Human pluripotent stem cells (hPSCs) are cells capable of self-renewal while differentiating into any cell type in the body. The differentiation into skeletal muscle progenitor cells (SMPCs), results in mature skeletal muscle cells. As the potential of deriving SMPCs from hPSCs has been further researched for medical application, the problem encountered is how to generate fully functional skeletal muscle cells from hPSCs experimentally. Currently, there are many existing protocols (e.g., HX, JC, MS Protocols) for such differentiation, known as myogenesis, but each has problems that need to overcome (e.g., time-consuming, not fully functional cells, resulted in other cell types). Our goal is to figure out what genes and metabolites are involved in the natural process of differentiation (from prenatal skeletal muscle progenitors to satellite cells) and apply the knowledge experimentally. To address this, we present the use of computational methods for the biological network-based integration of transcriptomic and metabolomic data. By analyzing data for both in vivo process and in vitro protocol results, we can find potential genes and metabolites differences, relating from the transcriptomic level to metabolomic level, revealing the inadequacy in the experimental protocols, presenting probable genes to improve the results. This involves RNA sequencing data analysis (e.g., single-cell, microarray) of human and mouse cells from embryonic to postnatal stages, with Seurat (pseudo-time analysis), Monocle, and AFFY, which processes/analyzes the data revealing potential genes. Then further analysis with UKIN, guided network propagation, to rank and identify the most probable genes, and perturb-Met to analyze metabolites involved. This is a time and cost-efficient way to find most probable genes directing the differentiation, which can then be tested and verified experimentally. If successful, it can be further developed into potential treatments much more effective and efficient than available medications and technologies for presently incurable musculoskeletal diseases.

Audio description (AD), an additional narration track that conveys essential visual information in a media work, is imperative for improving video accessibility for people who are blind or visually impaired. While large streaming services such as Netflix, Disney+, and AppleTV have begun offering AD on new titles, current processes are manually done. Movies can take more than 60 hours to describe, with a cost of $10-$70 per minute. To investigate the need for AD, we conducted extensive user research with people in the blind and visually impaired community. We learned when they use AD, how they use it, where they use it, and most importantly, what they value in a high quality audio description experience. Existing literature regarding AD does not address these questions of user preferences, no projects have specifically targeted the area of user-generated content or smaller budget video content, and there is minimal existing work on automating the AD process. We received over 100 survey responses and conducted 40 interviews with stakeholders, including leading industry accessibility experts. Of these interviewees, 18 identified as blind or low vision. Our findings show that the most prominent challenge is the lack of available AD. Some interview participants preferred brief descriptions, wanting to fill in the gaps with auditory information, while others favored longer, more expressive audio descriptions. Based on our interview insights, we developed an audio description software to automatically describe videos from a user-provided link. For our prototype, we identify key frames, use existing APIs from Microsoft and Google to describe and read the text from each frame, and use text-to-speech to generate a second audio track. Through this project, we have extended knowledge of audio description preferences and developed a service to provide automatic audio descriptions based on novel user insights.

Hate speech classifiers, which are machine learning models that detect hate speech, are important tools for content moderation online and help keep online communities safe. However, these models show evidence of bias against certain groups of people. For example, current research shows that state-of-the-art hate speech classifiers show significant bias towards African American Vernacular English (AAVE). In response, new research has produced novel datasets with more complex label spaces compared to previous hate speech classification datasets. By having additional labels (i.e. more than a single label determining whether the text is hateful or not), these datasets aim to capture more social context embedded in natural language. In this project, we investigate this aforementioned claim. In particular, we investigate which features of language are the most salient in the decisions they make based on the datasets that they are trained on. We use multiple methods to determine feature saliency to discover biases in hate speech classifiers. Through these methods, we introduce a set of words that hate speech classifiers determine as highly salient which represent these biases. We also investigate whether these sets of words differ across datasets with varying label space complexity to understand whether these biases persist regardless of dataset. Finally, we will discuss the implications of the results and address future directions for the project.

Quantum Machine Learning (QML) has shown early promise over the last few years. From simple AI algorithms to sophisticated neural networks, quantum computers have produced results that are as good as or better than their classical counterparts. However, all of these models have to deal with the memory bottleneck, which is caused by the limited number of qubits in near-term quantum devices. We instead propose a hybrid neural network that works by sandwiching any QML algorithm between two classical neural networks, using PyTorch. The design allows for an automatic scaling of quantum algorithms to inputs and outputs of any size, addressing the bottleneck issue, but it also provides an easy way of comparing classical algorithms to quantum ones and an expandability to other, more advanced classical scenarios. Additionally, the software supports the usage of configuration files, which allow for fast-paced testing of basic hypotheses, without the need of writing custom code.

