Pattern recognition is a cognitive task that humans are especially good at; identifying a visual or auditory stimulus (and being able to act accordingly) is an essential quality for the survival of living beings. With this in mind, it is natural to think that any neural model that aims at mimicking the mechanisms of the brain must include this characteristic in its dynamics.
A group of researchers, including scientists from IFISC (UIB-CSIC), has published a study in the journal Frontiers in Neuroscience in which they propose a new neuronal model of detection capable of learning and recognizing time patterns in impulses. Both the brain and the proposed neuronal model base their learning capacity on plasticity. The plasticity allows the regulation of the intensity of the connections between the neurons of the network. By means of the creation of new connections and the strengthening of others, it is obtained that the responses in front of the same stimulus vary along time, that is to say, the neural network learns.
The neuronal model proposed by the researchers is a variation of the well-known and proven Leaky Integrate-and-Fire (LIF) but with an added latency mechanism. This latency introduces a delay time between the incoming stimulus in the neuron and the neuron's output signal, thus allowing the intensity of the incoming signal to be coded as a different response times and increasing the computational capabilities of the neuron itself above its simplest version. Similar characteristics are found in real neuronal networks, such as those in charge of the auditory or visual systems. The novelty of the study lies in the addition of a new plasticity parameter: the network can not only adjust the intensity of the connection between neurons, but can also adjust the delay time between input and response.
In order to test the model, the researchers set up a neural network capable of recognizing pulses of patterns but also capable of self-regulating and learning new sequences without supervision. They used two types of data to train the neural network: on the one hand, they simulated a large number of artificial stimuli that the neural network processed to learn and classify. On the other hand, they also used real magneto-encephalogram data obtained from experimental subjects while performing a classification task.
The study concludes that the proposed model is biologically plausible, thus opening the door to a better understanding of how humans learn from repeated sequences in our sensory systems. In addition, the simplicity and low computational cost of the model proposed in the study would allow its implementation on a large scale for future applications in different areas, such as interfaces controlled by brain stimuli, biometric security or early detection of diseases.
Susi, G.; Tioro, L.; Canuet, L., López, M. A., Maestú, F.; Mirasso, C. R. and Pereda, E. Frontiers in Neuroscience 12, (2018). DOI: 10.3389/fnins.2018.00780