doi.bio/kahini_wadhawan

Kahini Wadhawan

Early Life and Education

Kahini Wadhawan is a researcher with a background in AI and machine learning. She studied at the University of Colorado Boulder.

Career and Research

Wadhawan is currently affiliated with IBM Research India, where she works as a Staff Research Scientist. Her research interests include large language models, causality, representation learning, biomedicine, proteins, AMP, small molecule, and drug discovery.

She has published extensively in the field of machine learning and AI, with notable works including:

• "Causally Fair Language Models Through Token-level Attribute Controlled Generation" (2023), which proposes a method to control the attributes of Language Models (LMs) for text generation tasks, achieving state-of-the-art performance for toxic degeneration.
• "Causal Graphs Underlying Generative Models: Path to Learning with Limited Data" (2022), which provides a simple algorithm to uncover causal graphs implied by generative models, aiding in interpreting latent representations.
• "Optimizing Molecules using Efficient Queries from Property Evaluations" (2022), where she proposes QMO, a query-based molecule optimization framework that exploits latent embeddings, outperforming existing methods in benchmark tasks.
• "Towards Interpreting Zoonotic Potential of Betacoronavirus Sequences with Attention" (2021), applying an attention-enhanced LSTM deep neural net classifier to predict zoonotic potential with 94% accuracy.
• "Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations" (2021), introducing CLaSS, a method for attribute-controlled molecule generation, which led to the discovery of novel antimicrobial peptides.
• "Text Style Transfer Using Partly-Shared Decoder" (2019), focusing on sentiment transfer and introducing a new decoder architecture to balance content preservation and style modification.
• "Co-regularized Alignment for Unsupervised Domain Adaptation" (2018), addressing the challenge of deep neural networks generalizing across different domains by aligning source and target distributions in diverse feature spaces.
• "PepCVAE: Semi-Supervised Targeted Design of Antimicrobial Peptide Sequences" (2018), a framework for designing novel antimicrobial peptide sequences by learning a rich latent space of biological peptide contexts.

Wadhawan has also been granted a patent for an "Unsupervised text style transfer method, system, and computer program product" in 2020.

Kahini Wadhawan

Early Life and Education

Kahini Wadhawan is a researcher who holds a Master's degree from the University of Colorado Boulder. Her Master's thesis, completed under the supervision of Professor Lawrence Hunter, focused on investigating word embeddings for their usability in the biomedical domain using Biomedical Pubmed Central data.

Career and Research

Wadhawan is currently affiliated with IBM Research India, where she works as a Staff Research Scientist. Her research interests include hate speech detection and mitigation, controlled text generation, large language models, causality, representation learning, biomedicine, proteins, AMP, small molecule, and drug discovery.

She has published extensively in the fields of machine learning and AI, with a particular focus on text and language models, as well as biomedical applications. Her notable works include:

• "Causally Fair Language Models Through Token-level Attribute Controlled Generation" (CFL), which proposes a method to control the attributes of Language Models (LMs) for text generation tasks, achieving state-of-the-art performance for toxic degeneration.
• "Causal Graphs Underlying Generative Models: Path to Learning with Limited Data", which provides a simple algorithm to uncover causal graphs implied by generative models, aiding in interpreting latent representations.
• "Optimizing Molecules using Efficient Queries from Property Evaluations", a query-based molecule optimization framework that exploits latent embeddings, demonstrating high consistency with external validations.
• "Towards Interpreting Zoonotic Potential of Betacoronavirus Sequences with Attention", applying an attention-enhanced LSTM deep neural net classifier to predict zoonotic potential with 94% accuracy.
• "Accelerated antimicrobial discovery via deep generative models and molecular dynamics simulations", introducing the CLaSS method for attribute-controlled molecule generation, leading to the discovery of novel antimicrobial peptides.
• "Effects of Naturalistic Variation in Goal-Oriented Dialog", investigating the impact of naturalistic variation on goal-oriented datasets and proposing new testbeds, resulting in a significant performance drop for state-of-the-art end-to-end neural methods.
• "Text Style Transfer Using Partly-Shared Decoder", expanding on existing approaches to text style transfer by focusing on sentiment transfer and introducing a new decoder architecture.
• "Interactive Visual Exploration of Latent Space (IVELS) for peptide auto-encoder model selection", presenting a tool for interactive visual exploration and model selection, aiding in evaluating how well generative models capture attributes of interest.
• "Co-regularized Alignment for Unsupervised Domain Adaptation", addressing the challenge of deep neural networks generalizing across different domains by aligning source and target distributions in multiple diverse feature spaces.
• "PepCVAE: Semi-Supervised Targeted Design of Antimicrobial Peptide Sequences", a framework for designing novel antimicrobial peptide sequences by learning a rich latent space of biological peptide contexts.

Wadhawan has also been granted a patent for an "Unsupervised text style transfer method, system, and computer program product" that can classify, translate, and re-write messages into a different style.