Researchers at Facebook parent’s Meta have trained a single AI model capable of processing speech, images, and text in the hope that these so-called multi-modal systems will power the company’s augmented reality and metaverse products.
The model, known as data2vec, can perform different tasks. Given an audio snippet, it can recognize speech. If it’s fed an image, it can classify objects. And when faced with text, it can check the grammar or analyse the writing’s tone and emotions.
AI algorithms are typically trained on one type of data, though data2vec is trained on three different modalities. It still, however, processes each form, whether its speech, images, and text, separately.
Meta believes these multi-modal models will help computers be more adaptable to blend physical and digital environments into one. “People experience the world through a combination of sight, sound and words, and systems like this could one day understand the world the way we do,” Meta CEO Mark Zuckerberg said in a statement to El Reg.
“This will all eventually get built into AR glasses with an AI assistant so, for example, it could help you cook dinner, noticing if you miss an ingredient, prompting you to turn down the heat, or more complex tasks.”
Data2vec is a transformer-based neural network and uses self-supervised learning to learn common patterns in audio, computer vision, and natural language processing. The model learns to operate with different types of data by learning how to predict how the representation of data it’s given; it knows it has to guess the next group of pixels when given an image, or the next speech utterance in audio, or fill in the words in a sentence.
The researchers used a mix of 16 Nvidia V100 and A100 GPUs to train data2vec on 960 hours of speech audio, millions of words from books and Wikipedia pages, and images from ImageNet-1K.
“We train separate models for each modality but the process through which the models learn is identical,” Alexei Baevski, a research engineer at Meta AI told The Register.
“We hope that it will enable future work to build high performing self-supervised models that combine modalities and are more effective than specialized models. Different modalities can add additional information to the same piece of content – for example body language from video, prosodic information from audio, and text can combine into a richer representation of a dialog. The algorithms that currently try to combine multi-modal information exist but they do not yet perform well enough to replace specialized algorithms and we hope our work will help change that.”
Baevski said in the future multi-modal systems could incorporate a larger range of data to model concepts like smell, 3D objects, or videos. He referred back to the idea of AR glasses helping wearers cook.
“Imagine having a model that has been trained on recordings of thousands of hours of cooking activity from various restaurants and chefs. Then, when you are cooking in a kitchen wearing your AR glasses that have access to this model, it’s able to overlay visual cues for what you need to do next, point out potential mistakes, or explain how adding a particular ingredient will affect the taste of your dish,” he told us.
Previous research on multi-modal systems have shown they can be prone to easy adversarial attacks. OpenAI’s CLIP model, for example, trained on images and text will identify an image of an apple incorrectly as an iPod if the word “iPod” are in the picture. It’s unclear, however, if data2vec suffers from similar weaknesses.
“We have not specifically analyzed how our models will react to adversarial examples but since our current models are trained separately for each modality, we believe that existing research on adversarial attack analysis for each modality would be applicable to our work as well,” Baevski said.
“In the future, we hope to use our work to enable high performance algorithms that combine modalities in one model and we plan to study how susceptible they are to adversarial attacks.”
When the researchers tested data2vec, it outperformed some top models that had been trained on a specific data type only on different types of tasks. The preliminary results are described in a paper [PDF], and the code has been published on GitHub.
“Data2vec demonstrates that the same self-supervised algorithm can work well in different modalities — and often better than the best existing algorithms,” the researchers explained in a blog post this week.
“This paves the way for more general self-supervised learning and brings us closer to a world where AI might use videos, articles, and audio recordings to learn about complicated subjects, such as the game of soccer or different ways to bake bread. We also hope data2vec will bring us closer to a world where computers need very little labeled data in order to accomplish tasks.” ®