Automatic Deep Neural Network Design by a Novel Stochastic Search Process

机器学习，特别是深度学习取得的令人鼓舞的进展，为不同的应用（如图像分类、对象检测、语音识别，甚至医疗应用）带来了更好的建模准确性。然而，这些模型的设计周期仍然依赖于人工，耗时，并且非常复杂的过程，这需要高水平的专业知识。 因此，使用这种耗时的手动过程来设计和开发专用且高效的神经网络模型是非常困难的，这减缓了它们在许多现实世界应用中的采用。这样的缺点变得更加严重，特别是当需要考虑可用硬件的计算能力和存储器的固有限制时。 神经结构搜索（NAS）方法提供了解决这些挑战的解决方案。NAS自动设计模型的底层图结构，通过搜索解空间来找到最优结构。尽管NAS方法具有很好的潜力，但它们仍处于起步阶段，计算复杂且仅限于特定应用。 拟议的研究计划旨在缓解这些挑战。将研究以下目标： i）研究随机分支预测策略以降低搜索空间架构的维度并相应地降低计算复杂度。 ii）描述基于新随机过程的深度学习模型的计算图内部的复杂信息。 iii）在设计过程中融入硬件特征，并开发一种新方法来为深度学习模型生成特定于硬件的优化计算图。 vi）通过考虑环境特性的新方法设计和开发更具适应性的机器学习系统。 拟议研究计划的结果将提高深度神经网络在自动驾驶、消费电子和医疗保健等不同行业的适用性。由于其自动化的性质，所提出的方法提供了显着减少的设计周期，这有利于开发应用程序的基础上，这种技术的最终用户的专业知识水平低得多。这种新方法还将促进加拿大作为人工智能领域的先驱国家在不同行业中采用深度学习模型。在拟议的研究计划中开发的知识和技术将通过与微软，英特尔和DarwinAI等工业合作伙伴的积极合作进行转移。这确保了深度神经网络的有效开发和部署，以用于现实世界的机器学习应用。 在该计划中接受培训的HQP将成为人工智能和机器学习领域的专家，使他们在目前缺乏专家的行业和学术界中处于人工智能领导角色的有利地位。

The promising advances made in machine learning and especially deep learning have led to better modeling accuracy for different applications such as image classification, object detection, speech recognition, and even medical applications. However, the design cycle of these models still rely on manual, time-consuming, and very complex process, which requires high-level expertise. Consequently, the design and development of specialized and efficient neural network models using such time-consuming manual processes are very difficult slowing down their adoption in many real-world applications. Such shortcoming becomes even more critical especially when it is necessary to take into account the inherent limitations in the computational power and the memory of the available hardware. Neural architecture search (NAS) methods offer a solution to address these challenges. NAS designs the underlying graph structure of models automatically by heuristically searching the solution space to find the most optimal structure. Despite their promising potentials, NAS approaches are still in their infancy, computationally complex and limited to specific applications. The proposed research program aims to mitigate these challenges. The following objectives will be investigated: i) Investigate stochastic branch prediction strategies to reduce the dimensionality of the search space architecture and accordingly the computational complexity. ii) Characterize the complex information inside the computational graph of a deep learning model based on the new stochastic process. iii) Incorporate hardware characteristics in the design process and develop a novel approach to generate hardware-specific optimized computational graphs for the deep learning models. vi) Design and develop more adaptive machine learning systems via the novel method which take environmental properties into account. The results of the proposed research program will increase the applicability of deep neural networks in different industries such as autonomous driving, consumer electronics and healthcare. Due to its automatic nature, the proposed approach offers dramatically reduced design cycles, which facilitates developing applications based on this technology with much less level of end-user expertise. This new approach will also boost the adoption of deep learning models in different industries in Canada as a pioneer country in the field of AI. The knowledge and technologies developed in the proposed research program will be transferred through active collaborations with industrial partners such as Microsoft, Intel and DarwinAI. This ensures the efficient development and deployment of deep neural networks for real-world Machine Learning applications. HQP trained in this program will become experts in the field of artificial intelligence and machine learning, putting them in a strong position for AI leadership roles in industry as well as in academia, where there is currently a shortage of experts.