Franklin Wang Email & Phone Number

California, United States

Projects:Nearest Neighbor Normalization Improves Multimodal Retrieval (in collaboration with Stanford NLP)- Discovered general debiasing method for multimodal retrieval tasks, consistently improving performance of all tested contrastive models (eg. CLIP, BLIP, BEiT-3, SigLIP) on text/image retrieval benchmarks without any training- Improved gender debiasing on image retrieval while improving retrieval accuracy (all previous debiasing methods decrease accuracy)A Multimodal Automated Interpretability Agent- Develop GPT-4-Vision agent for interpreting & experimenting on neurons in vision models, matching human expert experimenter performance- Applied approach to spurious feature removal on vision models, improving robustness to irrelevant background features and attaining near-SOTA performance on a spurious correlation benchmark w/o requiring any annotations (which current SOTA methods require)- Fine-tuned open-source LLMs on large multi-GPU clusters to act as judges and scorers for benchmark evaluationIntuitive Physics with Graph Neural Nets and Transformers- Designed transformer and GNN architectures to simulate solids & fluids using particle cloud representations- Ran experiments using data from ThreeDWorld to prepare for experiments on real-world data

San Francisco Bay Area

- Implement DPO into Anyscale training library and accelerate training with asynchronous preprocessing of reference model probabilities with Ray- Conduct case study on preference tuning for summarization: generated synthetic preference data with Q&A-based summarization metric to finetune Mistral-7B with multi-iteration online DPO, beating GPT-4o performance- Published blog post on scaling synthetic data & DPO training with Anyscale: https://www.anyscale.com/blog/direct-preference-optimization-with-synthetic-data

- Researched graph neural net approaches to modeling the dynamics of small molecules, leading to >300 times speedup compared to simulations- Created custom data loading caching system to significantly reduce redundant graph neural net calculations

- Contrastively train Bi-LSTM model using TensorFlow to improve sentence embeddings for empathy detection in call center transcripts- Experiment with multimodal (audio + text) models for emotion prediction

- Taught and develop the curriculum for 40 Python programming class sessions that have reached a total of over 300 students so far- Organized the Code2K 2021 Spring and Summer events, which reached over 200 students around the world (eg. Kazakhstan, Canada) and had 18 total workshop sessions ranging from machine learning to Discord bots- Developed the format of the event + led weekly meetings to check progress- Created and taught one of the workshops, which used Teachable Machine to allow middle school students to create their own programs using machine learning.

- Created and taught the curriculum which made complex topics in deep learning like gradient descent and linear algebra more accessible to high school students

- Used C++ and C# for mesh decimation, sensor tracking & registration, navigation visualization, and sensor accuracy evaluation- Worked frequently with quaternions, rotation matrices, and vectors

Computer Science And Mathematics Double Major

Education record