Alzheimer’s Disease (AD) is a neurodegenerative condition that affects more than 50 million individuals worldwide. The progression of AD is hallmarked by the buildup of beta-amyloid plaques and neurofibrillary tangles, leading to neuronal death on a large scale. In the early stages of AD, the hippocampus is disproportionately affected by this heavy neuronal loss. The genetic elements and pathways contributing to this AD pathology are still poorly understood. Excitatory synapse assembly (ESA) processes have been previously shown to be affected by the pathology of AD, leading us to investigate the expression patterns of genes involved in ESA in different brain regions. A gene ontology (GO) analysis was conducted to isolate individual genes involved in ESA and analyze their relation to beta-amyloid plaque and pTau neurofibrillary tangle severity in patients who presented signs of dementia due to AD. ESA gene expression was shown to be strongly correlated with increasing levels of pTau in all brain regions except for the hippocampus, where there was no correlation, in a linear regression analysis. This implies that the hippocampus has a unique response to AD pathology with regards to ESA gene expression among the different brain regions. The inability of hippocampal cells to express neuronal repair genes in the presence of severe neurofibrillary tangles requires further analysis and could eventually confer a novel target for AD therapeutics.

Interactive add-ons to computers offer the potential to enhance user experience and productivity, but these peripherals are often expensive and uncustomizable for normal users, designers, and differently abled individuals. AcousTickBoard uses cardboard, a ubiquitous household material due to widespread online shopping deliveries, to address accessibility in human-computer interaction (HCI) through tangible widgets that are both cheap and easily reconfigurable. First, we created the design of a pressable button made of cardboard that emits a “tick” sound when released. Then, we explored different ways of modifying the cardboard buttons in order to produce unique enough acoustic signatures to be independently recognized such that multiple buttons could be used at the same time. The modifications made to the buttons generally involved adding or subtracting material in order to manipulate the “tick”’s frequency and amplitude. Next, we developed a machine learning (ML) pipeline to recognize these different buttons through the computer’s microphone, with several modifications on the buttons producing detectably different audio signatures with an average of 90% accuracy per button. Finally, we conducted a user study in which participants replicated a subset of our button designs and provided both quantitative feedback in terms of how well their button worked as well as qualitative feedback in the form of survey questions. By understanding which modifications to the cardboard buttons yield the most accurate detection results, we can create and use the fabricated buttons as peripheral computer widgets. These low-cost widgets are easy to create and customize for educational purposes, temporary computer setups, and people with different abilities who need regular adjustments to their digital setups.

Misinformation and disinformation online — and on social media in particular — have become a topic of widespread concern. Recently, Facebook and Social Science One released a large, unique, privacy-preserving dataset to researchers that contains data on user interactions with URLs shared on Facebook, including how users interact with posts and demographic data from those users. We analyzed this data through the lens of U.S. mis/disinformation and cross-referenced three lists of domains: Third-Party Fact Checked URLs marked within the dataset, 2016 Election Fake News sites from Grinberg et al. (2019), and Broad Misinformation domains from Zeng et al. (2020). We calculated average interaction by user demographic group for these URL subsets, finding distributions for views, clicks, and shares among other metrics (taking into account the differential privacy noise added to the data). Furthermore, we compared interaction metrics between the three mis/disinformation subsets and baseline U.S. URLs, finding that posts containing mis/disinformation URLs draw significant user engagement. We also find that older and more politically conservative U.S. users are more likely to be exposed to (and ultimately re-share) potential mis/disinformation, but that those users who are exposed are roughly equally likely to click. We discuss the implications of our findings for platform interventions and further study towards reducing the spread of mis/disinformation on social media.

With the rise of social media and video sharing platforms, many people are turning to sources like Youtube as their main source of information. However, these platforms have become easy targets for misinformation campaigns. Credibility is hard to ascertain on a video sharing platform like Youtube come from the wide base of content creators. On a typical Google search for a topic, the leading sources are usually mainstream-media companies. However, on Youtube, it is significantly easier for an individual person, or channel, to overtake mainstream media and become popular. There is also no standard way for creators to display credibility factors or cite their sources on video-sharing platforms. This leads to creators using a workaround to cite sources or simply foregoing citations all together. The implications of this for viewers is that they are unable to quickly identify the credibility of the source without cross-referencing other places on the web, or lateral reading. Currently, the only standardized information displayed about a channel is its display name, profile picture, subscriber count, and occasionally a verification indicator. The verification indicator, in particular, can be misleading as it only indicates whether a channel is who they claim to be. This project is researching and developing citations for Youtube videos. Video citations will allow creators to display credibility within individual videos and allow viewers to conduct lateral reading with ease. Citations specific to this format could also take on advantages of social media platforms, specifically user ratings and collaborative creation of citations. There are many possibilities in how citations can be introduced that we will explore through this project, each with pros and cons. For instance, sharing and creating credible videos could contribute to the pre-existing communities currently on Youtube through crowd-sourcing citations.

