Hugging Face is increasingly becomes the “hub” of AI innovation. In this episode, Merve Noyan joins us to dive into this hub in more detail. We discuss automation around model cards, reproducibility, and the new community features. If you are wanting to engage with the wider AI community, this is the show for you!
Last week I logged the very impressive Imagen project, which smarter people than me have said is the SOTA for text-to-image synthesis. Now a WIP implementation is just a
pip install imagen-pytorch away.
Architecturally, it is actually much simpler than DALL-E2. It consists of a cascading DDPM conditioned on text embeddings from a large pretrained T5 model (attention network). It also contains dynamic clipping for improved classifier free guidance, noise level conditioning, and a memory efficient unet design.
Google researchers are giving DALL-E a run for its money:
Our key discovery is that generic large language models (e.g. T5), pretrained on text-only corpora, are surprisingly effective at encoding text for image synthesis: increasing the size of the language model in Imagen boosts both sample fidelity and image-text alignment much more than increasing the size of the image diffusion model.
Don’t all AI methods need a bunch of data to work? How could AI help document and revitalize endangered languages with “human-in-the-loop” or “active learning” methods? Sarah Moeller from the University of Florida joins us to discuss those and other related questions. She also shares many of her personal experiences working with languages in low resource settings.
AI is discovering new drugs. Sound like science fiction? Not at Absci! Sean and Joshua join us to discuss their AI-driven pipeline for drug discovery. We discuss the tech along with how it might change how we think about healthcare at the most fundamental level.
We all hear a lot about MLOps these days, but where does MLOps end and DevOps begin? Our friend Luis from OctoML joins us in this episode to discuss treating AI/ML models as regular software components (once they are trained and ready for deployment). We get into topics including optimization on various kinds of hardware and deployment of models at the edge.
Curious how OpenAI’s new DALL-E 2 manages to generate impressive artwork from short natural language prompts? This article breaks down the steps then focuses in on how it does the image generation step.
In the fourth “AI in Africa” spotlight episode, we welcome Leonida Mutuku and Godliver Owomugisha, two experts in applying advanced technology in agriculture. We had a great discussion about ending poverty, hunger, and inequality in Africa via AI innovation. The discussion touches on open data, relevant models, ethics, and more.
Abubakar Abid joins Daniel and Chris for a tour of Gradio and tells them about the project joining Hugging Face. What’s Gradio? The fastest way to demo your machine learning model with a friendly web interface, allowing non-technical users to access, use, and give feedback on models.
This last week has been a big week for AI news. BigScience is training a huge language model (while the world watches), and NVIDIA announced their latest “Hopper” GPUs. Chris and Daniel discuss these and other topics on this fully connected episode!
The term “foundation” model has been around since about the middle of last year when a research group at Stanford published the comprehensive report On the Opportunities and Risks of Foundation Models. The naming of these models created some strong reactions, both good and bad. In this episode, Chris and Daniel dive into the ideas behind the report.
What happens when your data operations grow to Internet-scale? How do thousands or millions of data producers and consumers efficiently, effectively, and productively interact with each other? How are varying formats, protocols, security levels, performance criteria, and use-case specific characteristics meshed into one unified data fabric? Chris and Daniel explore these questions in this illuminating and Fully-Connected discussion that brings this new data technology into the light.
Gary Marcus makes the case that deep learning has hit a wall:
Let me start by saying a few things that seem obvious,” Geoffrey Hinton, “Godfather” of deep learning, and one of the most celebrated scientists of our time, told a leading AI conference in Toronto in 2016. “If you work as a radiologist you’re like the coyote that’s already over the edge of the cliff but hasn’t looked down.” Deep learning is so well-suited to reading images from MRIs and CT scans, he reasoned, that people should “stop training radiologists now” and that it’s “just completely obvious within five years deep learning is going to do better.”
Fast forward to 2022, and not a single radiologist has been replaced.
But he doesn’t stop there. After laying out multiple examples of deep learning failures, he change tone:
For the first time in 40 years, I finally feel some optimism about AI.
Read the article to find out why that is.
Daniel and Chris talk with Lukas Egger, Head of Innovation Office and Strategic Projects at SAP Business Process Intelligence. Lukas describes what it takes to bring a culture of innovation into an organization, and how to infuse product development with that innovation culture. He also offers suggestions for how to mitigate challenges and blockers.
Alon from Greeneye and Moses from ClearML blew us away when they said that they are training 1000’s of models a year that get deployed to Kubernetes clusters on tractors. Yes… we said tractors, as in farming! This is a super cool discussion about MLOps solutions at scale for interesting use cases in agriculture.
From MIT researchers who have an AI system that rapidly predicts how two proteins will attach, to Facebook’s first high-performance self-supervised algorithm that works for speech, vision, and text, Daniel and Chris survey the AI landscape for notable milestones in the application of AI in industry and research.
In the third of the “AI in Africa” spotlight episodes, we welcome Kathleen Siminyu, who is building Kiswahili voice tools at Mozilla. We had a great discussion with Kathleen about creating more diverse voice and language datasets, involving local language communities in NLP work, and expanding grassroots ML/AI efforts across Africa.
In addition to being a Developer Advocate at Hugging Face, Thomas Simonini is building next-gen AI in games that can talk and have smart interactions with the player using Deep Reinforcement Learning (DRL) and Natural Language Processing (NLP). He also created a Deep Reinforcement Learning course that takes a DRL beginner to from zero to hero. Natalie and Chris explore what’s involved, and what the implications are, with a focus on the development path of the new AI data scientist.
From drug discovery at the Quebec AI Institute to improving capabilities with low-resourced languages at the Masakhane Research Foundation and Google AI, Bonaventure Dossou looks for opportunities to use his expertise in natural language processing to improve the world - and especially to help his homeland in the Benin Republic in Africa.
Alexey Palazhchenko joins Natalie to discuss the implications of GitHub’s Copilot on code generation. Go’s design lends itself nicely to computer generated authoring: thanks to
go fmt, there’s already only one Go style. This means AI-generated code will be consistent and seamless. Its focus on simplicity & readability make it tailor made for this new approach to software creation. Where might this take us?
You might know about MLPerf, a benchmark from MLCommons that measures how fast systems can train models to a target quality metric. However, MLCommons is working on so much more! David Kanter joins us in this episode to discuss two new speech datasets that are democratizing machine learning for speech via data scale and language/speaker diversity.
We have all seen how AI models fail, sometimes in spectacular ways. Yaron Singer joins us in this episode to discuss model vulnerabilities and automatic prevention of bad outcomes. By separating concerns and creating a “firewall” around your AI models, it’s possible to secure your AI workflows and prevent model failure.
In the last year, I’ve talked to ~30 companies in different industries about their challenges with real-time machine learning. I’ve also worked with quite a few to find the solutions. This post outlines the solutions for (1) online prediction and (2) continual learning, with step-by-step use cases, considerations, and technologies required for each level.
In the second of the “AI in Africa” spotlight episodes, we welcome guests from Radiant Earth to talk about machine learning for earth observation. They give us a glimpse into their amazing data and tooling for working with satellite imagery, and they talk about use cases including crop identification and tropical storm wind speed estimation.
We’re joined by Eran Yahav — talking about AI assistants for developers. Eran has been working on this problem for more than a decade. We talk about his path to now and how the idea for Tabnine came to life, this AI revolution taking place and the role it will play in developer productivity, and we talk about the elephant in the room - how Tabnine compares to GitHub Copilot, and what they’re doing to make Tabnine the AI assistant for every developer regardless of the IDE or editor you choose.