Neurobehavioral evidence suggests that human movement may be characterized by relatively stable individual differences (i.e. individual motor signatures or IMS). While most research has focused on the macroscopic level, all attempts to extract IMS have overlooked the fact that functionally relevant discontinuities are clearly visible when zooming into the microstructure of movements. These recurrent (2–3 Hz) speed breaks (sub-movements) reflect an intermittent motor control policy that might provide a far more robust way to identify IMSs. In this study, we show that individuals can be recognized from motion capture data using a neural network. In particular, we trained a classifier (a convolutional neural network) on a data set composed of time series recording the positions of index finger movements of 60 individuals; in tests, the neural network achieves an accuracy of 80%. We also investigated how different pre-processing techniques affect the accuracy in order to assess which motion features more strongly characterize each individual and, in particular, whether the presence of submovements in the data can improve the classifier’s performance.
Why Can Neural Networks Recognize Us by Our Finger Movements?
Galdi E. M.
;Alberti M.;D'Ausilio A.;
2023
Abstract
Neurobehavioral evidence suggests that human movement may be characterized by relatively stable individual differences (i.e. individual motor signatures or IMS). While most research has focused on the macroscopic level, all attempts to extract IMS have overlooked the fact that functionally relevant discontinuities are clearly visible when zooming into the microstructure of movements. These recurrent (2–3 Hz) speed breaks (sub-movements) reflect an intermittent motor control policy that might provide a far more robust way to identify IMSs. In this study, we show that individuals can be recognized from motion capture data using a neural network. In particular, we trained a classifier (a convolutional neural network) on a data set composed of time series recording the positions of index finger movements of 60 individuals; in tests, the neural network achieves an accuracy of 80%. We also investigated how different pre-processing techniques affect the accuracy in order to assess which motion features more strongly characterize each individual and, in particular, whether the presence of submovements in the data can improve the classifier’s performance.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.