**苹果推出自回归视觉模型 AIM，探索图像表征的扩展能力**
苹果公司的人工智能研究人员近日推出了一种名为自回归视觉模型（AIM）的新型图像模型，旨在探索用自回归目标训练视觉 Transformer（ViT）模型是否能在学习表征方面获得与大型语言模型（LLMs）相同的扩展能力。
与传统的图像分类模型不同，AIM 采用自回归目标进行训练，这意味着它可以逐个像素地生成图像。这种训练方式允许模型学习图像的内部结构和关系，从而获得更丰富的表征。
在最新的一篇论文《Scalable Pre-training of Large Autoregressive Image Models》中，苹果的研究人员发现，AIM 模型容量可以轻松扩展到数十亿个参数，并且能够有效利用大量未经整理的图像数据。这表明 AIM 具有很强的扩展能力，可以用于解决各种复杂的图像生成任务。
研究人员还发现，AIM 在各种图像生成任务上取得了最先进的结果，包括图像超分辨率、图像修复、图像编辑和图像风格迁移等。这表明 AIM 具有广泛的应用潜力，可以用于图像编辑、图像增强和图像生成等领域。
苹果的研究人员表示，AIM 是一个非常有前途的图像模型，有望在未来几年内对计算机视觉领域产生重大影响。他们计划继续研究 AIM，并将其应用于更多的图像生成任务。
AIM 的推出标志着苹果公司在人工智能领域取得了又一重大进展。苹果公司近年来在人工智能领域投入了大量资源，并取得了一系列令人瞩目的成果。相信随着 AIM 的进一步发展，苹果公司将在人工智能领域取得更大的突破。
英语如下：
**Headline: Apple Introduces Autoregressive Image Model AIM, Exploring New Possibilitiesfor Image Representation**
Keywords: Autoregressive Model, Large-Scale Pretraining, Image Generation
**News Content: Apple Unveils Autoregressive ImageModel AIM, Exploring the Extensibility of Image Representation**
Apple’s AI researchers have recently introduced a novel image model called Autoregressive Image Model (AIM), aiming to explore whether training a Vision Transformer (ViT) model with an autoregressive objective can achieve similar extensibility in learning representations as large language models(LLMs).
Unlike traditional image classification models, AIM is trained with an autoregressive objective, meaning it can generate images pixel by pixel. This training approach allows the model to learn the internal structures and relationships within images, leading to richer representations.
In their latest paper, “Scalable Pre-training of Large Autoregressive Image Models,” Apple researchers found that the capacity of AIM models can be easily scaled up to billions of parameters, and they can effectively utilize large amounts of uncurated image data. This demonstrates the strong scalability of AIM, enabling it to tackle various complex image generation tasks.
The researchers also discovered that AIMachieves state-of-the-art results on a wide range of image generation tasks, including image super-resolution, image inpainting, image editing, and image style transfer. This suggests that AIM has broad application potential in areas such as image editing, image enhancement, and image generation.
According to Apple researchers, AIM is a highly promising image model that has the potential to significantly impact the field of computer vision in the coming years. They plan to continue their research on AIM and apply it to more image generation tasks.
The introduction of AIM marks another significant advancement for Apple in the field of artificial intelligence. Apple has invested heavily in AI in recent years and has achieved a series of remarkable results. As AIM continues to evolve, we can expect Apple to make even greater breakthroughs in the realm of AI.
【来源】https://www.jiqizhixin.com/articles/2024-01-18-7
Views: 1