FRG: Collaborative Research: Robust, Efficient, and Private Deep Learning Algorithms

该项目开发强大，准确，高效的下一代深度学习算法，具有数据隐私和理论保证，可解决具有挑战性的人工智能（AI）问题。该方法将具有理论保证的对抗性攻击的鲁棒性。 该项目将把人工智能在性能和隐私方面的收益推向移动的设备。广泛的应用包括自动驾驶、药物和材料发现、医疗规划、国防、边缘隐私保护机器学习、联邦学习以及区块链。此外，开发的工具将显着扩展现有的科学模拟到超大规模和高维场景。该项目将每年在每个校区资助一名研究生。我们的方法是基于现代偏微分方程和优化算法和理论的理论原则，以实现值得信赖的深度学习。 该项目涉及新的算法和理论技术，以解决高维非凸，非光滑AI设置中的图形表示。特别是，该项目将研究（1）开发对抗性鲁棒深度学习算法及其理论基础;（2）利用新的随机优化和原则性神经网络单元设计辅助神经架构搜索来提高深度学习的准确性;（3）通过算法和硬件协同设计来推进深度神经网络压缩;（4）设计新的数据隐私机制，在效用和隐私之间进行最佳权衡;（5）发明新的定量分析工具，破解深度学习理论挑战的奥秘;（6）量化复杂深度学习算法的不确定性。该项目通过应用数学、计算机科学、数据科学和普通生物学、物理学、生物学和生物学等领域的合作教育和研究活动，在加州大学欧文分校、加州大学洛杉矶分校和犹他州大学培养了一批多样化的研究生和本科生。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查进行评估，被认为值得支持的搜索.

This project develops robust, accurate, and efficient next-generation deep learning algorithms with data privacy and theoretical guarantees for solving challenging artificial intelligence (AI) problems. The methods will have robustness to adversarial attacks with theoretical guarantees. The project will push artificial intelligence gains in performance and privacy to mobile devices. A broad range of applications includes autonomous driving, drug and material discovery, medical treatment planning, national defense, privacy-preserving machine learning at the edge, federated learning, and also blockchain. Moreover, the developed tools will significantly scale the existing scientific simulations to ultra-large scale and high-dimensional scenarios. This project will partially support one graduate student per year at each campus.Our approach toward trustworthy deep learning is theoretically principled by modern partial differential equations and optimization algorithms and theories. The project involves new algorithmic and theoretical techniques to tackle graph representation in high-dimensional non-convex, non-smooth AI settings. In particular, the project will study (1) developing adversarial robust deep learning algorithms and their theoretical foundations; (2) improving the accuracy of deep learning leveraging new stochastic optimization and principled neural network unit design assisted neural architecture search; (3) advancing deep neural networks compression with algorithms and hardware co-design; (4) designing new data privacy mechanisms to optimally tradeoff between utility and privacy; (5) inventing new quantitative analysis tools to decipher the mysteries of deep learning theoretical challenges; (6) quantifying uncertainties of sophisticated deep learning algorithms. The project trains a diverse body of graduate and undergraduate students at UC Irvine, UCLA, and University of Utah through collaborative education and research activities in applied mathematics, computer science, data science, and general biological, physical, and sociological disciplines.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.