Introduction

In this Leading with Data session, our guest is Sebastian Raschka, a passionate educator and AI researcher at Lightning AI. With a background in Statistics, he’s on a mission to make AI easy to understand for everyone. Whether it’s writing bestselling books or teaching at universities, Raschka’s goal is to make AI accessible to all.

You can listen to this episode of Leading with Data on popular platforms like Spotify, Google Podcasts, and Apple. Pick your favorite to enjoy the insightful content!

Key Insights from our Conversation with Sebastian Raschka

• Mixol and model merging are innovative strategies to optimize large language models without increasing data input.
• Building AI models from scratch is a valuable educational exercise that can lead to a deeper understanding of the technology.
• The transition from academia to industry involves a shift towards more practical applications and team-based problem-solving.
• Lightning AI Studios offers a cloud-based platform that simplifies the AI research process by providing a seamless environment for experimentation.
• Staying updated with AI advancements requires selective engagement with new research and discussions within the community.
• The focus on optimizing smaller large language models is expected to continue, with gradual improvements rather than groundbreaking changes.
• Practical applications of AI, such as language translation and specialized medical models, are areas where significant impact can be made.

Join our upcoming Leading with Data sessions for insightful discussions with AI and Data Science leaders!

Now, let’s look at the questions asked by Sebastian Raschka in the session and how he responded!

How are you approaching AI innovation differently this year?

This year, I’m particularly fascinated by the creative methods being employed to extract more from AI architectures. It’s not just about feeding more data into the system anymore. We’re seeing strategies like Mixol, which combines multiple modules to create a larger model that doesn’t use all modules during each inference step. It’s a smarter approach, moving away from brute force tactics. Another exciting area is model merging, where multiple models trained on different datasets are combined. This focus on optimizing what we call ‘small’ large language models is a trend I believe will continue for the coming months, as people strive to enhance performance.

Can you share some pivotal moments from your journey in AI and biotech?

One key decision around 2012 was taking a class in statistical pattern classification, which was heavily focused on Bayesian methods for predictive modeling. This class piqued my interest because I was working on predicting the activity of small molecules in biology. It led me to delve into statistics and machine learning, and I started using Python, which opened up a new world of possibilities. Working on biological problems with messy data taught me a lot about handling real-world data challenges. After completing my PhD, I transitioned from biology applications to more general deep learning research. Eventually, I moved to Lightning AI, where I’m currently focusing on creating educational material and running experiments on our platform.

What motivated you to write your book on Python machine learning, and how has it evolved?

The motivation behind my book on Python machine learning was to document everything I had learned and to create a resource I would have liked to read myself. It started to consolidate my knowledge and share my excitement about the field. Over time, the book has evolved with new editions, moving from Theano to TensorFlow and now to PyTorch. I’m also working on new books that cover advanced topics not typically found in introductory materials, such as evaluating large language models and using multiple GPUs for deep learning.

How do you balance the pros and cons of building AI models from scratch?

Building AI models from scratch is a time-consuming process that I undertake primarily for educational purposes. It provides a deep understanding of how things work, which is invaluable for learning and experimentation. While I wouldn’t recommend from-scratch implementations for industry applications due to efficiency concerns, the process is advantageous. It lets you understand the code intimately, make informed changes and troubleshoot effectively. Building a basic model from scratch can be a great starting point for those looking to enter AI research before moving on to more advanced libraries.

How do you incorporate generative AI assistance in your writing process?

I occasionally use large language models (LLMs) to assist with my writing, especially when I’m stuck on a particular sentence or need help with grammar correction. While I don’t rely on them extensively, they can be useful for polishing my work. Having a clear idea of what you want to write about is important, and LLMs can serve as helpful companions in refining the final output.

What are the considerations for postgraduate students when choosing between academia and industry?

Choosing between academia and industry is a personal decision, and each has pros and cons. In academia, you can design your research agenda and work with students, but you may face limitations in computing resources and the frustrations of the publication process. In industry, you often work in specialized teams where you can learn from colleagues with different expertise. The transition from academia to industry may involve focusing more on engineering and practical challenges, but it can also be rewarding to work on a platform that simplifies the research process.

Can you tell us more about the Lightning AI platform and its capabilities?

Lightning AI Studios is a cloud-based platform that provides a seamless environment for running experiments and developing applications. It’s like having your own computer in the cloud, with access to Visual Studio Code, Jupyter, and the ability to switch between different accelerators. The platform saves time and money by allowing you to work on a CPU for debugging and then switch to GPUs for training without reinstalling requirements or transferring data. It’s designed to remove the hassles of setting up and running AI experiments, making research more accessible and efficient.

What does your research workflow look like today, and how do you stay updated with AI advancements?

My research workflow involves waking up early to write, followed by a workday where I occasionally check Twitter and scan arXiv for new papers. Twitter helps filter through the noise to find interesting discussions and papers, while arXiv provides a comprehensive list of new research. I’m selective about what I read in detail, focusing on papers that I find exciting and summarizing them for my newsletter. Engaging with topics that genuinely interest you is important, as this will keep you motivated and consistent in your learning journey.

What breakthroughs or trends in AI research are you looking forward to in the next 12-18 months?

In large language models, I’m excited about the continued optimization of smaller models. Last year saw many open-source breakthroughs, and I believe we’ll see more innovative ways to enhance performance without necessarily increasing model size. Techniques like model merging and focusing on dataset quality are promising directions. While I don’t expect any groundbreaking changes this year, I anticipate gradual improvements that will make AI models more efficient and customizable with less data.

Is there a specific problem you hope to see solved using large language models?

While many people focus on achieving artificial general intelligence (AGI), I’m more interested in practical applications that can benefit from AI advancements. I’d like to see improvements in language translation, grammar detection, and writing assistance. Specialized models in fields like medicine could also solve many problems if we can overcome compute intensity bottlenecks and data hunger. Ultimately, I’m looking forward to seeing AI models become more efficient and capable of delivering high-quality results with less data.

Summing Up

In our chat with Sebastian Raschka, we discovered his dedication to simplifying AI education. His journey inspires us to keep learning and exploring the world of AI. Let’s follow his lead and make AI learning a breeze for everyone!

You can connect with Sebastian Raschka on LinkedIn.

For more engaging sessions on AI, data science, and GenAI, stay tuned with us on Leading with Data.

Check our upcoming sessions here.