AI has become a tool so that infrastructure can self-optimise, self-manage and self-secure. Organisations today rely heavily on AI’s ability to improve the quality of a video stream, streamlining interfaces to suit a person’s requirements and reducing latency. But for organisations to be able to leverage AI effectively, there are several bumps in the road that must be removed. Software engineers without a background in machine learning must fit in, mechanisms must be built to optimise for several product goals, causal connections must be separated from correlations in the data, and the system must be scalable to be able to train and monitor a big number of AI models.
Last week, Meta released an end-to-end AI platform called Looper that optimises, personalises and collects feedback with easy-to-use APIs. Looper monitors and supports a system through the entire machine learning lifecycle, starting from training a model, deployment and inference, ending with evaluation and tuning. This way, Looper can upgrade existing products to use AI for personalised optimisation instead of rebuilding existing products around AI models.
Sign up for your weekly dose of what's up in emerging technology.
Source: Meta AI
Looper makes life easier for those who avail Meta’s services and allows them to customise these choices according to their preferences. Having a product with an overloaded UI menu can make a product difficult to use despite the several functions it might offer. Also, prefetching content that will likely be watched by the user can make the user experience for a product better. The only condition is that accurate predictions of what the user likes must be made without pressuring the device’s hardware.
- Looper hosts and trains several types of models and decision policies.
- Looper works as a real-time model, unlike most AI platforms that do inference offline in batch mode.
- The platform helps with easy deployment and use of models for use cases where the data size is moderate, and the models are more complicated.
- Looper can work on a wide range of machine learning tasks, like classification, estimation, sequence and value prediction, and ranking and planning using supervised or reinforced learning. The AutoML tools collaborate with the model management infrastructure to select models and hyperparameters to balance the model’s quality, size and inference time. Looper looks at the route from the data sources to the impact of the product that is evaluated and optimised using casual experiments.
- Looper is a declarative AI system which indicates that product engineers have to state the functionality they need for the system to fill in the software implementation. The platform depends on strategy blueprint abstraction, which helps it maintain multiple versions of joint configurations combined for features, labels, models and decision policies. This creates comprehensive optimisation enabling coding-free management and keeps compatibility between different versions. Blueprints use an experiment optimisation system to push vertical optimisations of black-box product metrics.
- The platform helps engineers at Meta to track how the model is used in the tech stack while making tweaks to its modelling framework across aspects like metric selection and policy optimisation. Looper has extended the essential meaning of end-to-end into the software layer to optimise the model architecture and feature selection parameters in a multi-object tradeoff between model quality and computational resources. An engineer can adjust the importance given to different inputs during the real-time decision-making process.
Ease of deployment
Looper aims for quick onboarding, effective deployment and low maintenance of multiple smart strategies that are used to measure positive impact. It leverages AI platforms like PyTorch and Ax for machine learning tasks.
The platform works with models that can be re-trained and deployed fast on a large scale on shared infrastructure, unlike dense AI models for vision, speech and natural language processing, which perform offline inference with batch processing. Looper interprets user-interaction and system-interaction metadata as either labels for supervised learning or rewards for reinforcement learning.
Looper does an evaluation of the results when they are underwhelming, based on the product metrics. In one type of scenario, each decision is checked so that good and bad decisions can be used as examples on which a smart strategy learns using supervised learning. In another type of scenario, product metrics follow long-term objectives that cannot be narrowed down to certain specific decisions. The platform uses Meta’s monitoring system to detect side effects.
Source: Meta AI, where the y-axis is represented by the number of servers
Looper is presently in use by 90+ product teams at Meta to deploy a total of 690 models and make 4 million predictions per second.
Source: Research Paper for the blueprint of creating an end-to-end model lifecycle
Need for optimisation
Optimisation strategies usually get lost within organisations amidst targets and conversion rates. Platforms like Looper can help organisations with analysing their data and creating tailored experiences. Looper makes data and resources easily available, reduces engineering effort and ensures product impact.
- Looper can create a personalised experience based on the user’s history. Like, a product may show shopping-related content only to people who are likely to use it.
- It can improve user utility by displaying ranking order of items. It can offer a personalised feed of candidate items for the user.
- It can prefetch data depending on the most likely prediction made.
- The notifications or prompts are an option for users only if they find them useful.
- It can predict regression tasks using value estimation like latency or memory usage of a data query.