Human driving performance is a complex function of driver’s cognition such as situational awareness, and fatigue. We’ll present a shared control technique in which the autonomy captures the driver alertness through facial expressions and eye tracking and aids the human driver by ensuring safety and improved driving performance. To accomplish this, first we simulated multiple driving scenarios using MATLAB’s Driving Scenario Designer to replicate an automated vehicle. Our plan is to manually control the vehicle and autonomy infers the driver intent using clothoids, plausible programmed vehicle trajectories. Once the human intent is determined, the autonomy navigates to the appropriate clothoid. Specifically, the autonomy models plausible vehicle trajectories by generating clothoids, programmatically, and learning the human intention such as lane change by mapping the human input to an appropriate clothoid. After determining the human intent, the autonomy safely navigates the target clothoid by appropriately modifying the vehicle control input based on the driver alertness.

New materials are discovered now not only by physical lab experiments, but computational design can also contribute to discoveries. Machine learning has been one significant tool helping scientists and researchers to conduct experiments with new materials. By using various computing methods, material data can be recorded and computed. With new technologies, many models can be provided for scientists to efficiently screen materials with specific properties, reducing experimental time and cost. One challenge in increasing the impact of publishing machine learning models is in maintaining and using others codes. In this article, we will demonstrate how the cloud platform Foundry can be used to increase the useability of ML models and the ability of researchers to efficiently use other’s trained models. To do this we will focus on one material's property and dataset. Short information / background on solubility. We will retrain and assess this previous model/dataset using errobar analysis to better understand performance. By utilizing the Foundry framework we will demonstrate successful applications in material science through machine learning. Finally, we will evaluate the opportunities and challenges in material science brought by machine learning and big data.

An effective peptide-based cancer vaccine requires intracellular delivery of antigen peptides to activate tumor-killing immune responses. The Pun Lab developed the Virus-Inspired Polymers for Endosomal Release (VIPER) that facilitate intracellular delivery of antigen peptides and successfully applied VIPER in peptide cancer vaccines. VIPER can self-assemble into nanoparticles to enhance cellular uptake and intracellularly deliver drug by lysing the endosomes via conjugated melittin (a membrane-penetrating peptide). Current iterations of VIPER employ pyridyl disulfide (PDS) to conjugate melittin to the polymer backbone via a cleavable disulfide bond. However, previous work suggested that polymer-conjugate melittin is more potent at membrane lysis than free peptide. Here, I propose a more stable conjugation strategy: a VIPER variant that utilizes a pentafluorobenzyl moiety (VIPER-PFB) instead of PDS. We hypothesize that this modification can enhance endosomal trafficking, reduce toxicity, and improve efficacy of VIPER as a cancer peptide delivery platform. I synthesized VIPER-PFB and evaluated its endosome-lysing capability via an in vitro red blood cell lysis assay. The peptide vaccine delivery in vivo was assessed using a model antigen. Surprisingly, in vitro, VIPER-PFB demonstrated less endosomal lysis capabilities compared to VIPER. However, in vivo, VIPER-PFB significantly elevated cytotoxic T-cell response against our model antigen. VIPER-PFB also notably improved helper T-cell responses compared to VIPER, which can also enhance anticancer responses. While in vitro result is contrary to the mechanistic hypothesis, the in vivo result shows that VIPER-PFB is a promising cancer peptide delivery platform. These findings motivate deeper studies into delivery mechanism of VIPER-PFB.

Language learning transforms neural correlates within the brain. Electroencephalographic (EEG) alpha, beta, and gamma frequency band event-related desynchronization (ERD) and coherence reflect these changes. As one becomes more proficient in a skill—such as language interpretation and translation, alpha ERD increases, and its coherence becomes less widespread. The objective of this study is twofold: (1) to observe the relationship between sequentially learned Spanish language proficiency (Novice, Intermediate, and Advanced) and both ERD and coherence of frequency bands in the EEG spectrum, and (2) to implement machine learning techniques to predict Spanish language proficiency based on ERD and coherence features.

Due to the inherently biased data that machine learning relies on, the algorithm is likely to be partial. Technology laced with bias has been seen to skew hiring decisions in school and job applications, causing many to question the integrity of automated software. If left unchecked, this can not only perpetuate societal inequities but exacerbate them beyond scale. Previous work sought to certify robustness, ensuring the algorithm’s validity. However, these techniques are limited in predicting where the bias is occurring. This research aims to identify when a machine learning algorithm is not robust to small changes in training data. Empirical robustness will be assessed through label flipping in a supervised environment. Results will be given a consistency score and used to identify trends in program behavior. The project seeks to discover better methods of modifying data and the attributes prone to machine learning bias.