WiDS Cambridge
Poster Session 2024

​

Presenter: Rabab Alomairy

Affiliation: MIT

Title: Accelerating Force Calculations through Dimension Reduction in Neural Network Interatomic Potentials using Julia

Authors: Rabab Alomairy, Emmanuel Lujan, Spencer Wyant, Alan Edelman

Abstract: Quantum mechanics computations, particularly through Density Functional Theory, play a pivotal role in achieving accurate atomic force calculations. However, atomistic simulations face a significant challenge in accelerating force calculations, as these processes can extend over weeks or months, with ab initio calculations consuming a substantial portion of the overall time. To address the time-consuming nature of such calculations, surrogate machine learning models have emerged as rapid approximations, offering linear scalability with the number of atoms and promising accuracy across diverse atomistic systems. This study focuses on accelerating energy and force calculations, especially for larger systems and extended time frames. The research introduces dimension reduction in linear regression and neural network interatomic potentials, utilizing the capabilities of atomistic Julia packages like PotentialLearning.jl and InteratomicPotentials.jl. In addition, we quantify the impact in accuracy, speed, and stability of energy and force calculation on local/global descriptor of Atomic Cluster Expansion for Hf and HfO2 systems. The aim is to investigate phenomena, like the degradation of complex materials under extreme conditions, that are not directly observable experimentally. The work also explores the impact on GPU performance and accuracy across various neural network versions of state-of-the-art interatomic potentials, fitting them to a range of materials systems.

 

Presenter: Debankita Basu

Affiliation: Northeastern University

Title:  Unveiling Digital Truths in Deepfake Detection with CNN and Transformer Models

Author: Debankita Basu

Abstract: In the evolving landscape of digital media, the advent of Deep Fakes—sophisticated artificial intelligence-generated images and videos designed to replicate reality with alarming accuracy—has emerged as a formidable challenge to content authenticity. These manipulations, capable of altering public perception and infringing on personal security, underscore the pressing need for advanced detection methods. This research project embarks on a comparative analysis between two prominent machine learning architectures: Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs). The study meticulously evaluates the performance of these models under uniform experimental conditions to identify the more effective approach in discerning genuine from manipulated images. By leveraging a comprehensive dataset of real and AI-generated visuals, the aim is to understand the strengths and limitations of each model in the context of DeepFake detection. This exploration is crucial for developing robust, reliable tools capable of safeguarding digital content integrity against the sophisticated techniques employed in DeepFake creations. Through this comparative analysis, the research highlights the potential of machine learning in enhancing digital security and trust, paving the way for future research endeavors in the field.

 

Presenter: Julia Briden

Affiliation: MIT

Title: Constraint-Informed Learning for Warm Starting Trajectory Optimization

Authors: Julia Briden, Changrak Choi, Kyongsik Yun, Richard Linares, and Abhishek Cauligi

Abstract: Future spacecraft and surface robotic missions require increasingly capable autonomy stacks for exploring challenging and unstructured domains, and trajectory optimization will be a cornerstone of such autonomy stacks. However, the nonlinear optimization solvers required remain too slow for use on relatively resource-constrained flight-grade computers. In this work, we turn towards amortized optimization, a learning-based technique for accelerating optimization runtimes, and present TOAST: Trajectory Optimization with Merit Function Warm Starts. Offline, using data collected from a simulation, we train a neural network to learn a mapping to the full primal and dual solutions given the problem parameters. Crucially, we build upon recent results from decision-focused learning and present a set of decision-focused loss functions using the notion of merit functions for optimization problems. We show that training networks with such constraint-informed losses can better encode the structure of the trajectory optimization problem and jointly learn to reconstruct the primal-dual solution while yielding improved constraint satisfaction. Through numerical experiments on a Lunar rover problem, we demonstrate that TOAST outperforms benchmark approaches in terms of both computation times and network prediction constraint satisfaction.

 

Presenter: Sarah Caudill

Affiliation: University of Massachusetts Dartmouth

Title: Exploring the Dark Side of the Universe with Deep Learning

Author: Sarah Caudill

Abstract: The advanced detectors of the Laser Interferometer Gravitational-wave Observatory (LIGO) have been extremely successful in detecting binary black holes and neutron stars over the past nine years. These detections have revolutionized the field of astronomy, ushering in the era of multi-messenger astronomy with gravitational waves and enabling regular observations from the “dark side" of the universe. With these first detections, we are are probing new astrophysics and performing unprecedented tests of strong-field gravity. Nevertheless, the time has come to once again push beyond the frontier boundaries of the field. I demonstrate how deep-learning is paving the path forward for us. By employing a variety of deep-neural networks, we can search for more complex astrophysical signals, directly integrate data quality information, and perform previously prohibitive tasks orders of magnitude faster. I will describe these enhancements, discuss the challenges we faced, and present our plan for the next generation of gravitational-wave data analysis.

