Lewis Tunstall: Hugging Face, SetFit and Reinforcement Learning | Learning from Machine Learning #6
Oct 3, 2023
auto_awesome
Lewis Tunstall, machine learning engineer at Hugging Face and author of the best selling book Natural Language Processing with Transformers, talks about his journey from theoretical physics to machine learning, the benefits of the Hugging Face platform, exploring reinforcement learning and the TRL library, adapters in fine-tuning pre-trained transformers, the limitations of language models, and understanding biases in machine learning and the impact on society.
Focusing on one domain in machine learning maximizes learning rate and provides a strong foundation for branching out later.
Contributing to open source projects accelerates learning and provides feedback from experienced maintainers.
Maintaining a balanced perspective and focusing on problem-solving in machine learning helps avoid exhaustion and reminds us to cherish human connections.
Deep dives
Discovering the Power of Machine Learning
The speaker recounts their first encounter with machine learning through a friend's code that outperformed physicists' work in classification, leading to their fascination with the field and eventual career change.
The Birth of Hugging Face and Transformers
The guest shares their journey from academia to industry and how they got involved with Hugging Face, an open-source company. They discuss the development of the Transformer models and their collaboration in writing a popular book about Natural Language Processing with Transformers.
Set-Fit: Efficient Few-Shot Learning with Sentence Transformers
The guest explains the concept and benefits of Set-Fit, an approach that combines sentence transformers and linear classifiers to tackle the problem of limited labeled data. They discuss how it significantly reduces computational requirements while achieving competitive performance. They also touch on other topics like reinforcement learning, including Proximal Policy Optimization (PPO) and AI feedback in model evaluation.
Maximizing Learning Rate and Focusing on One Domain
When starting out in a career in machine learning, it is important to maximize your learning rate by focusing on one domain or problem area rather than trying to learn everything at once. Going deep in one area can provide a strong foundation for branching out later. Additionally, it is beneficial to move away from typical toy problems and tackle novel projects that require problem-solving in uncertain circumstances. Contributing to open source projects can greatly accelerate learning as it provides the opportunity to understand the internals of libraries and gain feedback from experienced maintainers.
Not Taking Life Too Seriously and Appreciating Human Connections
A career in machine learning has taught the importance of not taking life too seriously and maintaining a balanced perspective. Navigating debates and challenges in the field requires a level-headed approach to avoid exhaustion. By focusing on problem-solving in machine learning, it allows for a more relaxed and appreciative mindset in personal life. It reinforces the value of human connections and non-AI aspects of life, reminding us to cherish the things that make us human.
This episode features Lewis Tunstall, machine learning engineer at Hugging Face and author of the best selling book Natural Language Processing with Transformers. He currently focuses on one of the hottest topic in NLP right now reinforcement learning from human feedback (RLHF). Lewis holds a PhD in quantum physics and his research has taken him around the world and into some of the most impactful projects including the Large Hadron Collider, the world's largest and most powerful particle accelerator. Lewis shares his unique story from Quantum Physicist to Data Scientist to Machine Learning Engineer.
