiers to data, producing a model that maps features of neural activity to an output probability of later word recall. They have built a machine learning model that is specialized to each test subject. These classifiers monitor neural activity during memory encoding and trigger stimulation in real time in patterns of neural activity that have been learned by these models.
Results show memory can be improved
The investigations have shown that it is also possible to improve later recall by targeting the lateral temporal cortex. Memory dysfunctions can also be addressed through such a system. Humans remember or forget information depending on neural events that happen during encoding. The study of Functional magnetic resonance imaging (fMRI) signals has shown that activity is different in cortical and subcortical regions which hints if the information is going to be forgotten or remembered.
This research suggests that differences in specific individual’s neural activity can be studied, therefore making it easy to build models that modulate neural activity when there is a high chance that the brain won’t be able to encode successfully. This overall improves the performance by rescuing network activity. Similar research experiments have been tried in the past but with a difference. Most of the previous experiments were of open-loop design, meaning that stimulation was not delivered in response to ongoing neural activity or specific brain states.
Simple machine learning techniques show promise
The results of these experiments proved that subjects’ ability to recall words improved by an average of 15 percent. However, the success largely depends if the stimulation can be timed to rescue periods of poor encoding. This is a successful show of the power of machine learning algorithms that can not only learn to play games but help us decipher memory functions in the brain. Given the fact that machine learning algorithms used here were one of the simplest class of the algorithms, it simply implies that there is a lot of potentials that needs to be exploited.
The researchers also built a classifier that investigates the impact of the stimulation, by using machine learning. The future prospects could be collecting more data by implanting numerous and precise electrodes, that will help algorithms to study patterns in more detail, hence giving us much better models. Given the humongous amount of data, it would be interesting to see how advanced machine learning algorithms such as neural networks could be used for analysis. More the data, the better would be the performance by these models in improving memory.
Inscrutable machine learning algorithms
The researchers do realize that using some advanced algorithms that are available today, would be able to enhance performance. Using neural network algorithms would make it very difficult to know how the results were achieved. Making such progress would not be very useful since it would also add to complications around the brain processes. Hence they were careful in choosing the algorithms, as they worked with simple linear classifiers, that can be easily scrutinized.
New statistical and machine learning methods are produced every year, making more powerful tools accessible to neuroscientists to carry out studies related to the brain. Machine learning can also help neuroscience catch up with other scientific streams such as physics and chemistry. Models in physics are strong enough to predict unseen particles and other phenomena accurately, whereas brain models can still only describe basic forms of perception and behavior. Many neuroscientists around the world are looking at machine learning to be the framework that will help them study the brain. In coming years it may also supplement or even substitute other traditional frameworks like signal processing and information theory.