The WAVES Laboratory
Hayder Radha | radha@egr.msu.edu | www.egr.msu.edu/waves/index.html
The WAVES Lab focuses on the theory and application of multidimensional signal processing, communications, information theory, and related disciplines, which include machine learning and deep learning. Current application areas that are being investigated by the WAVES Lab include autonomous and connected vehicles, big-data, social networks, sustainability, visual communications, and wireless networks.
Autonomous and Connected Vehicles
The WAVES lab is currently leading a major effort in the area of autonomous and connected vehicles. One focus is on developing advanced deep learning solutions to enable vehicles “learn” how to drive using a variety of sensing modalities, including cameras, radars, and Lidars. Novel frameworks for the fusion and sharing of sensing data within an autonomous vehicle and/or among multiple connected vehicles are being explored. The ultimate goal is to enable vehicles achieve full awareness of their environments and to exploit that awareness for “eyes-off” and “mind-off” levels of autonomous driving. New approaches for optimizing active safety are being explored and developed.
Multi-vehicle sensor-data fusion among autonomous and connected vehicles is one area being explored by the WAVES Lab. In this example, an adjacent vehicle to the U-Haul van fuses critical visual information captured by the van (the pedestrian) with its own visual information for improved active safety.
Environmental Data Recovery and Image Super-resolution via Manifold Learning and Signal Sparsification
Manifolds are smooth low-dimensional surfaces (e.g., a sphere) that reside in high-dimensional spaces. We have been collaborating closely with environmental engineers and scientists to verify and exploit novel manifold-driven hypotheses regarding long-standing problems in sustainability. For example, we are studying and exploiting the observation that manifolds can capture many natural phenomena, such as the variation of temperature over a watershed or the complex distribution of wind speed and direction over a lake. Such observation opens the door for new compelling solutions in recovering invaluable environmental data that can’t be measured directly. In that context, in collaboration with the Computational Hydrology & Reactive Transport Modeling Laboratory (Directed by Professor Mantha Phanikumar), we have developed a new manifold-learning based framework for assimilating geophysical and meteorological data in earth system models. The manifold-based framework significantly improves the accuracy of earth system models.
More advanced methods for manifold learning are being explored and developed for image super-resolution (SR) applications. These manifold approaches are being combined with signal sprsification based frameworks to develop novel SR algorithms that are capable of achieving superior visual quality.
Big-Data Representation and Multi-/Hyper-spectral Image Compression using Tensors
Tensors, which are extensions of 1D vectors and 2D matrices, provide a powerful analytical tool for big-data analysis. Tensor decomposition also provides an efficient representation for massive data that exhibits correlation across multiple dimensions (e.g., spatial and temporal). We have been developing novel tensor-based approaches for representing big-data. One application area is multi-spectral and hyper-spectral image data analysis and compression.
Video Super-resolution in Conjunction with Video Compression
In collaboration with Google, the WAVES Lab has been developing novel approaches for video super-resolution embedded within video compression systems. This research, which has been supported through a Google Faculty Research Award, has generated significant improvements to the visual quality of video frames at the output of video codecs.
Recommendation Systems
Recommendation systems play a pivotal role and a broad range of applications. We have focused on challenging problems in this domain including the well-known “cold start” problem associated with recommending new items to new users (or simply recommending any item to a new user with an unknown profile). We have explored and developed novel algorithms based on matrix factorization using side information to address the cold-start challenge in recommendation systems.
