Building the next generation of intelligent machines with cooperative deep learning.

现代机器学习技术（如深度学习）的进步和普遍肯定，有助于在几个领域取得前所未有的表现，包括计算机视觉、自然语言处理和语音处理等。然而，大多数现有的深度学习系统只能解决一个或很少的问题。这种“孤立的”智能范式阻碍了当前技术对我们周围复杂世界的深入建模。相反，人类大脑显示，跨专门模块（又名“脑叶”）的合作对于解决高级和抽象问题至关重要。该研究计划旨在研究解决不同任务的人工神经网络如何合作解决具有挑战性的问题。这项研究将主要研究神经网络如何相互作用和通信，而不是专注于独立改进神经模块。目标是通过引入一种新颖的合作深度学习范式来扩展当前技术的功能，在这种范式中，神经模块不再实现“孤立”的智能，而是被安排在一个网络中，在这个网络中，它们逐渐学习如何合作、交流和交互。这项研究的一个自然应用场景是对话式人工智能。开发能够有效地与人类交谈的机器确实是一个具有挑战性的问题，它需要结合复杂的技术，例如关键字识别、语音增强、语音识别、口语理解和对话系统，等等。所提出的计划将通过逐步增加复杂性来逐步实现这一目标。研究将从少量模块开始（例如，2或3个神经网络）。团队稍后将扩展该范例以获得更复杂的解决方案。这项研究将以公开、透明和可重复的方式进行，以最大限度地为社区带来利益。这项研究的一个方便的开源工具包是SpeechBrain，它的开发是由主要申请人领导的一个大型国际网络进行的。这项发现资助计划的潜在成果是定义一种新的合作深度学习范式。这种方法最终可以实现人工智能代理网络的发展，其中神经生态系统的每个成员都有助于解决具有挑战性的问题。当电脑联网时，人类经历了一场革命。我们可以通过连接更多智能机器（如神经网络）实现另一个突破。该项目提供了令人兴奋的机会，培养4名高素质人才（HQP）（2名博士和2名硕士）和4名本科生。hqp将培养深度学习和对话式人工智能方面的专业知识，这些都是就业市场急需的能力。在这一领域培训HQP对于加拿大雄心勃勃的人工智能发展计划至关重要。

The progressive and pervasive affirmation of modern machine learning techniques such as deep learning has contributed to achieving unprecedented performance in several fields, including computer vision, natural language processing, and speech processing, to name a few. Most of the existing deep learning systems, however, solve one or very few problems only. This "isolated" intelligence paradigm prevents current technology from deeply modeling the complex world around us. The human brain, instead, shows that cooperation across specialized modules (aka "lobes") is essential to solving high-level and abstract problems. This research program proposes to study how artificial neural networks solving different tasks can cooperate to address challenging problems. Rather than focusing on improving neural modules independently, this research will mainly investigate how neural networks interact and communicate. The goal is to expand the functionalities of current technology by introducing a novel cooperative deep learning paradigm, where neural modules are not anymore implementing "isolated" intelligence but are arranged in a network where they progressively learn how to cooperate, communicate, and interact. A natural application scenario for this research is conversational AI. The development of machines able to effectively converse with humans is indeed a challenging problem that requires combining complex technologies, such as keyword spotting, speech enhancement, speech recognition, spoken language understanding, and dialog systems, to name a few. The proposed plan will accomplish this goal gradually by adding complexity progressively. The research will start with a small number of modules (e.g., 2 or 3 neural networks). The team will later scale the paradigm up to more complex solutions. This research will be conducted in an open, transparent, and reproducible way to maximize the benefits for the community. A convenient open-source toolkit for this research is SpeechBrain, whose development is conducted by a large international network led by the principal applicant. The potential outcome of this discovery grant program is the definition of a novel cooperative deep learning paradigm. This approach can eventually enable the development of networks of artificial intelligence agents where each member of the neural ecosystems contributes to solving challenging problems. Humanity experienced a revolution when computers were connected. We can achieve another breakthrough by interconnecting more intelligent machines such as neural networks. This program offers the exciting opportunity to train 4 Highly Qualified Personnel (HQP) (2 PhDs and 2 Master students) and 4 undergrad trainees. The HQPs will develop expertise in deep learning and conversational AI, which are competencies highly demanded in the job markets. Training HQP in this field is of paramount importance for the ambitious AI development plan of Canada.