Neuroscientists of Centrum Wiskunde & Informatica (CWI) and the Netherlands Institute for Neuroscience (NIN) have developed a biologically plausible neural network model that can learn to remember past events in order to use them in the future. The researchers developed their model by combining theoretical principles from machine learning with insights from neuroscience. The results were published on 5 March in PLoS Computational Biology.
Scientists understand how neurons, the smallest computational units of the brain, behave during tasks, but how brains learn to make efficient choices is unknown, in particular when the brain’s working memory is involved. The brain’s working memory is an essential part of intelligence that has been studied for a long time. To develop their model, the researchers created a parallel with complex animal behaviour-learning through reinforcement learning.
For example, in a well-known ‘working memory’ experiment by experimental neuroscientists, a monkey was trained to remember a particular spot on a computer screen where a circle was briefly displayed. If the monkey moved its eyes in the same direction and at the same spot after the circle had disappeared, it was rewarded with a treat. How did the monkey learn to make the right choice? The experiment showed that neurons in the monkey’s brain retained the information about the circle that had disappeared, as a kind of memory cells.
The developed neural network model tries to answer the question how these memory cells can learn to remember task related information. The researchers applied reinforcement learning techniques, such that the neural network model can remember which information is relevant and which isn’t by selecting the signals that turn out to be rewarding for storage in a working memory. The result of this research are relevant for the development of self-learning systems and provides new insights into the workings of connections between neurons.