Mohammad Aflah Khan Email & Phone Number

Working/worked broadly on LLM Memorization, Impact of PEFT on Memorization, Knowledge Acquisition in LLMs, Retrieval-Augmented Generation (RAG) Architectures and Optimizing for Pre-Training and Inference of LLMs- Built several copilot style tools for academics which are in use at several MPIs- Built and maintaining OpenChat (an internal chatbot), MaxCast (an internal Research Paper to Podcast conversion service) and MaxChat (an internal document chat service). All three were built from the ground up including hosting the models on-prem and tuning them for maximum throughput.- Led several cross-institute collaborations - Published and submitted works in A* conferences

Working towards analyzing and exploring influence of different components of LLMs in memorization, memorization in PEFT methods and investigating factual knowledge present in LLMs- Helped get over 50x speed up over baseline HF implementation for large scale inference tasks across 50 LLMs- Submitted works to ICML and HEAL@CHI

Working towards analyzing and exploring memorization and factual knowledge related properties of LLMs- Ran large scale experiments to figure out trends related to memorization of random strings- Submitted works to ICLR & SoLaR@NeurIPS

Worked on - - Pythia Model Suite (Accepted ICML'23) - Majorly contributed to the gender bias evals and intervention case study- Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon: An intuitive taxonomy to classify memorized sequences and then build predictors based on these classes

Delhi, India

Working on building solutions to curb Hate Speech spread in Social Media Platforms using State of the Art NLP Techniques.

Bengaluru, Karnataka, India

Worked in the Finance, Planning & Analysis Engineering division towards revamping the central hub of the department. Also built POCs based on user feedback to improve the search and access experience on the webapp

Teaching Assistant for Machine Learning (ECE363/ECE563/CSE343) course taken by roughly 70 students. The work includes designing exams and assignments as well as conducting tutorials and office hours.

Building Data Augmentation Techniques for KerasNLP a library under the Keras/TensorFlow Ecosystem which aims to build industry oriented NLP Solutions

Teaching Assistant for a course on Data Structures & Algorithms (CSE102) which is taken by roughly 600 students. The work includes taking tutorial sessions and designing questions for lab exams/assignments

India

Bachelor Of Technology - Btech, Computer Science Engineering, 9.63 Cgpa (Dept. Rank 2; Batch Rank 3)

Activities and Societies: BioBytes (Coordinator), Byld (Core Member & Coordinator), Cyborg (Interim Member), LDA (Member), ESYA'23.

11Th And 12Th Grade, 95% In All India Senior School Certificate Examination

Nursery To 10Th Grade, 95.8% In All India Secondary School Examination