May 11, 2023
AI has long been a crucial component of Meta’s ads system. We began with manual feature engineering for small models and progressed to building hundreds of deep neural network models with trillions of parameters. Each model is independently optimized for different goals — such as improving ad quality to provide better experiences for people or increasing conversion rates for a higher return on ad spend for our advertisers.
Meta continues to take bold steps to advance and deploy state-of-the-art AI and deliver a step change in our ads system performance. We’re rolling out more powerful AI models to improve performance across all ad types and ad surfaces. We aim to achieve this through deeper alignment with advertiser objectives, and utilizing the rapid expansion of high-growth areas, like short-form video, to provide enjoyable experiences to people — all while working to safeguard privacy.
Recently, we built and deployed Meta Lattice, a new model architecture that learns to predict an ad’s performance across a variety of datasets and optimization goals that were previously supported by numerous smaller, siloed models. It enhances Meta’s ads system in the following ways:
Better performance. Meta Lattice is capable of improving the performance of our ads system holistically. We’ve supercharged its performance with a high-capacity architecture that allows our ads system to more broadly and deeply understand new concepts and relationships in data and benefits advertisers through joint optimization of a large number of goals.
Early results from our deployment on Instagram show that knowledge-sharing across its different surfaces (e.g., Feed, Story, and Reels) and across various advertiser objectives (e.g., clicks, video views, and conversions) increased performance for advertisers. Joint optimization of value for people and advertisers resulted in better ad experiences for people, showing an ~8 percent improvement in ads quality.
Improved AI efficiency. Maintaining and advancing fewer, more powerful models makes our overall ads system more nimble in adopting future AI innovations, driving greater value for our advertisers. We also expect that transitioning to Meta Lattice will allow our fleet to improve compute efficiency, freeing up resources to explore new frontiers in AI.
Faster adaptability to the shifting market landscape. As people’s expectations of how their data is used continue to evolve, regulations and policies from governments and industry players are also changing. Evolving data use regulations and platform practices change the type and amount of data available to machine learning models. We've designed Meta Lattice to drive advertiser performance in the new digital advertising environment where we have access to less granular data. Additionally, Lattice is capable of generalizing learnings across domains and objectives, which is especially crucial when the model has limited data to train on. Fewer models also means we can proactively and efficiently update our models and adapt to the fast-evolving market landscape.
Before the use of this model architecture, Meta Ads model space was projected to grow substantially in the coming years as new surfaces, advertiser products, and privacy practices emerge. Maintaining a large model space often leads to slower proliferation of AI innovations and compute inefficiency.
We have built a holistic model architecture that incorporates disparate signals and balances performance across domains and objectives. Furthermore, balancing model performance with compute efficiency is a complex and compelling technical challenge.
To overcome these challenges, we built the following key components:
Holistic understanding of both advertisers’ and people's objectives. Meta Lattice can understand both the common usage patterns and the unique and latent people-advertiser engagement patterns from heterogeneous data sources, through multi-domain, multi-task learning and armed by sparse activation techniques. This mechanism is particularly useful for the “cold start” problem — people can receive more relevant ad recommendations on emerging products and surfaces, even though there is little data to learn from, through better generalization.
Handling delayed feedback. The engagement between an ad and a person viewing the ad can span from seconds (e.g., click, like) to days (e.g., considering a purchase, adding to a cart, and later making the purchase from a website or an app). Through multi-distribution modeling with temporal awareness, Meta Lattice can capture not only a person’s real-time intent from fresh signals but also long-term interest from slow, sparse, and delayed signals.
Balancing multiple domains and objectives. Meta Lattice is able to balance performance across multiple domains and objectives and reach a status where no objective can be further improved without hurting the rest (aka Pareto optimality). Techniques such as Pareto-front feature selection, MetaBalance , help avoid manually tuning the performance of thousands of different domains and tens of different objectives.
Advanced model scaling. Meta Lattice has trillions of parameters, is trained on hundreds of billions of examples from thousands of data domains, including Meta’s platform surfaces, and our advertiser-facing products. Our customized Deep and Hierarchical Ensemble Network model, built on top of a Transformers backbone, is highly scalable on GPUs.
Maximizing AI Capex efficiency. Previously, hundreds of models were separately trained, served, and optimized. Now we’ve introduced two levels of resource-sharing: (1) horizontal sharing across domains, objectives, and ranking stages through joint optimization; and (2) hierarchical sharing from large, high-capacity upstream models to lightweight downstream vertical models. Through resource-sharing enhancement, we can significantly reduce the amount of computational needs.
As businesses face continuous shifts in consumer behavior, economic slowdowns, and ongoing changes to industry data use practices and restrictions, we need smarter, more flexible AI systems that can address these challenges quickly and efficiently.
Meta Lattice is one way we’re using AI more broadly and deeply to enhance Meta’s ads system. The system will now continuously learn the essential characteristics that improve ad performance across various surfaces, objectives, and ad types simultaneously. Going forward, we’ll continue to further iterate on Meta Lattice. This new model architecture creates a more nimble system — one that is more adaptable to broader market changes, can quickly utilize new AI innovations, and operates more efficiently to deliver the results that help businesses grow.
We’re announcing updates to Facebook’s population density maps, which can be used to coordinate and improve the delivery of humanitarian aid around the world, including global COVID-19 vaccinations.
April 15, 2021
Working with Inria researchers, we’ve developed a self-supervised image representation method, DINO, which produces remarkable results when trained with Vision Transformers. We are also detailing PAWS, a new method for 10x more efficient training.
April 30, 2021