Current Research Interests of Prof. Keshab K. Parhi

Machine Learning: VLSI Architectures and Applications

Research is directed towards low-energy architectures for machine learning systems for both traditional machine learning as well as deep learning. Some key approaches include incremental-precision based feature computation and classification approaches. Other approaches include deep learning architectures using permuted diagonal inter-layer communication networks. Numerous applications of machine learning are investigated for healthcare including applications in neurology, psychiatry and ophthalmology.

Hardware Security: PUFs, TRNGs, and Reverse Engineering

The goal of this research is to use physical unclonable functions (PUFs) for counterfeit prevention of integrated circuit chips and devices, and to simultaneously authenticate devices and users. The objective is to design PUF circuits based on MUX and SRAM PUFs that cannot be hacked easily. Other objectives include designing digital circuits that are harder to reverse engineer.

Other projects in collaboration with Prof. Chris Kim involve design of integrated circuits for true random number generators and physically unclonable functions.

Seizure Prediction and Detection from EEG

Epilepsy is the second most common neurological disorder, which affects 0.6 to 0.8% of people in the world. Approximately 75% of the patients with epilepsy achieve partial or sufficient control over seizures from medication or resective surgery. However, for the remaining 25% of the patients, no treatment is currently available. If there is a way to predict occurrence of a seizure, it could sufficiently enhance the therapeutic possibilities, leading to a better quality of life of the patients.

The general goal of this project is to propose a patient-specific algorithm, which can predict occurrences of an epileptic seizure in advance. Specifically, this project intends to develop an algorithm to classify EEG (electroencephalogram) signals before a seizure onset from those during ordinary conditions with high sensitivity and a low rate of false positive.

In addition to predicting seizures, we are also working on design of classifiers for seizure detection.

This work is currently pursued in collaboration with Dr. Thomas Henry, MD, of Neurology.

Mental Disorders Discovery: Biomarkers and Brain Connectivity

The goal of the this research is to extract signal processing based features from magnetoencephalograms (MEG), functional magnetic resonance imaging (fMRI), and diffusion MRI (dMRI) to identify mental disorders such as schizophrenia, borderline personality disorder (BPD), major depressive disorder (MDD), and obsessive compulsive disorder (OCD). Various classifiers are used to identify biomarkers. We are very interested in understanding the changes in brain connectivity associated with these disorders. These research topics are carried out in collaboration with various faculty members in the Psychiatry department in Medical School at the University of Minnesota. In particular, our research efforts on BPD and MDD are carried out in collaboration with Dr. Kathryn Cullen, MD. For OCD, we collaborate with Dr. Gail Bernstein of the Psychiatry department. Our work on diffusion imaging and anatomical connectivity is in collaboration with Prof. Christophe Lenglet of the Center for Magnetic Resonance Research (CMRR). Work on MEG based language understanding in Schizophrenia was carried out in collaboration with Prof. Massoud Stephane, MD, formerly of Brain Science Center at the University of Minnesota.

Automated Retinal Imaging from Fundus and OCT Images

Diabetic retinopathy (DR) is the leading cause of blindness in people of working age in the developed world. The blindness due to diabetes costs US government and general public $500 million annually. A WHO collaborative study projected that the global diabetic burden is expected to increase to 221 million people by 2010. However treatment can prevent visual loss from sight-threatening retinopathy if detected early. In order to address the impact of diabetes, screening schemes are currently being put into place based on digital fundal photography. However, there are concerns regarding the cost of any screening scheme used for detecting sight-threatening diabetic retinopathy in the population. One of the greatest sources of expenditure in setting up any diabetic retinopathy screening program is the cost of financing trained manual graders. If automated detection programs are able to exclude a large number of those patients who have no diabetic retinopathy, it will reduce the workload of the trained graders and thus reduce cost.

Our current research interest is focused on automated diagnosis of diabetic retinopathy using digital fundus images. The images will be graded no DR, mild DR, moderate DR or severe DR based on the type, quantity and the area of lesions present in the eye. Feature extraction is the first step in the classification of these images. The challenge lies in extracting robust features as the image color varies from patient to patient.

In optical coherence tomography (OCT) imaging, the emphasis is on denoising, segmentation, and correlation of layer thicknesses to visual acuity.

This work is in collaboration with Dr. Dara Koozekanani, MD, PhD, of Ophthalmology department at the Medical School.

Molecular Signal Processing/DNA Computing

This research attempts to understand synthesis of digital signal processing functions through molecular reactions, where inputs and outputs are proteins or chemical molecules. One example of implementation involves DNA strands. Synthesizing signal processing functions in biochemical and biomolecular systems will enable biosensing, drug delivery, monitoring and controlling rate of therapy or treatment. Efforts are directed towards implementation of FIR and IIR digital filters, FFTs, and equalizers using chemical reactions. Efforts are also directed towards implementation of iterative computations through the molecular reactions.

High-Speed/Low-Power VLSI Digital Signal Processing Architectures

Various wireless communication technologies have led to tremendous increase in demand for mobile processing devices. Unlike wired devices that are optimized in favor of performance, minimization of power/energy consumption while maintaining a certain level of performance is a critical concern for wireless devices with limited energy capacity. Due to the prevailing use of intensive digital signal processing units and communication blocks in mobile devices, their low-power design is very important. Another area of interest is design of high speed decoders that can be operated at 10-100 gigabits per second. These decoders form the backbone of wired (copper and fiber) or wireless internet.

Yet another area involves use of stochastic computing for design of highly fault-tolerant DSP circuits for nanoscale CMOS technologies.

We are also pursuing low-power architectures for biomedical applications. Specifically efforts are directed towards low-power feature extractors and classifiers.

VLSI Polar Code Decoders

Polar codes, as the first provable capacity-achieving error-correcting codes, have received much attention in recent years. Our research is concerned with high-speed, low-latency and low-energy architectures for polar code decoders. Both successive-cancellation and belief-propagation decoders are considered.

Computing using Stochastic Logic

The goal of this research is to implement polynomials, functions, digital signal processing and machine learning systems using stochastic logic. Stochastic logic circuits are ideal where area is highly constrained and speed is not a primary concern. Our work has led to more efficient FIR digital filters using scaled stochastic implementations, stochastic implementations of polynomials and arithmetic functions using Horner's rule and factorization, and machine learning classifiers based on radial basis functions and artificial neural networks.