The Large Hadron Collider at CERN is built to accelerate particles to high speeds and make them collide. The curved trajectory of these particles is recorded by detecting electrical charges deposited on multiple layers of tracking equipment as a particle travels through the detector. The resulting patterns are then used to reconstruct the path traveled by the particle. The more the collisions, the higher the number of charge depositions to detect. This exponential increase in data is predicted to be the main issue with the next run of the LHC experiment, where we are set to test higher energy interactions. As traditional tracking algorithms do not scale well with this increase in data, supplementing them with machine learning provides a promising solution. Scaling high-energy particle tracking in the LHC to process petabytes of data is the focus of the Exa.TrkX project, which our study is part of. In our research, we study the robustness of Exa.Trkx models and algorithms against noise and misalignment. Robustness is judged by analyzing performance metrics like purity and efficiency of pairs of charge deposits or “doublets''. Purity is defined as the ratio of true-positives over positives, and efficiency is defined as the ratio of true positives over the number of true values. A true deposit belongs to the same trajectory as the one we are comparing it with. In this presentation, I will discuss how we proved robustness against noise by observing that the change in doublet purity and efficiency was a trivial decrease of 0.3 and 0.2 percent respectively. Our research makes sure that the Exa.TrkX models can be applied to actual LHC data. We do this by proving that the models are not affected by real-life impurities in the data like noise.

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.

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.

Data-driven predictions of the future often suffer from error propagation and overconfidence. In many scenarios where forecasting is practical, the data follows a quasi-periodic pattern, which means that data from one point in time relays useful information about the data one period later. Through a Koopman theoretic approach, we make use of this feature to make long-term probabilistic forecasts that do not suffer from error propagation. Overwhelmingly, data-driven forecasts attempt to predict the exact value of a quantity into the future; however, such point-forecasts fail to describe the uncertainty in that quantity. Even when models are designed to make probabilistic forecasts, they are often overconfident and rely on the model to supply all sources of uncertainty. By forecasting the parameters of a probability distribution describing a quantity, rather than the quantity itself, we are also able to overcome the problem of overconfidence. Furthermore, our model relays novel and useful information about temporal patterns in the uncertainty of a dataset. We apply our model to electric load data from the 2017 Global Energy Forecasting Competition and show significant improvements compared to competing forecasts. I implemented the approach and ran experiments to explore the ways in which Koopman theory can be leveraged to robustly mitigate error and overconfidence in electric load forecasting. Our contributions to probabilistic forecasting of energy demand have the potential to lessen global warming, and the approach can be used in forecasting problems more broadly.