The other relations can also be represented in the same way. One can freely define other types of spatial-temporal predicates. We can see that all of methods have better performance with the increase of training triplets. And our model achieves the best results. The parameters and FLOPs of all methods is shown in Table 1.

In this paper, we introduce a set of spatial-temporal logic rules as knowledge to explain human actions.

However, SSA T suffers from catastrophic overfit-ting (CO), a phenomenon that leads to a severely distorted classifier, making it vulnerable to multi-step adversarial attacks. In this work, we observe that some adversarial examples generated on the SSA T -trained network exhibit anomalous behaviour, that is, although these training samples are generated by the inner maximization process, their associated loss decreases instead, which we named abnormal adversarial examples (AAEs).

In this paper, we explore the use of affinity measures as features in graph neural networks, in particular measures arising from random walks, including effective resistance, hitting and commute times.