We then propose two methods to mitigate this problem.

ImageNet-100 labels, to define the positive samples. In Figure 1, we track the proxy task performance when progressively moving from MoCo to MoCo-v2, i.e . Figure 1 are for the same τ = 0 .2 . B.2 Hard negative mixing variants not discussed in the main text While developing MoCHi, we considered a number of different mixing strategies in feature space. We found the two strategies presented in Sections 4.1 and 4.2 of the main paper to For MoCHi, the "top" negatives are defined via the negative For MoCHi, in Section 4.2 we propose to synthesize MoCHi samples according to the percentage of the query they have.

We investigate the multiscale extension of the problem where the loss ranges of the experts are vastly different.

We study long-run market stability for repeated price competitions between two firms, where consumer demand depends on firms' posted prices and consumers'

We show that, surprisingly, the notion of optimal finite-time regret is not a uniquely defined property in this context and that, in general, it is decoupled from the asymptotic rate. We discuss alternative choices and propose a notion of finite-time optimality that we argue is meaningful .

In addition, current offline RL benchmarks test agents on the exact same environment where the data was collected.