Diversified Perspectives: Modes of Attention in Algorithmic Information Processing 多元化视角：算法信息处理中的注意力模式

Summary 摘要

Artificial intelligence took a significant leap toward human-like multisensory and associative capabilities with OpenAI’s ChatGPT model in 2022 when it implemented a self-attention mechanism. As AI technologies progress rapidly, the concept of attention offers a novel opportunity to re-envision information in machine learning models. This paper explores attention as a perceptual phenomenon with unique aesthetics and emphasizes its role in shaping percepts and the resulting information. Large language models that primarily rely on text can be understood as exhibiting object-based attention. The paper goes beyond this limited view, advocating for diversified perceptual modes in machines. By adopting multiple perceptual strategies, machines can help humans overcome their biases, inborn ignorance, and even hatred, thus fostering more humane/than human connections.

2022 年，人工智能在采用自我注意力机制时，通过采用 OpenAI 的 ChatGPT 模型，在类似人类的多感官并用能力和联想性思维能力方面，取得了重大飞跃。随着 AI 技术的快速发展，注意力理念具有了新的发展潜力，能够在机器学习模型中重新构想信息。本文将注意力作为一种具有独特美学的感知现象进行探讨，并强调了注意力在塑造各种感知及其产生的信息中的作用。关于那些主要依赖于文本的大型语言模型，可以将其理解为展示基于对象的注意力。本文超越了这种局限的观点，倡导机器中的多样化感知模式。通过采用多种感知策略，机器可以帮助人类克服偏好、与生俱来的愚昧，乃至仇恨，从而培养更人性化的、超越人本身的各种联系。