Adiabatic quantum computing for deep learning with boltzmann machines

使用玻尔兹曼机进行深度学习的绝热量子计算

• 批准号: 447518-2013
• 负责人: Bengio, Yoshua
• 金额: $ 14.17万
• 依托单位: Université de Montréal
• 依托单位国家: 加拿大
• 项目类别: Strategic Projects - Group
• 财政年份: 2014
• 资助国家: 加拿大
• 起止时间: 2014-01-01 至 2015-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=554573
• 关键词: Adiabatic; quantum; computing; deep; learning

Boltzmann machines are at the heart of the rapidly growing success of the deep learning approach to machine learning, but they require computationally expensive Monte-Carlo Markov Chains (MCMC) during training. On the other hand, D-Wave has built analog computing hardware that exploits quantum tunneling effects to efficiently performing the type of computation required to train Boltzmann machines. Replacing the software MCMC process by the more efficient hardware computation based on D-wave's tchnology has the potential to have profound impact on the scale of problems that are practically addressed with Boltzmann machines and could have wide-ranging implications for artificial intelligence research and applications. This project investigates the limitations of the current D-Wave hardware from the point of view of implementing deep Boltzmann machines, how to bypass these limitations, and how to take advantage of the strengths of this hardware to learn higher-level representations for which traditional software-based inference is challenging.

博尔兹曼机器是深度学习机器学习方法迅速取得成功的核心，但它们在训练期间需要计算昂贵的蒙特卡洛马尔可夫链(MCMC)。另一方面，D-Wave已经建立了模拟计算硬件，利用量子隧道效应来高效地执行训练Boltzmann机器所需的计算类型。用基于D-WAVE技术的更高效的硬件计算取代软件MCMC过程，可能会对用Boltzmann机器实际解决的问题的规模产生深远影响，并可能对人工智能研究和应用产生广泛的影响。本项目从实现深度Boltzmann机器的角度研究当前D-Wave硬件的限制，如何绕过这些限制，以及如何利用该硬件的优势来学习传统的基于软件的推理所具有的更高级别的表示。