Human navigation and spatial memory: from behavior and neural representations to assistive and rehabilitation tools

 Dr. shachar maidenbaum

Orienting ourselves, knowing where things are and how to get to them are critical parts of our lives. To perform these tasks we must generate and maintain internal representations of our surroundings, constantly updating them based on our perception. In my talk I will first review human spatial behavior and its underlying neural basis, and then focus on two specific representations that I recently found using intracranial recordings in humans while they performed spatial memory tasks in virtual reality. The first of these representations is a population signal for the human grid network which correlates with human spatial memory, and the second is a representation for heading direction. I will present theories for these signals’ underlying basis, their relation to other levels of neural representation, and their implications for navigation and designing environments. I will then demonstrate how these signals, and associated behavior, are modulated by external perceptual input (e.g. navigating with different levels of fog) and by internal idiothetic input (e.g. navigating in the real world with Augmented Reality). Next, I will turn to discussing more generally the potential of Augmented Reality for the field of rehabilitation, and demonstrate its advantages for creating tools for both practical rehabilitation of perceptual and spatial skills and for supporting and augmenting smart environments. Finally, I will conclude with a broader view of human spatial perception and interaction, and its variance across different populations and clinical conditions. 

Speaker: Michael Michelashvili

M.Sc. student under the supervision of Prof. Lior Wolf

Monday, January 27^th 2020 at 15:30

Room 011, Kitot Bldg., Faculty of Engineering

Semi-Supervised Monaural Singing Voice Separation With a Masking Network Trained on Synthetic Mixtures

Abstract

We study the problem of semi-supervised singing voice separation, in which the training data contains a set of samples of mixed music (singing and instrumental) and an unmatched set of instrumental music. Our solution employs a single mapping function g, which, applied to a mixed sample, recovers the underlying instrumental music, and, applied to an instrumental sample, returns the same sample. The network g is trained using purely instrumental samples, as well as on synthetic mixed samples that are created by mixing reconstructed singing voices with random instrumental samples. Our results indicate that we are on a par with or better than fully supervised methods, which are also provided with training samples of unmixed singing voices, and are better than other recent semi-supervised methods.