Sorry, you need to enable JavaScript to visit this website.

The (efficient and parsimonious) decomposition of higher-order tensors is a fundamental problem with numerous applications in a variety of fields. Several methods have been proposed in the literature to that end, with the Tucker and PARAFAC decompositions being the most prominent ones. Inspired by the latter, in this work we propose a multi-resolution low-rank tensor decomposition to describe (approximate) a tensor in a hierarchical fashion.

This paper introduces neural architecture search (NAS) for the automatic discovery of end-to-end keyword spotting (KWS) models in limited resource environments. We employ a differentiable NAS approach to optimize the structure of convolutional neural networks (CNNs) operating on raw audio waveforms. After a suitable KWS model is found with NAS, we conduct quantization of weights and activations to reduce the memory footprint. We conduct extensive experiments on the Google speech commands dataset.

In this paper, we develop a framework to achieve a desirable trade-off between fairness, inference accuracy and privacy protection in the inference as service scenario. Instead of sending raw data to the cloud, we conduct a random mapping of the data, which will increase privacy protection and mitigate bias but reduce inference accuracy. To properly address the trade-off, we formulate an optimization problem to find the optimal transformation map. As the problem is nonconvex in general, we develop an iterative algorithm to find the desired map.

In event-driven sampling (EDS) the signal is represented in terms of a series of spikes at non-uniform time locations. Owing to the limited dynamic range (DR), just like how conventional analog-to-digital converters (ADC) suffer from signal saturation, in EDS a similar manifestation is observed. Namely, when the input exceeds a threshold, no output spikes are generated. Recently, the Unlimited Sensing Framework (USF) was presented to overcome the DR limitation. In USF, the key idea is to fold the signal using a modulo non-linearity so that its DR is fixed.

Analog-to-digital converters (ADCs) are known to suffer from saturation and clipping for inputs exceeding their dynamic range (DR). Recently, the Unlimited Sampling approach addressed this problem by inserting a modulo non-linearity between the input and the ADC. Moreover, a new model called modulo-hysteresis was introduced to enable the recovery of different classes of inputs from noisy observations. ADCs are typically assumed to acquire the instantaneous input amplitude via an inner product with a Dirac delta function.

In this paper, we conduct a cross-dataset study on parametric and non-parametric raw-waveform based speaker embeddings through speaker verification experiments. In general, we observe a more significant performance degradation of these raw-waveform systems compared to spectral based systems. We then propose two strategies to improve the performance of raw-waveform based systems on cross-dataset tests. The first strategy is to change the real-valued filters into analytic filters to ensure shift-invariance.

The notion of a Moreau envelope is central to the analysis of first-order optimization algorithms for machine learning and signal processing. We define a compositional calculus adapted to Moreau envelopes and show how to apply it to deep networks, and, more broadly, to learning systems equipped with automatic differentiation and implemented in the spirit of differentiable programming.