Dealing with Noise in Speech Signal

by Hynek Hermansky, PhD

Besides a message, speech carries information from a number of additional sources, which introduce irrelevant variability (noise). As discussed in the talk, such noise comes in several distinct forms. Noise with predictable effects can be often suppressed analytically, and we discuss some techniquesfor doing so. Unpredictable effects of expected noise are typically successfully dealt with by extensive training of a machine using noisy data. Such multi-style training is currently the technique of choice in most practical situations, which is often hard to beat. However, we argue for alternative adaptive multi-stream approaches, which could in principle also deal with noises that are unexpected. Our current efforts in this direction are discussed.

Convexity, Sparsity, Nullity and All That…

by Hamid Krim, PhD

High dimensional data exhibit distinct properties compared to its low dimensional counterpart; this causes a common performance decrease and a formidable computational cost increase of traditional approaches. Novel methodologies are therefore needed to characterize data in high dimensional spaces. Considering the parsimonious degrees of freedom of high dimensional data compared to its dimensionality, we study the union-of-subspaces (UoS) model, as a generalization of the linear subspace model. The UoS model preserves the simplicity of the linear subspace model, and enjoys the additional ability to address nonlinear data. We show a sufficient condition to use l1 minimization to reveal the underlying UoS structure, and further propose a bi-sparsity model (RoSure) as an effective algorithm, to recover the given data characterized by the UoS model from errors/corruptions. As an interesting twist on the related problem of Dictionary Learning Problem, we discuss the sparse null space problem (SNS). Based on linear equality constraint, it first appeared in 1986 and has since inspired results, such as sparse basis pursuit, we investigate its relation to the analysis dictionary learning problem, and show that the SNS problem plays a central role, and may naturally be exploited to solve dictionary learning problems. Substantiating examples are provided, and the application and performance of these approaches are demonstrated on a wide range of problems, such as face clustering and video segmentation.

Object Search And Recognition With Mobile Robots

by Luis Enrique Sucar Succar, PhD

Searching for objects is a fundamental task for service robots in domestic environments. This problem includes several sub-problems: (i) estimating the most probable locations where to find the object in the environment, (ii) developing a strategy to explore efficiently the rooms in the environment, and (iii) finding and recognizing the object within each room. In this work we solve each sub-problem. We automatically estimate likely locations of desired objects by using information from Internet, combining four sources: Google, DBPedia, ConceptNet, and Word2Vec. We propose a new heuristic-based strategy for reducing the traveled distance following a two-step approach. Firstly, obtaining an exploration order of the rooms in the environment, and secondly, searching for the object in each room by positioning the robot through a set of viewpoints. To obtain the exploration order of the rooms it is proposed a simple heuristic based on the distance from the robot to the room, the probability of finding the object therein and the room area. For the exploration in each room, it is proposed the integration of horizontal at surfaces locations into the robot map for pose generation. Experiments were performed in simulation, and with a real robot to search an object in a six-room environment.

Integrating Spatial Information in Hyperspectral Unmixing

by Miguel Velez Reyes, PhD

Hyperspectral imaging (HSI) is an imaging technology that provides fully registered spatial and high spectral resolution (radiance, reflectance, or emission) information of the scene in the field of view of the sensor. Hyperspectral remote sensing is currently undergoing a revolution with the appearance and blooming development of hyperspectral imaging sensors available across a number of platforms such as UAV, stand-off, and commercial/military airborne and space-borne systems. It is possible to capture information about a region of interest at high resolution in the spatial, spectral and temporal domains. Therefore it is of great interest the development of models and algorithms for hyperspectral image processing that fully exploit the information in all three domains. An important problem in hyperspectral image processing is hyperspectral unmixing. Spectral signatures collected with hyperspectral remote sensors can be modeled as the convex combination of the spectral signatures of the materials in the sensor field of view. The distinct materials associated with the surface are called endmembers, and the fractions in which they appear in a pixel are called abundances. The unmixing problem refers to the inverse problem of finding the number of endmembers, their spectral signatures, and their spatial abundances from a given hyperspectral image scene. This is an ill-posed inverse problem. The talk will present how spatial information can be used to constraint the unmixing problem and improve the quality of the retrieved model from the data. These ideas are exploited in developing an extension to the nonnegative matrix factorization that adapts to the spatial characteristics of the image to improved hyperspectral unmixing. We illustrate potential benefits of these approaches with real data sets.