Lorenz Kuhn

PhD student in Computer Science

University of Oxford

I am currently a DPhil (PhD) student in Computer Science working with Prof. Yarin Gal at the University of Oxford. I’m interested in improving our theoretical understanding of deep learning, and making deep learning safer and more reliable for real-world applications.

I recently wrote my Master’s Thesis on pruning and generalization in deep neural networks in collaboration with Prof. Yarin Gal at the University of Oxford and Prof. Andreas Krause at ETH Zürich. Even more recently, at Cohere, I worked on making large language models faster and more efficient.

Previously, I obtained a MSc in CS from ETH Zürich, and a BSc in CS from ETHZ and Imperial College London. I undertook research on medical recommendation systems at IBM Research and ETHZ, and worked as a data scientist for BCG Gamma and QantEv, an Entrepreneur First-backed InsureTech start up.

Publications

Robustness to Pruning Predicts Generalization in Deep Neural Networks

Preprint Existing generalization measures that aim to capture a model’s simplicity based on parameter counts or norms fail to …

Lorenz Kuhn, Clare Lyle, Aidan N. Gomez, Jonas Rothfuss, Yarin Gal

PDF

Efficient Smoothing of Dilated Convolutions for Image Segmentation

Preprint Dilated Convolutions have been shown to be highly useful for the task of image segmentation. By introducing gaps into …

Thomas Ziegler, Manuel Fritsche, Lorenz Kuhn, Konstantin Donhauser

PDF

Patient Risk Assessment and Warning Symptom Detection Using Deep Attention-Based Neural Networks

Health Text Mining and Information Analysis workshop at EMNLP 2018 We present an operational component of a real-world patient triage …

Ivan Girardi, Pengfei Ji, An-phi Nguyen, Nora Hollenstein, Adam Ivankay, Lorenz Kuhn, Chiara Marchiori, Ce Zhang

PDF

Implicit Negative Feedback in Clinical Information Retrieval

Medical Information Retrieval Workshop at ACM SIGIR 2016 In this paper, we reflect on ways to improve the quality of bio-medical …

Lorenz Kuhn, Carsten Eickhoff

PDF

Experience

March 2021 – August 2021

Toronto

Machine Learning Researcher

Cohere

At Cohere, I:

Conducted extensive experiments on efficient transformer architectures.
Coordinated a team effort to improve the quality of our models which lead to multiple new production models that strongly outperformed our previous generation of models.

June 2019 – August 2019

Munich

Visiting Data Scientist

BCG Gamma

At BCG Gamma, I:

Developed a text classification algorithm based on our analysis of our client’s business needs.
Reduced loading times in our prototype from two minutes to a few seconds by introducing cache warming to our Continuous Deployment pipeline.

March 2019 – June 2019

Paris

Data Science Intern

QantEv

QantEv is an insurance tech start up which emerged from the Entrepreneur First program. During my time at QantEv, I:

Developed and implemented Optimal Transport models to predict which health care providers the patients in a given region will use.
Implemented state-of-the-art text classification methods to automatically annotate text description of medical services with standardised codes (CPT).
Made large contributions to the front and back end of a web app, working with Flask and ReactJS.

August 2017 – January 2018

Zurich

Research Intern

IBM Research

Contributed to the development of an Artificial Intelligence Medical Recommendation system:

Proposed a 10x speedup for existing system.
Designed, implemented and tested novel classification algorithms leading to a 25% improvement of recommendation accuracy.
Lead project on redesign of data management system.

Projects

Bringing Boosting Black Box Variational Inference to Pyro

In this project, we contributed an implementation of Boosting Black Box Variational Inference to the Pyro probabilistic programming library.

Efficient Smoothing of Dilated Convolutions for Image Segmentation

In this project we introduce low cost methods of improving dilated convolutions in an image segmentation application. We achieve comparable results to state-of-the-art segmentation performance while being computationally more efficient than previously proposed methods.

Collaborative Filtering: Stacking Collaborative Filtering and Neural Networks for Improved Recommendations

For this project we leverage matrix factorization and neural network methods to build a recommender system for movies.