Pranay Manocha "Perceptually consistent Speech Quality Evaluation Metrics"

Abstract: Many audio processing tasks require perceptual assessment. The “gold standard” of obtaining human judgments is time-consuming, expensive, and cannot be used as an optimization criterion. On the other hand, automated metrics are efficient to compute but often correlate poorly with human judgment, particularly for audio differences at the threshold of human detection. In this talk, I am going to present a new method of quality evaluation where we fit a deep neural network to a new large dataset of crowdsourced human judgments. Subjects are prompted to answer a straightforward, objective question: are two recordings identical or not? These pairs are algorithmically generated under a variety of perturbations, including noise, reverb, and compression artifacts; the perturbation space is probed with the goal of efficiently identifying the just-noticeable difference (JND) level of the subject. We show that the resulting learned metric is well-calibrated with human judgments, outperforming baseline methods, as well as shows several traits of human auditory perception.

Carlos Velazquez Vargas "Exploring human learning and planning on grid navigation using arbitrary mappings"

Abstract: From learning to play video games to using novel tools, humans are able to acquire a variety of complex mappings between their actions and arbitrary outcomes. In addition, once they have learned such mappings, they often have to use them sequentially to achieve goals, i.e., planning. In this work we study how the learning of a novel mapping interacts with planning in the context of grid navigation. In order to do so, we developed a computer-based game where subjects have to move a cursor from start to target locations using the keys of their keyboard. Importantly, to more closely resemble the complexity of the mappings that people acquire in their lives, the cursor movement was determined by a non-trivial rule inspired by the movement of the piece of chess known as the Knight. In Experiment 1, we show that participants were able to improve their performance in our task, though not arriving optimally to the targets in the majority of trials. Additionally, we proposed two computational models where the mapping-learning and planning interaction occurs in the form of an action value computation, and then evaluated their fit to our data. Finally, in Experiment 2, we showed that exposing participants to the mapping component of the task without having to plan, provides a performance improvement when exposed to the full task later. Crucially, this improvement does not occur if subjects are exposed to the planning component of the task prior to doing it fully.