深度学习的10年回顾与展望

摘要：过去10 年深度学习在算法、算力、数据方面获得了长足发展，使人工智能（AI）技术突破商用限制，行业应用场景日益广泛，产业规模持续扩大。在基础模型方面出现了卷积、注意力机制等关键突破；在学习方法方面，强化学习、自监督学习、大模型并行训练等使模型学习能力大大加强。新型AI 计算芯片不断涌现，使计算能效提升百倍。未来10 年，深度学习若要保持可持续的指数增长态势，绿色、高效、安全将成为新的核心要素。空间计算、近似计算等技术有望使AI 芯片效能继续获得百倍提升。一系列生态融合工具的出现将解决目前日趋严峻的生态碎片化问题。AI 安全、可信将成为AI 技术应用的基本要求。

关键词：深度学习；AI 芯片；推理加速；可信AI；开源

Abstract: In the past ten years, deep learning has made great progress in algorithm, computing power, and data, which has enabled artificial intelligence (AI) technology to meet commercial requirements, and has an increasingly wide range of application in various kinds of business, and the scale of the industry has continued to expand. In terms of basic models, there have been key breakthroughs such as convolution and attention mechanisms; in terms of learning methods, technologies such as reinforcement learning, self-supervised learning, and parallel training of large-scale model have greatly enhanced performance. New AI chips continue to emerge, and computing energy efficiency has increased by a hundredfold. In the next ten years, deep learning will maintain a sustainable exponential growth trend, and green, efficient, and safe will become the new core elements. Spatial computing, approximate computing and other technologies are expected to continue to improve the performance of AI chips by a hundredfold. Some integration tools will appear to solve the increasingly severe ecological fragmentation problem. AI security and trustworthiness will become the basic requirements for the application of AI technology.

Keywords: deep learning; AI chip; inference accelerating; trusted AI; open source