SHF: Small: Synthesizing Mixed Discrete/Continuous Programs with the Neurosymbolic Librarian

SHF：小型：与神经符号图书馆员综合混合离散/连续程序

• 批准号: 2310350
• 负责人: Kevin Ellis
• 金额: $ 60万
• 依托单位: Cornell University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2023
• 资助国家: 美国
• 起止时间: 2023-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2310350&HistoricalAwards=false
• 关键词: SHF; Small; Synthesizing; Mixed; Discrete

This project concerns automatically generating computer programs that can use neural networks and other machine learning models as subroutines. Programs like these are important because they form the cornerstone of modern machine learning systems, and also because symbolic programs and neural networks are complementary in their abilities, so such systems could learn to solve many diverse problems using a mixture of neural networks and symbolic code. However such programs are difficult to automatically generate or synthesize. The project’s novelties are new strategies that makes it much easier to generate such programs by using machine learning. The project's impact is a step toward systems that could learn to solve new problems using a mixture of neural networks and symbolic code, as well as a step toward Artificial Intelligence (AI) systems that could assist the development of further AI systems.From a technical perspective the project presents a way of jointly generating symbolic code and neural network weights both using gradient descent, by relaxing the discrete space of symbolic code into a continuous form. Because convergence of this relaxation can be difficult, the investigator proposes learning to generate the code in a multitask setting, which allows learning across many problems to aid convergence. The results will be showcased on generating 3-dimensional graphics programs, mixing implicit neural representations of geometry with discrete graphics primitives, as well as a few-shot learning domain.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.

该项目涉及自动生成计算机程序，这些程序可以使用神经网络和其他机器学习模型作为子程序。像这样的程序很重要，因为它们构成了现代机器学习系统的基石，也因为符号程序和神经网络在能力上是互补的，所以这样的系统可以学习使用神经网络和符号代码的混合来解决许多不同的问题。然而，这样的程序很难自动生成或合成。该项目的新颖之处在于新的策略，使使用机器学习生成此类程序变得更加容易。该项目的影响是朝着能够使用神经网络和符号代码的混合来学习解决新问题的系统迈出了一步，同时也朝着能够帮助进一步开发人工智能系统的人工智能（AI）系统迈出了一步。从技术角度来看，该项目提出了一种使用梯度下降联合生成符号代码和神经网络权重的方法，通过将符号代码的离散空间放松为连续形式。由于这种放松的收敛可能很困难，研究人员建议在多任务设置中学习生成代码，这允许在许多问题上学习以帮助收敛。结果将展示在生成三维图形程序，混合隐式神经几何表示与离散图形原语，以及几个镜头的学习域。这个奖项反映了NSF的法定使命，并已被认为是值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估的支持。