 

Presenter: Elisabetta Cornacchia

Affiliation: MIT

Title: Provable Advantage of Curriculum Learning on Parity Targets with Mixed Inputs

Authors: Emmanuel Abbe, Elisabetta Cornacchia, Aryo Lotfi

Abstract: Experimental results have shown that curriculum learning, i.e., presenting simpler examples before more complex ones, can improve the efficiency of learning. Some recent theoretical results also showed that changing the sampling distribution can help neural networks learn parities, with formal results only for large learning rates and one-step arguments. Here we show a separation result in the number of training steps with standard (bounded) learning rates on a common sample distribution: if the data distribution is a mixture of sparse and dense inputs, there exists a regime in which a 2-layer ReLU neural network trained by a curriculum noisy-GD (or SGD) algorithm that uses sparse examples first, can learn parities of sufficiently large degree, while any fully connected neural network of possibly larger width or depth trained by noisy-GD on the unordered samples cannot learn without additional steps. We also provide experimental results supporting the qualitative separation beyond the specific regime of the theoretical results. Join work with E. Abbe and A. Lotfi.

 

Presenter: Yifu Ding

Affiliation: MIT

Title: Optimizing data-driven coal retrofitting solutions with geospatial synthetic power system datasets for India

Authors: Yifu Ding, Jansen Wong, Serena Patel, Guiyan Zang, Rob Stoner

Abstract: India has committed to achieving net-zero emissions by 2070 and installing 500 GW renewable energy capacity by 2030. Nevertheless, coal electricity still contributed over 70% of the total power generation as of 2021. These coal power plants face the risk of stranded assets. Early coal retirement and renewable capacity expansion could lead to the lock-in effect due to the lack of flexibility and low coal prices. This research will optimize the retrofitting solutions for 284 operational coal power plants in the Indian power system, leveraging the geospatial synthetic power system dataset. The machine learning clustering approaches will apply to predict and cluster coal power plants' operational characteristics.

 

Presenter: Clarissa Lauditi

Affiliation: Harvard University

Title: The star-shaped space of solutions of the spherical negative perceptron

Authors: Brandon Livio Annesi, Clarissa Lauditi, Carlo Lucibello, Enrico Malatesta, Gabriele Perugini, Fabrizio Pittorino, Luca Saglietti

Abstract: Empirical studies on the landscape of neural networks have shown that low-energy configurations are often found in complex connected structures, where zero-energy paths between pairs of distant solutions can be constructed. Here, we consider the spherical negative perceptron, a prototypical nonconvex neural network model framed as a continuous constraint satisfaction problem. We introduce a general analytical method for computing energy barriers in the simplex with vertex configurations sampled from the equilibrium. We find that in the overparametrized regime the solution manifold displays simple connectivity properties. There exists a large geodesically convex component that is attractive for a wide range of optimization dynamics. Inside this region we identify a subset of atypical high-margin solutions that are geodesically connected with most other solutions, giving rise to a star-shaped geometry. We analytically characterize the organization of the connected space of solutions and show numerical evidence of a transition, at larger constraint densities, where the aforementioned simple geodesic connectivity breaks down. Finally, we explain what kind of bias this could induce in the training dynamics of local optimization dynamics.

 

Presenter: Sijia ‘Nancy’ Li

Affiliation: Harvard University

Title: Understanding Defensive Strategies for Adversarial Attacks on Large Vision Language Models

Authors: Danning Lai, Sijia Li, Tiantong Li, Yong Zhang, Hanlin Zhu

Abstract: Recent advancements in vision language models (VLMs) have enabled the integration of multiple modalities to solve more complex problems, such as image captioning and visual question answering. Despite the great abilities in these multimodal tasks, VLMs are still susceptible to adversarial attacks, raising profound concerns about their reliability and safety. One study introduced the Behavior Matching method, exposing vulnerabilities to string attack, leak context attack, and jailbreak attack (Bailey et al., 2023). To ensure the model's robustness, we need defensive strategies to protect the model from producing illegal or incorrect results. In this project, we aim to investigate the defensive strategies in the vision domain for VLMs and visualize their effectiveness from image embedding space. Our experiments show that JPEG compression and cropping are the most effective defense strategies against VLM adversarial attacks. Visualizing images in the embedding space, we also demonstrate that clean, adversarial, and defense images lie in distinct surfaces in the embedding space. After applying defenses to adversarial images, the embeddings of defense images shifted back towards clean images. Our findings advance the understanding and enhancement of VLM robustness.

