Explore Meta FAIR’s Latest AI Innovations: Chameleon, Multi-Token Prediction, and AudioSeal

Meta’s Fundamental AI Research, or FAIR team, has been one of the first research teams to extend the concept of Open Research towards advancing AI. The latest development at Meta FAIR has clearly been stated, whereby it has released six new research artefacts in the public domain with an emphasis on invention, imagination, productivity, and accountability. These are recent models such as Meta Chameleon, Multi-Token Prediction, Meta Joint Audio and Symbolic Conditioning for Temporally Controlled Text-to-Music Generation (JASCO) AraAJ, and AudioSeal for the identification of AI fake speech.

Chameleon

Meta Chameleon is a family of models that can use text and images as inputs and outputs and can return any combination of text and images with the same encoding/decoding architecture for both inputs and outputs. While most current models of late-fusion are applied through the diffusion-based learning process, Meta Chameleon tokenizes both text and images, in a more integrated manner that also allows for easier planning, control, and future adaptation.

Multi Token Prediction

Almost all the contemporary developed LLMs are of the type that simulates the next word sequence, which is quite naïve and resource-consuming. To this end, Meta FAIR has recently introduced a new concept termed multi-token prediction that trains designed language models to guess the next several words. It enhances model performance and training effectiveness to enable control of higher speed in this method.

Also Read: Apple Unveils ‘Apple Intelligence’ AI, Limited Developer Access This Summer

JASCO

Creative businesses such as GenAI now allow individuals to input a set of text prompts and get back music clips generated by generative AI. Of the new features, Meta FAIR’s model is JASCO, which can receive conditioning inputs, for example, particular chords or beats, for enhanced control of generated music outputs. Specifically, by applying layers of information bottleneck, when combined with temporal blurring, JASCO can extract relevant information concerning the given controls and the corresponding inputs, which allows the model to not only understand symbolic conditions but also the conditions from audio input.

AudioSeal

As AI tools become more prevalent, ensuring their responsible use is paramount. To this end, Meta FAIR has developed AudioSeal, an audio watermarking technique designed specifically for the localized detection of AI-generated speech. AudioSeal revamps classical audio watermarking by focusing on detecting AI-generated content rather than steganography. This approach allows for faster and more efficient detection, making it highly suitable for large-scale and real-time applications.

Meta’s latest releases indicate that the team is committed to AI technology developments and is in collaboration with the global AI community. With these innovative models and datasets, META targets inspiring new AI developments.

Meta AI Introduces Pixel Transformers for Enhanced Computer Vision