This website uses third party cookies to improve your experience. If you continue browsing or close this notice, you will accept their use.

Machine Learning and Optimization for Neuroscience

22 May 2018 at 12:00 PM - 1:00 PM

Room 104, Campus on Viale Romania, 32

Speaker: Veronica Piccialli, Tor Vergata

Abstract: In neuroscience, a fundamental theme is the study of brain functioning, for different scopes, such as neurorehabilitation, diagnosis support, and brain activity monitoring in general. The detection of brain state changes plays a fundamental role in this field. Evoked potentials, for example, which are the electrical responses recorded from the brain after specific stimulations, are widely used by researchers and clinicians to support scientific hypotheses, to make diagnoses or to build communication protocols. The study of evoked potentials translates into a hard binary classification problem. On the one hand, solving this problem requires an efficient machine learning algorithm for online training when the evoked potential is used to build a communication interface. On the other, there is also an interest into feature selection techniques in order not only to improve the classifier performances but also to better understand (offline) the underlying cerebral processes that generated the data in order to support scientific hypotheses or diagnoses. In this work, we give contributions in both these aspects. Indeed, in [1],[2] we showed how the filter RRelieFF has the ability to produce a feature ranking that is physiologically correct, both in healthy subjects and in patients that are affected by Amyotrophic Lateral Sclerosis (ALS). This ranking helps in both increasing the classifier accuracy and reducing the dimension of the dataset.

Ongoing research is devoted to the online classification method.  Our method is based on a standard linear classifier, but we exploit in an innovative way the produced hyperplane considering that the protocol is made of repetitions. Combining the knowledge on the protocol with the knowledge on the machine learning algorithm, we obtain a significant increase of the accuracy and we define a mechanism to reduce the number of repetitions for the patient in the test phase, boosting the communication speed without losing the accuracy. Our approach, then, leads to a huge benefit for the patients, and potentially to a new dynamically adjusted protocol. Furthermore, we are also defining an ad hoc machine learning model that already incorporate into the training phase the knowledge on the protocol, adding ad hoc constraints to the standard ones. This allows to further increase the accuracy in the classification and the communication efficiency.

Joint work with Luigi Bianchi, Matteo Cosmi, Giampaolo Liuzzi and Chiara Liti.

[1]L. Bianchi, C. Liti, V. Piccialli (2016). Features reduction for P300 Spellers, Proceedings of the Sixth International Brain-Computer Interface Meeting: BCI Past, Present, and Future, May 30 - June 3, 2016, Asilomar Conference Center, Pacific Grove, California, USA, DOI:10.3217/978-3-85125-467-9-26.

[2] L. Bianchi, M. Cosmi, C. Liti, V. Piccialli (2017). Can feature selection be used to detect physiological components in P300 based BCI for Amyotrophic lateral sclerosis patients? Proceedings of the 7th Graz Brain-Computer Interface Conference 2017, September, Graz, DOI: 10.3217/978-3-85125-533-1-51