 

Presenter: Shrestha Mohanty

Affiliation: MIT

Title: Asking Clarifying Questions for Effective Collaboration in Grounded Instruction-Based Agent Interactions

Authors: Shrestha Mohanty, Negar Arabzadeh, Julia Kiseleva, Artem Zholus, Milagro Teruel, Ahmed Awadallah, Yuxuan Sun, Kavya Srinet, Arthur Szlam

Abstract: Motivated by the adaptability of human intelligence across various tasks and multi-modal environments, the research community is actively engaged in developing interactive agents capable of engaging in natural conversations with humans and assisting them in real-world tasks. These agents need the ability to request feedback in the form of situated clarifying questions when communication breaks down or instructions are unclear. This paper delves into an extensive investigation of the production of clarifying questions within the context of human-centered AI instruction-based interaction, using a Minecraft environment as a grounding framework. The unique challenges presented by this scenario include the agent's requirement to navigate and complete tasks in a complex, virtual environment, relying on natural language instructions and action states. Specifically, we make the following contributions: i) A crowd-sourcing tool for collecting grounded language instructions along with clarifying questions in times when instructions are not clear at scale with low costs; ii) A substantial dataset of grounded language instructions accompanied by clarifying questions; and iii) Several baselines for requesting feedback in case of unclear instructions. These contributions are suitable as a foundation for further research.

 

Presenter: Swati Padmanabhan

Affiliation: MIT

Title: Computing Approximate Lp Sensitivities

Authors: Swati Padmanabhan, David P. Woodruff, Qiuyi R. Zhang

Abstract: Recent works in dimensionality reduction for regression tasks have introduced the notion of sensitivity, an estimate of the importance of a specific datapoint in a dataset, offering provable guarantees on the quality of the approximation after removing low-sensitivity datapoints via subsampling. However, fast algorithms for approximating ℓp sensitivities, which we show is equivalent to approximate ℓp regression, are known for only the ℓ2 setting, in which they are termed leverage scores. In this work, we provide efficient algorithms for approximating ℓp sensitivities and related summary statistics of a given matrix.

 

Presenter: Ágata Piffer Braga

Affiliation: University of Massachusetts Dartmouth

Title: Multi Instrumental Data Combination for Coastal Fluid Dynamics

Author: Ágata Piffer Braga

Abstract: As a physical oceanographer deeply engaged in data science, I rely heavily on data analysis in my daily work. In my poster presentation, I will share a segment of my PhD. research methodology, which involved integrating a diverse oceanic dataset. This included combining data from Autonomous Underwater Vehicles (AUVs) measuring salinity, velocity, temperature, and turbulence, along with georeferenced drone footage and data collected from boats. My research focuses on understanding the complex fluid dynamics of river plumes, where freshwater meets the salty ocean, carrying sediment and other materials to the coast. This understanding is crucial for effective coastal management and maintaining water quality. In sharing this work, I aim to underscore the intrinsic role of data science in physical oceanography, facilitated by intensive Python coding and scientific computing. Additionally, I seek to shed light on my unique learning journey, which has diverged from the experiences of many male colleagues in the physical sciences.

 

Presenter: Jessica Quaye

Affiliation: Harvard University

Title: Improving safety of T2I generative AI models through crowdsourcing

Author: Jessica Quaye

Abstract: With the rise of text-to-image (T2I) generative AI models reaching wide audiences, it is critical to evaluate model robustness against non-obvious attacks to mitigate the generation of offensive images. By focusing on “implicitly adversarial prompts” (those that trigger safety violations for non-obvious reasons), we isolate a set of safety issues that are usually hidden in blindspots of T2I models. We present to you the Adversarial Nibbler Challenge, a red-teaming methodology for crowdsourcing a diverse set of implicitly adversarial prompts for T2I models. We provide a systematic study of novel attack strategies, discuss safety failures revealed by challenge participants, and release a companion visualization tool for easy exploration. We are confident that this work will enable proactive, iterative safety assessments and promote responsible development of T2I models.

