Apple researchers have shared their work on building a multimodal artificial intelligence (AI) large language model (LLM), in a pre-print paper. Published on an online portal on March 14, the paper highlights how it was able to achieve the advanced capabilities of multimodality and make the foundation model train on both text-only data as well as images. The new advancements in AI for the Cupertino-based tech giant come following CEO Tim Cook’s remarks made during the company’s earning calls where he said that AI features could arrive later this year.
The pre-print version of the research paper has been published on arXiv, an open-access online repository of scholarly papers. However, the papers posted here are not peer-reviewed. While the paper itself does not mention Apple, most of the researchers mentioned are affiliated with the company’s machine learning (ML) division, leading to the belief that the project is also affiliated with the iPhone maker.
As per the researchers, they are working on MM1, a family of multimodal models containing up to 30 billion parameters. Calling it a “performant multimodal LLM (MLLM), the authors of the paper highlighted that image encoders, the vision language connector, and other architecture components and data choices were made to create the AI model which is capable of understanding both text as well as image-based inputs.
Giving an example, the paper stated, “We demonstrate that for large-scale multimodal pre-training using a careful mix of image-caption, interleaved image-text, and text-only data is crucial for achieving state-of-the-art (SOTA) few-shot results across multiple benchmarks, compared to other published pre-training results.”
To break it down, the AI model is currently in the pre-training phase, which means it is not trained enough to give the desired outputs. This is the stage when the algorithm and the AI architecture are used to design the workflow of the model and how it processes data, eventually. The team of Apple researchers were able to add computer vision to the model using image encoders and a vision language connector. Then, when testing with a mix of just images, image and text, and text-only data set, the team found that the results were competitive compared to existing models at the same stage.
While the breakthrough is significant, this research paper is not enough to ascertain that a multimodal AI chatbot will be added to Apple’s operating system. At this stage, it is difficult to even say whether the AI model is multimodal while taking inputs or in giving output as well (whether it can generate AI images or not). But if the results are confirmed to be consistent after peer review, it can be said that the tech giant has taken another big step towards building a native generative AI foundation model.
