Activities
Divisions
Performances
Activities
Divisions
Performances
At last, Apple has rolled out MM1, their AI system designed for generating text and images. The AI system is also capable of making contextual predictions due to its extensive pre-training in multiple modes.
Following a period of gossip and conjecture about their future AI ventures and multimodal AI models, Apple's research team has created a group of expansive multimodal language models named MM1. These models have the capability to handle and produce both textual and visual information, as detailed in a research paper demonstrated last week.
The research conducted in Apple's laboratories was focused on developing efficient large language models that are multimodal (MLLMs). This was achieved through meticulous examination and removal of different architectural elements, data sources, and training methods.
The study revealed that the clarity of the image and the effectiveness of the visual encoder greatly affected the performance of the model. However, the particular technique used to merge visual and text data was not as significant.
They also found out that a deliberate combination of various data forms was essential. Few-shot learning was aided by intermixed image-text documents, traditional captioned images enhanced zero-shot performance, and the inclusion of text-only data sustained robust language comprehension skills.
MM1 has the capability to make predictions within a given context due to its extensive multimodal pre-training. This qualifies MM1 to tally items and adhere to tailored formatting, reference sections of images and carry out Optical Character Recognition (OCR), exhibit common sense and familiarity with daily objects, as well as conduct simple mathematical operations.
From the knowledge gained, the group created the MM1 model family, which includes models with parameters ranging from three billion to 30 billion. This includes dense and mixture-of-experts types. When the training was expanded, MM1 reached top-tier performance on several multimodal benchmarks during the pre-training phase.
After additional tuning using a curated dataset of 1 million examples, the ultimate MM1 models showed competitive results across 12 multimodal tasks, including visual question answering and captioning. Importantly, MM1 was able to carry out multi-image reasoning and few-shot learning, crucial functions made possible by the team's meticulous multimodal pre-training strategy.
This study expands on earlier investigations into fields such as CLIP that learn visual representation from naturally occurring language supervision, and autoregressive models like GPT designed for text creation. Nonetheless, it is among the pioneering comprehensive researches that are particularly concentrated on large-scale multimodal pre-training.
The investigators are optimistic that their findings will speed up advancements, considering that Apple is allegedly discussing incorporating Google's Gemini generative AI models into forthcoming iPhone software.
Look for us on YouTube
Top Programs
Associated Reports
Microsoft recruits Mustafa Suleyman, cofounder of DeepMind, to head its new consumer AI division
South Korean chip makers Samsung and Rebellions aim to surpass NVIDIA
The upcoming Apple Watch model is set to include a much-anticipated feature
NVIDIA introduces its new AI powerhouse, the Blackwell B200 superchip, boasting 30 times the power of their present top-tier H100
Microsoft has brought on board Mustafa Suleyman, co-founder of DeepMind, to spearhead their new consumer AI squad
Samsung and Rebellions, South Korean chip producers, have set their sights on outperforming NVIDIA
The eagerly awaited feature is finally going to be included in the next Apple Watch edition
NVIDIA presents their new Blackwell B200 AI superchip, alleging it to be 30 times more potent than their existing flagship model, the H100
Available on YouTube
All rights reserved by Firstpost. Copyright remains effective until 2024.
+ There are no comments
Add yours