Phillip Lippe Email & Phone Number

United States

Amsterdam, North Holland, Netherlands

Senior Research Scientist at NXAI in the LLM Team (joined temporarily to bridge the Visa application time to join Google DeepMind in the US). I focused on large-scale pretraining, designing a JAX-based code base which can scale to training 70B models. I conducted research on a 7B model scale, distributed over 256 H100 GPUs, and focused on architecture research for advancing the xLSTM as a Transformer alternative. Further, I wrote and optimized Pallas and Triton Kernels on H100 GPUs.

Amsterdam, North Holland, Netherlands

I was a PhD student in the QUVA Lab supervised by Efstratios Gavves, Taco Cohen, Sara Magliacane, and Yuki Asano. I was also part of the ELLIS PhD program in cooperation with Qualcomm. My research focused on the intersection of causality and machine learning, particularly on causal representation learning and temporal data.

Amsterdam

I have been a teaching assistant for the following graduate courses of the MSc Artificial Intelligence at the University of Amsterdam:- Deep Learning (2019, 2020, 2021, 2022)- Natural Language Processing 1 (2019)- Fairness, Accountability, Confidentiality and Transparency in AI (2020)- Information Retrieval 1 (2020)- Advanced Topics in Computational Semantics (2020, 2021)- Foundation Models (2024)I have given lectures in the courses:- Deep Learning (2020, 2021, 2022)- Advanced Topics in Computational Semantics (2020, 2021, 2022, 2024)- Foundation Models (2024)For the Deep Learning course, I have created a public series of implementation-based tutorials, which is continuously updated with new content. The tutorials have ~40k page visits per month with >2k GitHub stars, and are part of the official PyTorch Lightning documentation. For the Foundation Models course, I have created 10 tutorials on distributed training at scale (e.g. data, pipeline, tensor parallelism). For more details, see the link below.

Amsterdam Area, Netherlands

Research Assistant in the AIRLab (ILPS) at the University of Amsterdam in cooperation with Ahold Delhaize (full-time till August, part-time till December). My research focused on advancing task-oriented dialogue systems to conduct human-like conversations with diverse, natural responses. The results are summarized in our paper "Simultaneously Improving Utility and User Experience in Task-oriented Dialogue Systems".

Amsterdam, North Holland, Netherlands

Student Research Intern at Google DeepMind. I was supervised by Mostafa Dehghani and focused on generative multi-modal pretraining. This included training models at scales of up to 4 billion parameters, distributed across 512 devices. The developed model is now internally used in the Gemini foundation model project. Tasks included:- Designing a generative pretraining objective for datasets with interleaved text-image sequences (e.g. blog posts).- Parallelizing models over up to 512 devices with tensor and data parallelism.- Training Generative Multi-Modal Large Language Models (VLM) with up to 4 billion parameters.- Setting up evaluations and demo of developed generative model.- Investigating scaling laws for VLMs.

Amsterdam, North Holland, Netherlands

Research Intern in the AI4Science lab at Microsoft Research. I was supervised by Johannes Brandstetter and focused on improving scientific simulations like weather modeling with neural solvers. We developed a Diffusion-like Neural PDE Solver which provides significantly longer accurate rollouts in complex, chaotic PDEs like Navier Stokes and weather dynamics. Our research was published at NeurIPS 2023. Tasks included:- Analyzing limitations of current methods and challenges of long stable predictions of Neural PDE solvers.- Developing PDE-Refiner, a novel method which, based on noise-based refinement, provides significantly longer accurate predictions.- Implementing and Training Video Diffusion Models on 1D and 2D PDEs.- Running large hyperparameter searches and model comparisons.

Amsterdam Area, Netherlands

Research Assistant (part-time) in the field of high-order automated theorem proving developing a hammer for Lean. A hammer reduces high-order formulas into first-order clauses to efficiently check for possible proofs and makes manual formalizations much easier. The research was part of the Lean Forward project (https://lean-forward.github.io) supervised by Jasmin Christian Blanchette.

Stuttgart Und Umgebung, Deutschland

I have conducted my Bachelor studies in cooperation with Daimler AG. See the internships below for projects conducted during the Bachelor.

Stuttgart Vaihingen, Baden-Württemberg, Germany

Joining the Image Understanding team for autonomous driving at Daimler AG for an internship, with the goal of writing a Bachelor Thesis. The topic of the project was "Hierarchical multi-label object detection of rare classes for autonomous driving"Tasks:- Development of hierarchical multi-label classification with object attributes for bounding box detection and semantic segmentation- Development of a hierarchical metric- Development of a metric loss function based on hierarchical IoU- Optimization of training on rare classes with strongly imbalanced data

San Francisco Bay Und Umgebung

Joined the trajectory planning team for autonomous driving at MBRDNA. The topic of my project was "Deep Prediction Learning for Autonomous Driving"Tasks:- Development of a recurrent network architecture- Handling Generative Adversarial Neural Networks (GAN)- Software engineering with Scrum

Stuttgart, Baden-Württemberg, Germany

Joined the Image Understanding team for autonomous driving at Daimler AG for an internship on "Efficient training of neural networks for detecting special vehicles".Tasks: - Development of efficient training methods- Bounding Box Detection with SSD and YOLO9000 - Modelling a dataset for semi-supervised learning

Phd, Artificial Intelligence

My research focuses on Causality and Machine Learning, in particular Causal Representation Learning for temporal data. Supervisors.

Master Of Science - Msc, Artificial Intelligence, 9.5, Cum Laude (Dutch Grading System, 1-10 With 10 Being Best)

Master study programme in Artificial Intelligence with strong research orientation. Courses include Machine Learning and Deep Learning.

Bachelor Of Engineering - Be, It-Automotive, 1.0 (German Grading System, 1.0-4.0 With 1.0 Being The Best)

The study program was organized in cooperation with Daimler. The field of study was Computer Science with a focus on the application.

Schüleruni, Mathematics And Computer Science

Activities and Societies: SchoolUni BochumVisiting and completing university courses in mathematics and computer science as a high-school.