Arvind Seshan

Arvind Seshan

Modern applications such as voice recognition rely on the ability to compare signals to pre-recorded ones to classify them. However, this comparison typically needs to ignore differences due to signal noise, temporal offset, signal magnitude, and other external factors. The Dynamic Time Warping (DTW) algorithm quantifies this similarity by finding corresponding regions between the signals and non-linearly warping one signal by stretching and shrinking it. Unfortunately, searching through all "warps" of a signal to find the best corresponding regions is computationally expensive. The FastDTW algorithm improves performance, but sacrifices accuracy by only considering small signal warps. My goal is to improve the speed of DTW while maintaining high accuracy. My key insight is that in any particular application domain, signals exhibit specific types of variation. For example, the accelerometer signal measured for two different people would differ based on their stride length and weight. My system, called Machine Learning DTW (MLDTW), uses machine learning to learn the types of warps that are common in a particular domain. It then uses the learned model to improve DTW performance by limiting the search of potential warps appropriately. My results show that compared to FastDTW, MLDTW is at least as fast and reduces errors by 60% on average across four different data sets. These improvements will significantly impact a wide variety of applications (e.g. health monitoring) and enable more scalable processing of multivariate, higher frequency, and longer signal recordings.

Arvind Seshan

Determining neuronal identity in imaging data is an essential task in neuroscience, facilitating the comparison of neural activity across organisms. Cross-organism comparison, in turn, enables a wide variety of research including whole-brain analysis of functional networks and linking the activity of specific neurons to behavior or environmental stimuli. The recent development of three-dimensional, pan-neuronal imaging with single-cell resolution within Caenorhabditis elegans has brought neuron identification, tracking, and activity monitoring all within reach. The nematode C. elegans is often used as a model organism to study neuronal activity due to factors such as its transparency and well-understood nervous system. The principal barrier to high-accuracy neuron identification is that in adult C. elegans, the position of neuronal cell bodies is not stereotyped. Existing approaches to address this issue use genetically encoded markers as an additional identifying feature. For example, the NeuroPAL strain uses multicolored fluorescent reporters. However, this approach has limited use due to the negative effects of excessive genetic modification. In this study, I propose an alternative neuronal identification technique using only single-color fluorescent images. I designed a novel neural network based classifier that automatically labels sensory neurons using an iterative, landmark-based neuron identification process inspired by the manual annotation procedures that humans employ. This design labels sensory neurons in C. elegans with 91.61% accuracy.

Arvind Seshan

While computers play an increasingly important role in every aspect of our lives, their inability to understand what tasks users are physically performing makes a wide range of applications, including health monitoring and context-specific assistance, difficult or impossible. With Human Activity Recognition (HAR), applications could track if a patient took his pills and detect the behavioral changes associated with diseases such as Alzheimer's. Current systems for HAR require diverse sensors (e.g., cameras, microphones, proximity sensors, and accelerometers) placed throughout the environment to provide detailed observations needed for high-accuracy HAR. The difficulty of instrumenting an environment with these sensors makes this approach impractical. This project considers whether recent advances in indoor localization (Wi-Fi Round Trip Time) enable high-accuracy HAR using only a smartphone. My design, called Location-Enhanced HAR (LEHAR), uses machine learning to combine acceleration, audio, and location data to detect common human activities. A LEHAR prototype, designed to recognize a dozen common activities conducted in a typical household, achieved an F1-score of 0.965. In contrast, existing approaches, which use only acceleration or audio data, obtained F1-scores of 0.660 and 0.865, respectively, on the same activities. In addition, the F1-score of existing designs dropped significantly as more activities were added for recognition, while LEHAR was able to maintain high accuracy. The results show that using a combination of acceleration, audio, and Wi-Fi Round Trip Time localization can enable a highly accurate and easily deployable HAR system.

Arvind Seshan

Millions of people around the world face motor impairments due to Parkinson's, cerebral palsy, muscular dystrophy and other physical disabilities. The goal of this project is to increase the usable surface-area of devices for users with these disabilities by creating a simple, inexpensive, and portable way to enable high accuracy touch interaction with large surfaces such as a table or even a wall. This project uses a novel approach that analyzes the acoustic signals at four piezoelectric microphones placed on the interactive surface to identify sounds related to the same event (e.g., a finger tap) at each of the microphones. ALTo (Acoustic Localized Touch) uses the results of this signal processing to compute the time difference of arrival (TDOA) across the microphones. The collected TDOA data is used to compute an approximate location of a sound source (e.g., a finger tap) using a collection of hyperbolic equations. An experimental evaluation of a system prototype was used to identify a number of software and signal processing optimizations needed to significantly improve accuracy and create a usable system. The results of the research indicate that it is possible to detect the location of a touch with high accuracy. The ALTo prototype achieves an accuracy of 1.45cm in the x-direction and 2.72cm the y-direction which is within the range for the target usage (i.e., those with motor impairments).