The Biggest Week in AI

This week exploded with enough announcements and news to fill hundreds of pages– with PyTorch 2.0 and V5 Midjourney released just as we were about to hit publish! New products using GPT-4 poured into our timelines along with criticism about the accompanying 'technical report– including our very own CEO’s bold perspective featured in VentureBeat. We’re celebrating the release of our biggest update to PyTorch Lightning yet, and the world of ML research continued pushing the limits of language models and GANs and more. Let’s dive in!

Featured Story

PyTorch Lightning was launched four years ago and has exceeded initial expectations with 4 million monthly downloads and a vibrant community. Over the last year, the community has been more active than ever, giving us feedback and helping shape the direction of PyTorch Lightning. Yesterday, we revealed the biggest update to our open-source libraries yet!

We're introducing PyTorch Lightning 2.0—as well as?Fabric, a new library—to continue unlocking unprecedented scale, collaboration, and iteration for researchers and developers.

Highlights Include:

• Commitment to backward compatibility in the 2.0 series
• Simplified abstraction layers, removed legacy functionality, integrations out of the main repo. This improves the project's readability and debugging experience.
• Introducing Fabric! Scale any PyTorch model with just a few lines of code.

Dig Deeper:

• Release notes
• Our CTO’s blog about the development of this release
• The FAQ

Research Highlights

• Meet in the Middle: A New Pre-training Paradigm. Researchers from Microsoft Azure AI proposed a new language model pre-training paradigm that uses left-to-right and right-to-left sequences during training and inference. The authors introduced wwo techniques: aligning the predictions of left-to-right and right-to-left models and combining their outputs through bidirectional inference. Their proposed paradigm claims to improve training data efficiency and model capabilities in infilling tasks.
• Scaling up GANs for Text-to-Image Synthesis. A new GAN architecture called GigaGAN has been proposed by researchers from POSTECH, Carnegie Mellon University, and Adobe Research. GigaGAN aims to demonstrate the viability of using GANs for text-to-image synthesis. The authors claim this architecture can synthesize high-resolution images, is faster at inference time, and supports various latent space editing applications.
• Eliciting Latent Predictions from Transformers with the Tuned Lens. The "tuned lens" method has been developed by researchers from Eleuther AI and others to analyze transformers and understand how model predictions are refined layer by layer. The authors clain that it decodes every hidden state into a distribution over the vocabulary, making use of an affine probe for each block in a frozen pretrained model. This technique has been claimed to be tested on various autoregressive language models with up to 20B parameters, showing to be more predictive, reliable, and unbiased than the earlier "logit lens" technique. Causal experiments claimed that the "tuned lens" uses similar features to the model itself, and it can be used to detect malicious inputs accurately.

ML Engineering Highlights

With OpenAI’s announcement of GPT-4 came a flurry of interesting use cases including:

• Be My Eyes presented Virtual Volunteer?, a digital visual assistant that will answer any question about an image and provide instantaneous visual assistance in real-time within the app.
• Duolingo released Duolingo Max, a subscription tier above “Super” that gives you access to your own personal, AI-powered language tutor through Explain My Answer and Roleplay, two features developed with GPT-4.
• Intercom introduced Fin, a “ChatGPT for customer service” bot that immediately answers customer questions, reducing support volume and resolution times.

Open Source Highlight

PyTorch, the popular deep learning framework, has released its latest version, PyTorch 2.0. The update offers several improvements, including better support for mobile and web deployment, faster performance, and improved usability. PyTorch 2.0 includes key features such as dynamic computation graphs and autograd and provides code examples for training neural networks. The update also introduces new features such as support for TorchScript and ONNX and provides guidance on upgrading from previous versions of PyTorch.

Tutorial of the Week

This week, we’re all about migrating to PyTorch Lightning 2.0. To that end, we’ve put together two helpful resources for you. The first is a migration guide that’ll walk you through upgrading from 1.x to 2.0. The next is an FAQ that includes some of the most commonly asked questions related to PyTorch Lightning 2.0.

Community Spotlight

• This open community PR introduces a better filtering mechanism to fix checkpoint finding by only looking at files with the correct extension. Simple, clean, and efficient solution. Kudos to GitHub user yassersouri!
• Now this, we love to see. This open community PR customizes gradio components with on-brand Lightning colors. These new components can be used with lightning run app lit_gradio_interface_demo.py in the?tests dir. Great stuff, Irina Kogay!
• Finally, this merged community PR introduces a method to check whether a Fabric model has already been set up. It fixes an earlier issue introduced by our team where the need to determine whether a Fabric model had already been set up was raised. Great work, Atharva Phatak!

Lightning AI Highlights

• Join the adrenaline-fueled PyTorch Lightning Bicoastal Hackathon in Palo Alto or NYC on April 1st! Collaborate with top developers, learn from mentors, and showcase your skills. With a $2,500 prize and community exposure, ignite your PyTorch passion and push the limits of what's possible with PyTorch Lightning. Don't miss this chance to code, innovate, and create!
• Learn Distributed Model Training with PyTorch from Dev Advocate Aniket Maurya during his “Code With Me” Discord livestream on March 17th. Train PyTorch in a distributed fashion on MPS, Cuda, and TPUs and discover different strategies with PyTorch Lightning. Level up your coding game and sign up now for this live coding journey!
• PyTorch Lightning 2.0 and Lightning Fabric are here. Take a look at our refreshed website and the brand new home for all of your open-source deep learning needs, complete with everything from basic guidance to in-depth tutorials.

Don’t Miss the Submission Deadline

• AutoML-Conf 2023: The international conference on automated machine learning. Sep 12-15, 2023. (Berlin, Germany). Submission deadline: Fri Mar 24 2023 04:59:59 GMT-0700
• BMVC 2023: 34th annual conference on machine vision, image processing, and pattern recognition. Nov 20 - 24, 2023. (Aberdeen, United Kingdom). Submission deadline: Fri May 12 2023 16:59:59 GMT-0700
• NeurIPS 2023: 37th conference on Neural Information Processing Systems. Dec 10 - 16, 2023. (New Orleans, Louisiana). Submission deadline: Wed May 17 2023 13:00:00 GMT-0700

Want to learn more from Lightning AI? “Subscribe” to make sure you don’t miss the latest flashes of inspiration, news, tutorials, educational courses, and other AI-driven resources from around the industry. Thanks for reading!