 

Presenter: Shreyaa Raghavan

Affiliation: MIT

Title: Identifying Stop-and-Go Congestion with Data-Driven Traffic Reconstruction

Authors: Shreyaa Raghavan, Edgar Ramirez Sanchez, Cathy Wu

Abstract: Identifying stop-and-go events (SAGs) in traffic flow presents an important avenue for advancing data-driven research for climate change mitigation and sustainability, owing to their substantial impact on carbon emissions, travel time, fuel consumption, and roadway safety. SAGs occur when drivers repeatedly accelerate and decelerate in highway traffic jams and are estimated to account for a significant portion (33-50%) of the adverse effects of highway driving. However, insufficient attention has been paid to precisely quantifying when, where, and how often these SAGs take place––necessary for downstream decision-making, such as intervention design and policy analysis. A key challenge is that the traffic data available to researchers and governments are typically sparse and aggregated to a granularity that obscures SAGs. To overcome such limitations, this study thus explores the use of traffic reconstruction techniques to generate rich traffic data. Then, we introduce a kernel-based method for identifying spatiotemporal features, such as SAGs, in traffic and leverage bootstrapping to quantify the uncertainty of the reconstruction process. Experimental results on California highway data demonstrate the promise of the method for capturing stop-and-go events. This work contributes to a foundation for data-driven decision-making to advance the sustainability of traffic and transportation systems.

 

Presenter: Maria Sol Rosito

Affiliation: Institute for Astronomy and Space Physics and University of Buenos Aires

Title: A data-driven approach to galaxy classification using unsupervised learning

Authors: M. S. Rosito, L. A. Bignone, P. B. Tissera, S. E. Pedrosa

Abstract: Harnessing the power of statistics and artificial intelligence, this study introduces a method developed for classifying galactic morphologies. With the unprecedented growth of astronomical data sets, the adoption of automatic and efficient classification tools, particularly those based on unsupervised approaches, became imperative. High-dimensional kinematic images of galaxies obtained from the EAGLE cosmological simulation serve as inputs for the method, which employs a combination of dimensionality reduction (UMAP) and clustering (HDBSCAN) algorithms. Compressing input information, which involves velocity, dispersion, and flux, into bidimensional points facilitates the definition of data-driven, well-defined groups of galaxies that capture inherent variations in their fundamental properties. Comprehensive statistical analysis validates the meaningfulness of this novel classification from physical and observational perspectives. Our experiments yield relevant astronomical results, focused on the segregation of galaxies with different amounts of rotation and/or rotation orientation, the refinement of clusters based on 3D-shapes, the method's robustness when simulating different observational scenarios, and the importance of utilizing realistic training data sets. This work exemplifies the crucial role of data science in enhancing our understanding of nature and encourages its ongoing integration into astronomical research methodologies.

 

Presenter: Michelle Vaccaro

Affiliation: MIT

Title: Evaluating Human-AI Collaboration

Authors: Michelle Vaccaro, Abdullah Almaatouq, Thomas Malone

Abstract: People increasingly work with artificial intelligence (AI) tools to accomplish goals in fields including medicine, finance, and law, as well as in daily activities such as traveling, shopping, and communicating. Despite a wide body of experimental studies involving humans and AI-systems working with each other, we still lack a conceptual understanding of factors that impact the performance of human-AI systems, which hinders our ability to create them in ways that enable the best results. To address this gap, we conducted a systematic literature review and meta-analysis of studies involving human-AI collaboration in which we compare the performance of the human-AI system to a baseline of the human alone or AI alone, whichever of the two performs better. We found that, on average, the human-AI combination performed significantly worse than this baseline with an overall effect size of Hedge’s g = -0.2679, (95% confidence interval (CI) [-0.4475, -0.0884]). We also identified several factors that significantly impacted the effectiveness of human-AI collaboration: the type of task, AI, and division of labor involved in the experiment. Our work thus sheds needed light on potential designs of future experiments and AI tools that unlock greater human-AI synergy.

