Siddhesh Bangar

- Spearheaded the creation of RanLayNet, a synthetic document dataset with automated labeling, aimed at enhancing layout detection model adaptability to diverse document formats.
- Conducted empirical experiments demonstrating that a deep layout identification model trained on RanLayNet outperformed models trained solely on real documents, showcasing improved adaptability and robustness.
- Developed domain adaptation approaches to address challenges arising from differences between source… Show more

- Spearheaded the creation of RanLayNet, a synthetic document dataset with automated labeling, aimed at enhancing layout detection model adaptability to diverse document formats.
- Conducted empirical experiments demonstrating that a deep layout identification model trained on RanLayNet outperformed models trained solely on real documents, showcasing improved adaptability and robustness.
- Developed domain adaptation approaches to address challenges arising from differences between source and target domains, achieving optimal model performance, as evidenced by a 0.398 and 0.588 mAP95 score in the scientific document domain for the TABLE class. Show less

• Developed Dynamic Context Learner (DCL) model GEC-DCL for optimal context breakpoints, outperforming fixed-length models.
• Achieved significant F0.5 score improvements of 77% increase, 19.61% boost, and 10.49% rise, respectively, on CoNLL2014, BEA-Dev, and FCE-Test datasets.
• Validated DCL's performance in scientific writing and scholarly publications with cutting-edge results on S2ORC dataset.

• Introduced an innovative end-to-end pipeline TC-OCR that integrates state-of-the-art deep learning models, which introduced an application to capture and transfer any table structure and content from one platform to another.
• Achieving a 25% improvement in OCR accuracy, with an impressive IOU of 0.96 for Table class.

Our Contributions
- Introduces a novel schema for organizing previous work on controllable text generation.
- Identifies seven crucial components in the text creation process.
- Focuses on controlled generation for scientific literature.
- Describes modulation strategies for each of the seven components.
- Presents a theoretical analysis and qualitative evaluation of these modulation methods.
- Provides insights into potential new architectures by combining these… Show more

Our Contributions
- Introduces a novel schema for organizing previous work on controllable text generation.
- Identifies seven crucial components in the text creation process.
- Focuses on controlled generation for scientific literature.
- Describes modulation strategies for each of the seven components.
- Presents a theoretical analysis and qualitative evaluation of these modulation methods.
- Provides insights into potential new architectures by combining these components.
- Anticipates future research that will empirically compare the strengths and utility of these methods. Show less

Our contributions:
- Propose a novel black-box denoised smoothing (DS) defense mechanism, ZO-RUDS, by prepending our RDUNet to the black-box model, ensuring black-box defense.
- Also, propose ZO-AE-RUDS in which RDUNet followed by autoencoder (AE) is prepended to the black-box model.