MIT doctoral thesis | robust and reliable intelligent system using neural symbol learning

Thesis link ：https://dspace.mit.edu/bitstream/handle/1721.1/143249/Inala-jinala-PhD-EECS-2022-thesis.pdf?sequence=1&isAllowed=y

This paper shows that , Looking at intelligent systems from the perspective of neural symbolic models has several advantages over traditional deep learning methods . Neural symbolic model contains symbolic procedural structure , Like a cycle 、 Conditioned and continuous neural components . The symbolic part makes the model interpretable 、 Generalization and robustness , The neural part deals with the complexity of intelligent systems . To be specific , This paper presents two kinds of neural symbolic models —— State machines and neural symbols transformers, The autonomous system based on reinforcement learning and multi robot system are taken as examples to evaluate them . These case studies show , The neural symbolic model of learning is human readable , It can be extrapolated to invisible scenes , And can deal with the robust goals in the specification . In order to effectively learn these neural symbolic models , We introduce a neural symbol learning algorithm using the latest technology of machine learning and program synthesis .