Brief: Introduce a learning resource the task of preconditioned visual language inference and rationalization (PVLIR).

Brief: Proposed a novel synthetic captioning method to incorporate traffic domain knowledge into video-language models, to improve Traffic question/answering performance of selective video-language models by 20%

Brief: Propose an improved model for reasoning with preconditions through minimum supervision.

Brief: Propose suite of tasks to evaluate preconditioned inference of commonsense knowledge in SOTA models

Brief: Introduced and implemented AutoUnload and AutoUnload Advisor for Oracle Heatwave

Brief: Build an automatic pipeline to calibrate and probe visible aspects of commonsense in visual language models.

Brief: Human-Assisted Neuro-Symbolic Architecture for Multi-modal Intelligent Traffic Monitoring

Brief: In this work, we investigate and consolidate various resources of commonsense knowledge into a first integrated commonsense knowledge graph.

Brief: Evaluate reasoning with preconditions in SOTA models by proposing a multiple-choice question answering dataset

Brief: End-to-End Automated Machine Learning (E2E-AutoML) seeks to automate the machine learning pipeline generation process. In this work we present DSBox, an expandable, multi-modal E2E-AutoML system, that can go directly from raw data to ML pipeline with minimal human supervision. We show DSBox is able to outperform human experts and beat state-of-the-art AutoML systems (ML-Plan, Auto-sklearn, TPOT) through evaluation on 468 diverse datasets with multiple media and task types.

Brief: Aim of this research is to understand tabular data based on actual contents and the meta-data of table (e.g. styling, headers) and map the content to entities and relations from knowledge graphs such as dbpedia and wikidata.

Brief: Here, we investigated spatio-temporal data analysis methods in motor-based early ASD (autism spectrum disorder) detection application. The challenge here is processing the stream multi-modal (Motion, fMRI, EEG, and DTI) information with the goal of detecting early signs of ASD in the human subject. We propose two solutions, with the ability to achieve 80% accuracy. Additionally, we have extracted a set of entropy-based features that can help explain the decision-making process of the network and improve confidence in the results.

Brief: In this work, we build upon previous work on using Left-Right consistency for depth estimation. The conventional approach in such methods is using a series of hand-designed loss metrics to measure the consistency between the left-eye and right-image, which limits the capabilities of such models. In this work, we borrow from Generative Adversarial Networks (GAN) and propose a method for LR-Consistency that learns the consistency/loss function with goal of discriminating the inconsistent Images.

Brief: In this work, we studied deep learning applications in elucidating the relationships among painting. The painting dataset consists of three periods of Pre-Baroque, Baroque, and post-baroque.

Brief: Investigate deep learning applications as the computational intelligence medium for classification of music pieces according to style and historical period, such as Baroque, Classical and Romantic periods.

Brief: Big data local-to-global methods in analysis and prediction of dynamics in atmospheric chemistry spatiotemporal data.

Brief: This research focuses on our proposed hardware architecture on a highly scalable, shared-memory, MonteCarlo Tree Search (MCTS) based Blokus-Duo solver. Our design is inspired from parallel MCTS algorithms and is potentially capable of obtaining maximum possible parallelism from MCTSalgorithm. We also combine MCTS with set of pruning heuristics to increase both the memory and Logic Element (LE) utilizations.

Brief: This research focuses on our proposed hardware architecture on a highly scalable, shared-memory, MonteCarlo Tree Search (MCTS) based Blokus-Duo solver. Our design is inspired from parallel MCTS algorithms and is potentially capable of obtaining maximum possible parallelism from MCTSalgorithm. We also combine MCTS with set of pruning heuristics to increase both the memory and Logic Element (LE) utilization.

Brief: In this work, we address the problem of scheduling in high-level synthesis application. We propose a new scheduling algorithm based on Monte-Carlo (MC) simulation that uses MC Tree Search (MCTS)method as the heart of our scheme handles the schedulingto produce optimized scheduling with both time and resource constraints.

Brief: In this work, we investigated SRAM memory architectures to work in subthreshold voltage range for low-power applications. We have proposed a set of cell architectures that improve read, write, and hold stability compared to state-of-the-art cell architectures. Finally, we developed our in-house memory compiler tool that assembles such memory cells into full SRAM memory with arbitrary size.