Insights, breakthroughs, and practical tips from my involvement in machine learning research.https://lars-quaedvlieg.github.io/https://lars-quaedvlieg.github.io/blog/foundation-model-decision-making/https://lars-quaedvlieg.github.io/blog/foundation-model-decision-making/In this post, I'm hoping to give some sort of timeline of advances made in fields like reinforcement learning and robotics that I believe could be important to know about or might be important to realize AI embodied in the real world.e will start by reviewing some classical papers that combine Transformers with reinforcement learning for multi-task control. Then, we will look at some more recent advances that are used in simulated open-ended environments, and from there on we will move on to advances in control for real-world robotics with robotics foundation models.Fri, 10 May 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-var-inf/https://lars-quaedvlieg.github.io/blog/cs330-stanford-var-inf/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. This post will talk about variational inference, which is a way of approximating complex distributions through Bayesian inference. We will go from talking about latent variable models all the way to amortized variational inference!Sat, 30 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-upt-rbm/https://lars-quaedvlieg.github.io/blog/cs330-stanford-upt-rbm/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this post is to introduce to widely-used methods for unsupervised pre-training, which is essential in many fields nowadays, most notably in the development of foundation models. We also introduce methods that help with efficient fine-tuning of pre-trained models!Tue, 19 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-upt-fsl-cl/https://lars-quaedvlieg.github.io/blog/cs330-stanford-upt-fsl-cl/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to understand the intuition, design choices, and implementation of contrastive learning for unsupervised representation learning. We will also talk about the relationship between contrastive learning and meta learning!Sat, 16 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-fsl-ml/https://lars-quaedvlieg.github.io/blog/cs330-stanford-fsl-ml/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to to understand the third form of meta learning: non-parametric few-shot learning. We will also compare the three different methods of meta learning. Finally, we give practical examples of meta learning, in domains such as imitation learning, drug discovery, motion prediction, and language generation!Thu, 14 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-obml/https://lars-quaedvlieg.github.io/blog/cs330-stanford-obml/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to understand the basics of optimization-based meta learning techniques. You will also learn about the trade-offs between black-box and optimization-based meta learning!Sun, 10 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-bbml-icl/https://lars-quaedvlieg.github.io/blog/cs330-stanford-bbml-icl/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to learn how to implement black-box meta-learning techniques. We will also talk about a case study of GPT-3!Sun, 03 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-tl-ml/https://lars-quaedvlieg.github.io/blog/cs330-stanford-tl-ml/This lecture is part of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to learn how to transfer knowledge from one task to another, discuss what it means for two tasks to share a common structure, and start thinking about meta learning.Sun, 03 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-mtl/https://lars-quaedvlieg.github.io/blog/cs330-stanford-mtl/This is the first lecture of the CS-330 Deep Multi-Task and Meta Learning course, taught by Chelsea Finn in Fall 2023 at Stanford. The goal of this lecture is to understand the key design decisions when building multi-task learning systems.Sat, 02 Mar 2024 00:00:00 GMThttps://lars-quaedvlieg.github.io/blog/cs330-stanford-introduction/https://lars-quaedvlieg.github.io/blog/cs330-stanford-introduction/I have been incredibly interested in the recent wave of multimodal foundation models, especially in robotics and sequential decision-making. Since I never had a formal introduction to this topic, I decided to audit the Deep Multi-Task and Meta Learning course, which is taught yearly by Chelsea Finn at Stanford. I will mainly document my takes on the lectures, hopefully making it a nice read for people who would like to learn more about this topic!Fri, 01 Mar 2024 00:00:00 GMT