 Motor neuroprosthetics are brain-computer interfaces (BCIs) that record neural signals from an individual, use a decoding algorithm to predict the individual's intended movement, and perform that movement using a prosthetic device. As the neural code for movements is still being uncovered, current devices are still limited in their ability to create naturalistic movements. Development of motor decoding algorithms that can predict more naturalistic movement may enhance these prosthetics. In this project, we aim to implement and elaborate on newly published motor decoding algorithms and then compare alternative approaches to predicting the direction of movement in upper body joints (initially just the wrist) based on electrocorticography (ECoG) recordings from the surface of the human brain. Our dataset uses neural data collected during one week of continuous ECoG recordings from four subjects undergoing chronic, in-patient seizure monitoring, and includes hours of naturalistic movement, previously mapped from concurrent video to provide joint tracking of the head, shoulder, elbow and wrist. Carrying out this study on existing ECoG recordings provides a basis for extending our decoding work to future, real-time neural recordings. We have identified electrodes that correspond to hand-related areas in motor and sensory cortices for our initial focus on motor decoding of wrist movement. The goal of our project is to explore motor decoding of dimensionally-reduced spectral features using various neural network architectures. This will allow for more accurate responses in motor neuroprosthetics since they rely on motor decoding of naturalistic movements.

Internet censorship is accomplished through technological means, it is sometimes enforced via automated keyword filtering with a keyword list, implemented on a platform or state level. Individual users who discuss “sensitive” issues online may find their social media posts are made invisible or removed and may even have their accounts shut down. In this project, I experimented with using images as carrier of text data to send sensitive inforamtion over a censored network and made a publicly available data-hiding web tool for this project. A PRNG-based, spatial domain embedding method was used to embed the text data, the carrier image is then sent over a network where a keyword-based censor was detected. The result of the experiment shows that steganography can be used to combat keyword-based internet censorship, which promotes the free exchange of information. 

Biotechnology and DNA sequencing have become commonplace in their integration into a multitude of fields. Forensics has employed DNA sequencing in order to identify crime scene victims. Direct to consumer (DTC) DNA testing has allowed consumers to determine nuances regarding their ancestry and health. Medicine has employed sequencing and wet lab systems in order to further drug development and create targeted treatments. Previous literature has shown the potential for DNA sequencing and biotechnology systems to host unique cybersecurity vulnerabilities rooting in their functionalities. For example, researchers have shown the potential for misuse in the genetic genealogy database GedMatch. Often used by law enforcement, databases such as GedMatch allow individuals to perform relationship and family tracking, using DNA sequencing results to match potential relatives. Researchers have shown that this database and its underlying algorithms could be capitalized on by an adversary to spoof their own identity or create false relationships. In coming years, we expect genetic inference models such as ancestry analysis and DNA Phenotyping, the extrapolation of an individual’s physical features based on their genetic information, to become commonplace in the aforementioned fields amongst others. Given the potency for issues that researchers have found in other models, we look to advance the body of knowledge surrounding the validity and robustness of these genetic inference models being used. We perform an overview of a series of models that determine traits such as ancestry and physical features of individuals from DNA and explore how these models could be exploited or tricked into returning incorrect results. Finally, we discuss the implications of these vulnerabilities, both for the biotechnology space and cyber-physical systems as a whole.

Misinformation, media containing misleading or inaccurate information, is an increasingly prevalent and complex issue in society. There has been lots of previous work to classify misinformation, but none have contextualized it in terms of its harms to people, groups, and society. Misinformation can also have disparate harms and impacts on different groups: political disinformation campaigns often target underprivileged groups to attempt to disenfranchise them, while recent coronavirus misinformation has significantly affected marginalized groups. Misinformation may vary in the scope of their societal harm: some harassment may target specific individuals with misinformation, while others can cause a broader societal effect, such as through the loss of trust in public institutions. We propose to develop a taxonomy that classifies types of misinformation according to their potential for harm to aid efforts to address the effects of misinformation effectively. We also start with specific examples from two domains: elections and public health. We aim to interview fact-checkers early about what factors they consider when deciding to fact-check specific content, as they often must triage and select incoming media, and harness their intuitions in terms of potential negative impacts of misinformation. After this, we will develop a survey and survey a broad demographic of people to obtain initial results. From the survey data, we will develop a taxonomy of harms related to misinformation and iterate on the taxonomy with more people to get feedback. Our harm taxonomy can help fact-checkers triage incoming misinformation and prioritize which needs to be checked first. It also offers an explicit characterization of different types of intended harms, which may be useful when considering what kind of response is warranted. Finally, it lays a groundwork for improvement of machine learning systems that could better aid human review.

