| AI Technology Base Camp (ID: rgznai100)
Meta At the recent discussion on "Builder Bot with Artificial Intelligence ", Meta showed the latest AI black technology "Builder Bot", and at this meeting, Meta announced an artificial intelligence plan on building the meta universe, including a universal language translation system, a dialogue AI system CAIRaoke, an artificial intelligence recommendation system TorchRec, etc.
Builder Bot
At this discussion, Zuckerberg shared that they are working on a new class of generative AI models that will allow people to describe a world where artificial intelligence automatically generates aspects of it. For the first time, the process of creating a virtual space using the artificial intelligence system Builder Bot is demonstrated.
appears in the form of 3D avatars on the islands of the metaverse, and issues voice commands to create beaches, etc. The system can follow the commands to change the climate and add different scenarios.
In this world, you only need voice commands to get out in seconds. What's more hot is that Builder Bot can also play different music.
At present, the Builder Bot is still under testing and is not open yet, because Meta is not sure where its limitations are. Another part of
At the discussion was speech recognition technology, Meta said it is studying artificial intelligence to allow people to have more natural conversations with voice assistant .
Dialogue AI system CAIRaoke
In the demonstration, CAIRaoke plays a role in a family scene. When people are cooking, the voice assistant will issue a prompt warning that the salt has been added to the pot. The AI assistant also noticed that there was less salt, so he prompted to add more.
In a blog published by Meta, researchers interpret the technology behind CAIRaoke. CAIRaoke reduces the work required to add a new domain. In the specification approach, extending to new domains requires sequential construction and changes to each module before the next module can be reliably trained. This technology improves development and training speeds, allowing us to fine-tune other models with less data.
The company also said it has used the model in its video calling device Portal and looks forward to integrating it into hardware devices with AR and VR.
real-time translator
Meta is developing a universal voice translator designed to create translation software for "all languages in the world". The company has previously set the goal of translating all written languages for its AI systems.
Zuckerberg said: "Being able to communicate with anyone in any language is a superpower that people dream of, and artificial intelligence will achieve this in our lifetime."
Although current translation tools can be well applied to common languages such as English, Mandarin , Spanish , about 20% of the world's population will not speak the languages covered by these systems. Usually, these underserved languages do not have an easy-to-access written text corpus, and these corpus are also necessary to train AI systems, and some languages do not even have standardized writing systems at all.
Meta says it hopes to overcome these challenges by deploying new machine learning technology in two specific areas. The first, called No Language Left Behind, will focus on building AI models that can learn translation languages with fewer training examples. The second is a universal voice translator, designed to build systems that directly translate voice from one language to another in real time without the need for written components as an intermediary (written intermediary is a common technique for many translation applications).
Meta is building a new advanced AI model that can be language-learning from fewer examples, and Meta will use it to enable expert translations in hundreds of languages, from Asturian , Luganda to Urdu .Meta is also building a new universal real-time voice translator to support languages and spoken languages without a standard writing system.
In order to support translation of low-resource languages and create building blocks for translation of more languages in the future, Meta is expanding its own automatic dataset creation technology. One is based on LASER and the other is an open source toolkit, which now contains over 125 languages written in 28 different scripts. Based on LASER, researchers have built systems such as ccMatrix and ccAligned, which can find parallel text in different languages on the Internet. Because there is very little data available for low-resource languages, Meta has created a new training method that enables LASER to focus on specific language subgroups, such as Bantu, and learn from smaller datasets.
This enables LASER to operate efficiently across languages at a large scale, and recently extended LASER to the field of speech processing.
In order to improve the performance of the machine translation model, a large amount of resources were invested to create large-capacity and efficient training models. By increasing the model size and automatic path learning function, different symbols can use different expert abilities. In order to extend the text-based machine translation model to hundreds of languages, Meta extracted another language-centric multilingual translation system that was even better than the best bilingual translation.
In a blog post that announced the news, Meta researchers have not provided a timeline to complete these projects, nor have they provided a major roadmap to achieve their goals. Instead, the company simply emphasizes the possibility of universal language translation.
Meta It is also envisioned that this technology will greatly benefit its global products, further expanding its reach and transforming it into an indispensable communication tool for millions of people. As the blog post writes, universal translation software will become a killer application for future wearable devices, such as the AR glasses that Meta is building, and will also break the boundaries between immersive VR and AR reality spaces (Meta is also building).
Conclusion
Indeed, the advances in machine learning in recent years have greatly improved the speed and accuracy of machine translation. From Google to Apple , many large tech companies are now providing users with free AI translation tools for work and travel, which undoubtedly provides immeasurable benefits worldwide.
But there will be some technical problems. Critics point out that machine translation ignores some crucial nuances when humans speak, there will be some bias in the output, and point out some strange, unintentional syntax errors that can only occur in computers. For those who use non-common languages, they are worried about losing their grasp of their language and culture if the translation is entirely controlled by large tech companies.
Although Meta has long been eager to enter the global market, the company's own products still tend to countries where it provides most of its revenue.
Although a universal translator is a very good wish, Meta needs not only to prove its technology and capabilities, but also to be a company, it also needs to study its applications fairly.
At this discussion meeting, the information received was still full. What do you think about this?
Reference link:
https://www.theverge.com/2022/2/23/22947368/meta-facebook-ai-universal-speech-translation-projecthttml2
https://ai.facebook.com/blog/teaching -ai-to-translate-100s-of-spoken-and-writen-languages-in-real-time
https://www.reuters.com/technology/metaverse-event-metas-zuckerberg-unveils-work-improve-how-humans-chat-ai-2022-02-23/
