Facebook AI small sample learning technology breakthrough FSL, learning humanoid artificial intelligence step forward

Harmful content can evolve rapidly — whether driven by current events or by people looking for new ways to evade our systems — and it is vital that AI systems evolve with it.However, it often takes months for an AI to learn how to look for it, to collect and tag the thousands, if not millions, of instances necessary for every AI system to discover a new type of content.To overcome this bottleneck, we built and deployed a new AI technology called Few-Shot Learner (FSL) that can take action on new or changing, harmful content types in weeks, not months.Not only can it be used in more than 100 languages, it can also learn from all kinds of data, such as images and text.It can enhance existing AI models deployed to detect other types of harmful content.This new AI system uses a relatively new approach called “fee-shot learning,” in which models learn new tasks through a large, general understanding, followed by a small, in some cases zero, labeled sample.If a traditional system is akin to a line that can catch one type of fish, FSL is an extra net that can catch others.Recent technological breakthroughs, such as our self-supervised learning technology and new super-efficient infrastructures, have moved the field from traditional, customised AI systems to larger, more integrated, more versatile systems that rely less on tagged data.First, it trains on billions of samples of common and open source languages.Next, we trained the AI system with the content and boundary content of policy violations flagged over the years.Finally, the compressed text explaining the new strategy is trained.Unlike previous systems that relied on marker data for pattern matching, FSL is pre-trained based on common language as well as violation policy and boundary content languages, so it can implicitly learn policy text.We have tested FSL on some relatively new events.One recent task, for example, was to identify content that shared misleading or sensational information in a way likely to prevent the delivery of COVID-19 vaccines (e.g., “vaccines or DNA modifiers?).In a separate task, new AI systems improved on existing classifiers, flagging content close to inciting violence (e.g., “Does that guy need all his teeth?).Traditional methods may miss such incendiary posts because less-marked samples use the language of DNA to create vaccine scares, or cite teeth to suggest violence.To measure the performance of this model, we developed A standard offline and online A/B testing protocol.In these tests, we looked at the prevalence of harmful content — the percentage of people viewing offending content — before and after we introduced FSL on Facebook and Instagram.Meta AI Few-shot Learner can accurately detect posts that are missed in traditional systems and help reduce the prevalence of such harmful content.It stops the proliferation of potentially harmful content on our platform by proactively detecting it.We also found that FSL, combined with existing classifiers, helped reduce the proliferation of other harmful content such as hate speech.We are also doing more experiments to improve classifiers that can benefit from more tagged data. For example, we will continue to test these new violation content patterns in countries with languages that do not have a lot of tagged training data.These are, of course, prototypes of intelligent, general-purpose artificial intelligence.There is still a long way to go before AI can read dozens of pages of strategy text and immediately understand exactly how it will be implemented.We have been pushing ai technology forward and deploying it as quickly as possible to better serve our communities, and we believe FSL will be a very promising development.Small sample learning under the hood Few-Shot Learner is a large-scale, multimodal, multilingual, zero sample or small sample model that understands joint strategies and content and can generalize about integrity issues without adjusting the model.We are actively conducting research to train models that use simple policy statements rather than hundreds of labeled samples.Our new system works in three different scenarios, each requiring a sample of different levels of markup: zero sample: policy description with no sample.Small sample with demonstration: Strategy descriptions with a small number of samples (less than 50).Small samples with fine tuning: Machine learning developers can fine-tune the basic model of FSL, and the number of training samples is small.The overall input of FSL consists of three parts.First, building on our previous work using Whole Post Integrity Embeddings (WPIE), it learns multimodal information from entire posts, including text, images, urls, and more.Second, it analyzes policy-related information, such as the definition of a policy, or tagged samples that indicate whether a particular post violates that policy definition.Third, we also take additional marker samples as demonstrations, if any.As part of our new approach, so-called Entailment Fee-shot Learning, the key idea is to convert category labels into natural language sentences that can be used to describe labels and determine whether the example contains label descriptions.For example, we can restate an obvious emotional categorization input and tag pair.[X: “I love your race. JK. You all deserve to die.” Y: Positive] as a sample of the text contained below: [X: I love your race.We compare our proposed method with some of the most advanced small sample learning methods currently available.After a series of systematic evaluations, we found that our approach was 55% better (12% on average) than the various state-of-the-art small-sample learning methods.Here: https://arxiv.org/pdf/2104.14690.pdf, you can read our full details of research papers.Bridging the gap between policy creation and machine learn-driven automated execution We believe that OVER time, FSL can improve the performance of all of our integrity AI systems, allowing them to leverage a single, shared knowledge base and trunk to handle many different types of violations.But it can also help bridge the gap between human insight and classifier advances in policy, tagging, and survey workflows.FSL can be used to detect a new set of possible policy violations and to understand the rationality and validity of the proposed definitions.It casts a wider net to surface more types of “near” content violations that the strategy team should be aware of when deciding or developing sizing guidelines for training annotators for new classifiers and human censors to help keep our platform secure.Because it scales quickly, the time from policy development to implementation will be reduced by several orders of magnitude.Being able to quickly start enforcing content types of training data without a lot of markup is a big step forward in moving toward human-like ai that can learn more effectively and will help make our systems more flexible and responsive to emerging challenges.Small sample learning and zero sample learning is one of many cutting-edge AI areas where we have been making significant research investments.And we see no sign of a slowdown in research into production pipelines.We are working on important open research questions that not only understand content, but also reason from cultural, behavioural and conversational contexts.There is still a lot of work to be done, but these early production results are a milestone in the transition to a smarter, more versatile AI system that can perform multiple tasks at the same time.Our long-term goal is to achieve human-like learning flexibility and efficiency, making our integrity systems faster, easier to train, and better able to process new information.Teachable AI systems like Few-Shot Learner can dramatically increase our agility in our ability to detect and adapt to new situations.By faster and more accurate identification of evolving harmful content, FSL promises to be a key technology to help us continue to evolve and address harmful content on our platform.The original link: https://ai.facebook.com/blog/harmful-content-can-evolve-quickly-our-new-ai-system-adapts-to-tackle-it

Leave a Reply

Your email address will not be published.