Our nonconvex optimization formulation solves for a filter h on the unit sphere that produces sparse outputyi ~h.

Acommon formulation of curiosity-driven exploration uses the difference between the real future and the future predicted by a learned model [1].

Figure 5: Graph-Q-SAT's MRIR improvement (10 model calls) results in the wall clock time reduction. We call the middle part'the core'. The output of the core is concatenated with the output of the encoder andgetsfedtothecoreagain. We also plan to release the experimental code and the modified version of MiniSat to use as a gym environment. Encoder and Decoder are independent graph networks,i.e.

For VaTeX captioning and retrieval, we use the latest v1.1 version3, which contains 25,991 videos for training and 6,000 videos for public testing. The statistics can be found in Table 1. Visual genome synsets are pairs, where the keys are noisy natural language phrases and the values are the mapped WordNet synsets [6]. Ifavisualtokenoccurs in multiple frames, we use the averaged frame indexas its temporal indicator. Specifically,for UniVL, we set the number of epoches to be50 and the linear warmup steps to be40.

C.1 Ablationstudy Ablation study results are provided in Table 3.