One of the most challenging question types in VQA is when answering the\nquestion requires outside knowledge not present in the image. In this work we\nstudy open-domain knowledge, the setting when the knowledge required to answer\na question is not given/annotated, neither at training nor test time. We tap\ninto two types of knowledge representations and reasoning. First, implicit\nknowledge which can be learned effectively from unsupervised language\npre-training and supervised training data with transformer-based models.\nSecond, explicit, symbolic knowledge encoded in knowledge bases. Our approach\ncombines both - exploiting the powerful implicit reasoning of transformer\nmodels for answer prediction, and integrating symbolic representations from a\nknowledge graph, while never losing their explicit semantics to an implicit\nembedding. We combine diverse sources of knowledge to cover the wide variety of\nknowledge needed to solve knowledge-based questions. We show our approach,\nKRISP (Knowledge Reasoning with Implicit and Symbolic rePresentations),\nsignificantly outperforms state-of-the-art on OK-VQA, the largest available\ndataset for open-domain knowledge-based VQA. We show with extensive ablations\nthat while our model successfully exploits implicit knowledge reasoning, the\nsymbolic answer module which explicitly connects the knowledge graph to the\nanswer vocabulary is critical to the performance of our method and generalizes\nto rare answers.\n