Meta AI researchers created an end-to-end machine learning platform called Looper, with easy-to-use APIs for decision making and feedback collection

This article is based on the research paper 'LOOPER: AN END-TO-END ML PLATFORM FOR PRODUCT DECISIONS' and Meta AI article. All credit for this research goes to the researchers of this paper 👏👏👏

Please don't forget to join our ML Subreddit

From improving user experience to making computing infrastructure more efficient, AI is a crucial aspect of making today’s software systems and products as efficient as possible. Artificial intelligence is often more effective than even the precisely engineered heuristic tactics of humans today, whether it’s reducing latency, improving the quality of a video stream, or streamlining interfaces. to respond to requests from a specific person. But, to use AI more effectively in various products, several challenges must be overcome: the system must accommodate software engineers with no experience in machine learning; it should provide optimization mechanisms for a variety of product purposes, which may differ from closed-form machine learning loss functions; it must distinguish causal links from data correlations; and it must scale efficiently to train, host, and monitor a large number of AI models.

Meta Researchers is developing “Looper”, an end-to-end AI platform that was designed with easy-to-use APIs for optimization, personalization, and feedback collection to meet these needs. Looper can be used to support the entire machine learning lifecycle, from model training through deployment and inference to product evaluation and optimization. Looper allows us to modify existing products to leverage AI for custom optimizations rather than having to rebuild them around AI models. Currently, the Looper platform hosts 700 AI models and produces 4 million AI outputs every second.

Provide intelligent policies to applications

Every day, billions of people use Meta’s various services, each with their own interests and preferences.

Looper allows us to scale many of them “out of the box” on an unparalleled scale while avoiding the need for complex, specialized code.

Providing a user with dozens of options in a UI menu can make a product unappealing, regardless of its value. However, various people have distinct menu preferences. Similarly, opportunistic preloading of items that a user might view on a mobile device can significantly improve the user experience of the product, but without overwhelming the hardware capabilities of the device, one must properly predict what would be the more interesting.

  • Looper provides many features to facilitate real-time smart strategies in a scalable way: Looper is designed for use cases with modest data amounts and model complexity that require ease of use and rapid model deployment .
  • It supports a wide range of model and host types and trains a large number of decision models and policies.
  • Its flexibility to apply supervised or reinforcement learning facilitates a wide range of machine learning applications (classification, estimation, value and sequence prediction, ranking and planning). Meta (AutoML) automation tools select models and hyperparameters to balance quality, size, inference time, and other factors when used with a model management framework. Looper evaluates and optimizes everything from data sources to product impact to causal experiences.
  • Looper works in real time, unlike many other AI platforms that perform offline inference in batch mode. Many AI systems work with consistent data, such as pixels or text, but different products can have very different metadata, which comes from different places.
  • A/B testing can examine a variety of models and decision rules, such as those used by contextual bandits to simulate uncertainty in predictions on one or more targets, or reinforcement learning can be used to maximize long-term cumulative goals.
  • Unlike typical end-to-end AI systems, Looper allows meta engineers and others to see how a model is used in the software stack and experiment with different parts of the modeling framework, from selecting metrics to policy optimization.

Intelligent Strategy Deployment Platform

Unlike heavy AI models for vision, voice, and natural language processing, which prefer offline inference and batch processing, Looper uses models that can be quickly recycled and deployed in large numbers on an infrastructure. shared. The software translates metadata of user-system interactions into supervised learning labels or reinforcement learning rewards.

Looper aims for rapid integration, stable deployment and low-effort maintenance of various smart techniques, with positive benefits assessed and optimized directly in the application. Looper uses current horizontal AI platforms, such as PyTorch and Ax, with interchangeable models for machine learning tasks, and it separates application code from platform code.

Source: https://ai.facebook.com/blog/looper-meta-ai-optimization-platform-for-engineers/

Adoption and impact of a smart strategy

The vertical machine learning platform supports medium-sized models from horizontal platforms to improve many aspects of software systems. These models are easily deployed and maintained without the need for model-specific infrastructure. At Meta, Looper is used by over 90 product teams, with 690 models making 4 million predictions per second.

The range of AI skills among product teams ranged from novices to experienced AI engineers, with AI engineers making up only 15% of teams using the Looper platform. An easy-to-use AI platform is typically the deciding factor of adoption for teams without prior production AI experience, and investment in AI continues once utility has been demonstrated. Behind high-level services, Meta’s platform handles issues related to software upgrades, logging, monitoring, and other issues, unlocking significant productivity gains. A smart strategies platform increases the productivity of experienced AI developers by automating time-consuming tasks such as writing database queries, building data pipelines, and setting up monitoring and alerts. It allows product makers to deploy more AI use cases than targeted solutions. Irrespective of past AI experience, platform users implemented the first machine learning models in just days, quickly acquired training data, updated their models, and launched new products in only a few months.

There are significant opportunities to integrate innovative self-optimizing product choice strategies into software systems to improve user experience, optimize resource utilization, and support additional functions. Looper, the AI ​​platform, simplifies the deployment of smart strategies at scale by solving the pain points of product-driven, end-to-end machine learning systems. It offers immediate and measurable benefits in terms of data availability, ease of configuration, careful use of available resources, reduction of engineering effort, and assurance of product impact. The broad support for evaluating effect through causal inference and indirect resource measures is particularly appealing to platform users.

Looper makes smart tactics more accessible to software engineers, enabling product teams to service, design, deploy and upgrade AI-powered capabilities.

Article: https://arxiv.org/pdf/2110.07554.pdf

Source: https://ai.facebook.com/blog/looper-meta-ai-optimization-platform-for-engineers/

James G. Williams