 

Presenter: Erin Walk

Affiliation: MIT

Title: Partisan Segregation in Online Content Viewing

Author: Erin Walk

Abstract: As polarization in the U.S. becomes more extreme, and individuals spend more time engaging with news and other content online, researchers have become interested in both the extent of content segregation across parties as well as where individuals get their political information. This paper explores how and when individuals consume news online, whether partisan news behaviors are consistent across platforms, and whether certain individuals are consuming increasingly extreme content over time. In addition, I consider segregation of content viewership across platforms extending this work beyond news viewership. I use a panel of data from provider MFour to analyze YouTube viewing behaviors using text classification of transcripts to label videos as political or non-political as well as by political leaning. I impute panelist partisanship based on affiliations from the Cooperative Congressional Election Survey (CCES), which includes similar demographic characteristics. In addition, I combine the panel information with data from Spotify and web browsing as well as additional information collected using the YouTube API. Preliminary results indicate that several areas outside of political news are highly polarized, including gaming channels and other entertainment.

 

Presenter: Cynthia Zeng

Affiliation: MIT

Title: Multimodal Machine Learning and Climate Change Adaptation

Authors: Cynthia Zeng

Abstract: Climate change is escalating the frequency and severity of natural disasters worldwide, necessitating urgent societal adaptation. In this talk, I present a multimodal machine learning (ML) framework designed to predict natural disasters. Traditionally, weather forecasting has depended on dynamical equations for over a century. However, recent advancements in artificial intelligence are revolutionizing this domain. The innovative multimodal ML framework leverages processing techniques from computer vision, natural language processing, time series signal processing techniques to integrate various data types, such as satellite imagery, textual information, and tabular data, to generate both short-term and long-term forecasts. Our first case study demonstrates that, for 24-hour hurricane forecasting, our ML models achieve results that are competitive with those produced by established national weather forecasting agencies. In our second case study, we explore the potential to create global models with a multi-year scope for assessing flood risks. Artificial intelligence will fundamentally change the way our interaction with weather, and these ML-driven risk assessments will have profound impacts on urban planning, infrastructure investment, renewable energy planning, and insurance policy.

 

Presenter: Jiaqi Zhang

Affiliation: MIT

Title: Membership Testing in Markov Equivalence Classes via Independence Queries

Authors: Jiaqi Zhang, Kiran Shiragur, Caroline Uhler

Abstract: Understanding causal relationships between variables is a fundamental problem with broad impacts in numerous scientific fields. While extensive research have been dedicated to _learning_ causal graphs from data, its complementary concept of _testing_ causal relationships has remained largely unexplored. In our work, we take the initiative to formally delve into the _testing_ aspect of causal discovery. While _learning_ involves the task of recovering the Markov equivalence class (MEC) of the underlying causal graph from observational data, our _testing_ counterpart addresses a critical question: _Given a specific MEC, can we determine if the underlying causal graph belongs to it with observational data?_ We explore constraint-based testing methods by establishing bounds on the required number of conditional independence tests. Our bounds are in terms of the size of the maximum clique ($s'$) and the size of the maximum undirected clique ($s$) of the given MEC. In the worst case, we show a lower bound of $\exp(\Omega(s))$ independence tests. We then give an algorithm that resolves the task with $\exp(O(s'))$ independence tests. Notably, our lower and upper bounds coincide when considering moral MECs ($s'=s$). Compared to _learning_, where algorithms often use a number of independence tests that is exponential in the maximum in-degree, our results show that _testing_ is a relatively more manageable task. In particular, it requires exponentially less independence tests in graphs featuring high in-degrees and small clique sizes.

 

Presenter: Xiaohan Zhao

Affiliation: Harvard University

Title: Quantifying Differences in LLM and Human Feedback on Mental Health Related Surveys

Authors: Phoebe Cheng, Yixian Gan, Haitian Liu, Catherine Ni, Xiaohan Zhao

Abstract: Our research pioneers the quantification of difference between LLM (large language model) and human-generated feedback on mental health surveys, addressing a gap in existing literature. Given the absence of existing dataset, we curated the first benchmark dataset for the detection of LLM-generated text in mental health domain. This involved simulating human subjects using LLM and employing text augmentation techniques. On the detection front, we carefully examined the linguistic differences between LLM and human-generated text, uncovering valuable insights. We implemented models from different categories, i.e. tree-based models, transformer-based classifiers and in-content few-shot learning models, achieving impressive performance in detecting LLM responses. By addressing both the generation and detection aspects of fake responses, our work provides a holistic framework for enhancing the authenticity and reliability of survey data in the context of mental health research.