He-Yen (Louis) Hsieh

Hi, I'm He-Yen (Louis) Hsieh, a Ph.D. candidate in Computer Science at Harvard University, advised by Prof. H. T. Kung.

I enjoy building machine learning algorithms that make AI models smaller, faster, and easier to adapt for real-world deployment. My work spans efficient AI, generative AI, large language models, and computer vision, with recent projects in model adaptation, post-training quantization, knowledge distillation, efficient inference, gaze estimation, and personalization.

What I enjoy most is turning research ideas into working AI systems: designing new algorithms, building prototypes, and making new ideas practical.

Before and during my Ph.D., I have been fortunate to work with researchers at Intel, Reality Labs Research at Meta, Adobe Research, and Academia Sinica. My work has been published at CVPR, AAAI, ECCV, WACV, ICIP, and ICASSP.

I'm always happy to connect with people interested in efficient AI, model compression, generative models, or research ideas for making AI smaller, faster, easier to adapt, and still accurate.


Visualizing Ideas

I enjoy using visual diagrams, animations, and interactive demos to distill technical ideas so they are easier to understand. For example, during my time as a teaching assistant, I built a step-by-step visualization of Dijkstra’s algorithm to explain graph search. I also like making research more accessible through demos, including GazeGen and Gen4Gen.

Step-by-step visualization of Dijkstra's algorithm.


Research Roadmap

My research centers on making AI practical under limited data, compute, and memory. I started from computer vision problems such as detection, video understanding, and anomaly detection, then expanded into gaze estimation, generative AI, and efficient learning systems. More recently, my work focuses on efficient adaptation and compression for vision and language models.

Efficient AI
Smaller, faster, easier to adapt, and still accurate.

Efficient Model Adaptation

Alfa — Efficient few-shot model adaptation

2026

Adapts gaze models to new users and domains with only 5 samples by reusing pre-trained knowledge, making the model 5× smaller while keeping or improving accuracy. The same low-rank adaptation idea also extends to diffusion-based language models for reasoning.

Use when
  • A pre-trained model needs compact adaptation.
  • Only a few target samples are available.
  • More trainable capacity is needed without increasing inference-time model size.
  • AAAI'26 · Model adaptation · Gaze estimation

    DFT Gaze — Few-shot gaze personalization

    2025

    Compresses gaze estimation to a tiny 281K-parameter model, keeping 90% of the precision of a 10× larger model while supporting real-time personalization on edge devices.

    Use when
  • A large vision model is too slow for real-time deployment.
  • Personalization is needed with only a few user samples.
  • A compact model must stay accurate under edge-device constraints.
  • ICIP'25 (Spotlight) · 5-shot · Model compression

    Efficient AI Deployment

    Calibration-free PTQ for LLMs

    Ongoing

    Makes 3-/4-bit LLM deployment easier when calibration data is unavailable or hard to collect.

    Use when
  • Make small LLMs even smaller with low-bit quantization.
  • Quantize without calibration data for deployment budgets.
  • LLMs · PTQ · Quantization

    DFT Gaze — Compact vision model design

    2025

    Compresses gaze estimation to a tiny 281K-parameter model, keeping 90% of the precision of a 10× larger model while supporting real-time personalization on edge devices.

    Use when
  • A large vision model is too slow for real-time deployment.
  • Personalization is needed with only a few user samples.
  • A compact model must stay accurate under edge-device constraints.
  • ICIP'25 (Spotlight) · Distillation · Edge AI

    EDIT — Faster diffusion language model inference

    2025

    Speeds up diffusion-based language model reasoning by stopping inference once generation has stabilized.

    Use when
    • Diffusion-based language models use more inference steps than needed.
    • Reasoning tasks need faster generation under test-time compute budgets.
    NeurIPS OPT'25 · LLMs · Efficient inference

    Generative AI

    Personalized text-to-image generation

    2023

    Few-shot personalized generation from only a few user-provided examples.

    Use when
    • Personalized generation is needed from only a few examples.
    • Multiple user-provided concepts must appear in one generated image.
    Personalization · Text-to-image · Diffusion models · Few-shot generation

    Gen4Gen — Generative AI for training data synthesis

    2025

    Builds a generative pipeline to create diverse synthetic training data for stronger downstream vision models.

    Use when
    • Generative models can be used to build a data synthesis pipeline.
    • Real training data is limited, expensive, or hard to collect.
    BMVC'25 · Generative AI · Data synthesis · Few-shot generation

    GazeGen — Gaze-guided generation

    2024

    Turns gaze into a natural control signal for interactive image and video generation.

    Use when
    • User intent should guide generation without heavy manual interaction.
    • Gaze can serve as a lightweight control signal for interactive AI systems.
    Generative AI · Gaze estimation · Interaction · Few-shot learning

    Computer Vision

    S3R — Video anomaly detection

    2022

    Detects abnormal events in videos with weaker supervision, reducing the need for heavy manual annotation.

    Use when
    • Video understanding needs to scale without dense frame-level labels.
    • Rare events must be detected from compact normal-pattern representations.
    ECCV'22 · Video anomaly detection

    Adaptive Image Transformer — One-shot object detection

    2021

    Recognizes new object categories from only one example, making detection more flexible.

    Use when
    • New object categories must be detected from only one example.
    • Full training data is unavailable for every new class.
    CVPR'21 · Object detection · One-shot learning

    Temporal action detection

    2020–2023

    Finds actions in long videos from only one example, using zoom-in attention to refine coarse-to-fine action locations.

    Use when
    • Long videos need action localization with limited examples.
    • Coarse video cues must be refined into precise temporal action boundaries.
    ICASSP'23 · One-shot learning · Action detection · Zoom-in attention

    How Can AI Work with Less?

    Choose a constraint and see how my projects help an existing AI model work with less.

    Your Existing AI Model
    Across model types
    Vision Language Generation Video
    What constraint does your AI face?
    Less Data How it works

    Alfa

    1 Start
    2 Constraint
    3 Method
    4 Outcome
    Why it matters