Emily Liu

Emily Liu

Machine learning engineer

Papers and Publications

Learning Genetic Perturbation Effects with Variational Causal Inference

Published:

Understanding how genes interact and respond to perturbations is crucial for uncovering the mechanisms of cells and identifying potential ways to treat diseases. Recent advances in sequencing technologies now allow us to measure how individual cells react when specific genes are altered. However, making sense of this complex data requires advanced computational tools. In our work, we address the challenge of predicting how cells respond to potentially new untested genetic perturbations. We noticed that while deep learning models perform well on data measured before, they struggle with making predictions on new cases. On the other hand, models based on biological understanding can, in theory, make better predictions, but they often rely on overly simple assumptions that do not hold with real-world data. We developed a new method that combines the strengths of both approaches. Our model, called SCCVAE, uses knowledge of gene networks together with deep learning to better predict how cells will respond to gene changes. It can simulate new experiments and help identify groups of genes that work together. This tool could be valuable for researchers studying perturbational changes, as well as gene functions and diseases.

Recommended citation: Learning Genetic Perturbation Effects with Variational Causal Inference. Emily Liu, Jiaqi Zhang, Caroline Uhler.
Download Paper

Relative Advantage Debiasing for Watch-Time Prediction in Short-Video Recommendation

Published:

This work tackles the challenge of using watch time as a signal for user preference in video recommendation systems. Raw watch time is heavily influenced by factors like video length, popularity, and individual user habits, which can distort what users actually enjoy. We introduce a relative-advantage debiasing framework that compares a user’s watch time to reference distributions conditioned on similar users and items, producing a quantile-based, more reliable measure of preference. Our two-stage architecture cleanly separates the tasks of distribution estimation and preference learning, and we further develop distributional embeddings that make quantile prediction efficient without storing historical data. Both offline and live experiments show substantial gains in accuracy and robustness over existing approaches.

Recommended citation: Relative Advantage Debiasing for Watch-Time Prediction in Short-Video Recommendation. Emily Liu, Kuan Han, Minfeng Zhan, Bocheng Zhao, Guanyu Mu, Yang Song. Association for the Advancement of Artificial Intelligence (AAAI) Conference 2026.
Download Paper

Leveraging Intermediate Neural Collapse with Simplex ETFs for Efficient Deep Neural Networks

Published:

This paper examines the phenomenon of neural collapse, where network activations, class means, and linear classifier weights converge to a simplex equiangular tight frame (ETF) during training, enhancing interpretability, robustness, and generalization. Building on findings that neural collapse extends beyond the final layer in fully connected networks, the authors propose two novel methods: Adaptive-ETF, which enforces simplex ETF constraints across all layers beyond a certain depth, and ETF-Transformer, which applies these constraints to feedforward layers in transformer blocks. Both methods maintain performance while significantly reducing trainable parameters, offering efficient alternatives for network training and regularization.

Recommended citation: Leveraging Intermediate Neural Collapse with Simplex ETFs for Efficient Deep Neural Networks (2024). Emily Liu. NeurIPS 2024 Workshop on Mathematics of Modern Machine Learning.
Download Paper

Semi-Supervised Neural Processes for Articulated Object Interaction

Published:

This paper addresses the challenge of limited labeled action data for robotic object manipulation by introducing the Semi-Supervised Neural Process (SSNP). SSNP combines small amounts of labeled interaction data with abundant unlabeled visual data, using a jointly trained reward-prediction and autoencoding framework to extract task-relevant features. This approach reduces the need for extensive retraining and computational resources while improving generalization. The model outperforms other semi-supervised methods in a door-opening task, achieving superior performance with significantly less data.

Recommended citation: Semi-Supervised Neural Processes for Articulated Object Interaction. Emily Liu, Michael Noseworthy, Nicholas Roy. RSS 2024 Workshop on Structural Priors as Inductive Biases for Learning Robot Dynamics.
Download Paper | Download Slides

Structured Latent Variable Models for Articulated Object Interaction

Published:

This paper explores how a robot can learn a low-dimensional representation of doors from videos of them opening or closing, enabling the inference of door-related parameters and interaction outcomes. Instead of relying solely on labeled datasets, the study employs a semi-supervised approach using the Neural Statistician, a structured latent variable model that separates shared context-level variables (common across all images of the same door) from instance-level variables (specific to each image). The model effectively generates realistic door image embeddings, which outperform context-free baselines in tasks such as predicting door parameters and optimizing actions in a visual bandit door-opening scenario, demonstrating its utility for more efficient and accurate robotic interaction.

Recommended citation: Structured Latent Variable Models for Articulated Object Interaction. Emily Liu, Michael Noseworthy, Nicholas Roy. 2023.
Download Paper