Usually, to compute the mean of a probability distribution, one takes samples from the distribution and computes the sample mean. However, if the samples are corrupted by an adversary, or if they contain outliers, then these estimates can differ significantly from the true mean. In 2019, an algorithm was developed to detect outliers in a sample, which allowed for robust mean estimation algorithms - that is, mean estimation in the presence of outliers or corruptions. In 2020, a robust covariance estimation algorithm was developed, using the 2019 mean estimation algorithms. The proofs in the 2020 paper suggest that we can create robust covariance estimation algorithms using robust mean estimation algorithms. So it’s natural to ask whether we can create a robust and differentially private covariance estimation algorithm (that is, the algorithm is robust and doesn’t leak information about users whose data appears in the dataset), as long as we have a robust mean estimation algorithm that is differentially private. Building on a 2021 work from researchers at the University of Washington that introduces the first efficient algorithms for private and robust mean estimators, we have created private and robust covariance estimators. Our approach is analogous to the 2020 robust (but not private) covariance estimation algorithm that builds on the 2019 robust (but not private) mean estimation algorithm, with the required modifications. This will be a crucial part of designing federated systems, where privacy is a main constraint and robustness is necessary as data comes from multiple sources, not all of which can be trusted.

Wireless power through magnetic resonance between coils of wire has enabled a new charging paradigm in a variety of domains, from robotics to biomedical implants. As wireless power systems move from simplistic to more perfomant architectures comprising of many coils, the design complexity scales very quickly. This is due to the difficulty in simulating and modeling the magnetic fields that form the backbone of the wireless power transfer, as in the multi-coil case the computational complexity quickly exceeds the capacity of even high end servers. To enable the development of next generation wireless power devices, we developed the Mostly Printed Field Characterization System (MPFCS), a robotic scanner that collects high-fidelity, high-resolution magnetic field data. However, while the system creates useful visualizations for wireless power, it does not provide a mathematical model that would allow for the precise optimization and rigorous understanding of the fields that engineers often need. Addressing that, we present physics-driven machine learning methods that combine electromagnetic theory with data collected from the MPFCS to build simplified mathematical models for these magnetic fields. We provide, for the first time, a characterization of fields for systems that were previously too complex to analyze effectively by hand or through computation. Preliminary evaluation of the data shows that there is very little error compared to simulated values. Based on the algorithm's performance on similar problems, this suggests promising final results. This work provides a deeper understanding and design tool to build and iterate on next generation devices, leading to both accelerated prototyping and novel research directions. 

Adolescents are vulnerable to high levels of stress in their lives that usually result from school, relationships, and family life. Approximately 27% of US teens report very high levels of daily stress, and 31% report feeling overwhelmed from negative stress. The data of fluctuating stress levels throughout the day can facilitate formulating effective stress measurement and reduction techniques for teens, which is imperative to support this vulnerable population. Using social robots to collect these data and offer mental-health counseling can be very cost-effective and scalable to intervene in critical mental health situations in adolescents worldwide. Today’s teens are the first generation to spend a lifetime living and increasingly experiencing human-computer interaction. According to far-seeing scholars, it is critical to innovate, design, and prototype technologies enhancing the connection between humans and robots targeting next generations. A wide array of research in human-robot interaction (HRI) focuses on specific age groups, where assistive technologies are mostly used to help the populations of elderly people and young children. However, very little research has been conducted to address teen-stress, or teen-robot interaction. My research in Project EMAR (Ecological Momentary Assessment Robot) aims to collect in-the-moment data from teens to map their stress and mood levels and offer counseling through a social robot EMAR focusing specifically on interventions for Dialectic Behavioral Therapy (DBT), and Acceptance and Commitment Therapy (ACT), two evidence-based therapies with efficacy treating teenager’s mental health. EMAR interacts with teens intimately through activities containing sounds, texts and images to collect data measuring moods and stress with EMA (Ecological Momentary Assessment) technique. EMA allows adolescents to report on sensations, feelings and behaviors close in time to actual experiences. Thus, EMA effectively minimizes recall bias, maximizes ecological validity, and allows study of micro processes that influence behavior in real-world contexts.