Our approach is purely based on 2D convolutions. Nevertheless, it3 outperforms or performs comparably to many more costly 3D models. We thank the reviewers for pointing out some related (or missing) references. The12 Timeception layers involve group convolutions at different time scales while our TAM layers only use depthwise13 convolution. As a result, the Timeception has significantly more parameters than the TAM (10% vs. 0.1% of the14 totalmodelparameters).

To promote better performance-bandwidth trade-off for multi-agent perception, weproposeanovel distilledcollaborationgraph (DiscoGraph)tomodeltrainable, pose-aware, and adaptive collaboration among agents. Our key novelties lie in twoaspects.

One icy morning in February, nearly 200 people gathered in a church in downtown Richmond, Va. Most had awakened before dawn and driven in from across the state. There were Republicans and Democrats from rural farms and D.C. exurbs. They shared one goal: to fight back against AI development in a region with the largest concentration of data centers in the world. "Aren't you tired of being ignored by both parties, and having your quality of life and your environment absolutely destroyed by corporate greed?" state senator Danica Roem said, to a standing ovation. The activists--wearing homemade shirts with slogans like Boondoggle: Data Center in Botetourt County--marched to the state capitol and spent the day testifying to lawmakers about their fears over data centers' impacts on electricity, water, noise pollution, and more. Some lawmakers pledged to help: "You're getting a sh-t deal," state delegate John McAuliff told activists. The phrase captured many people's feelings toward the AI industry as a whole. Not much unites Americans these days.

Inthis paper,we employadversarial training to improve the performance of randomized smoothing.

We derive a tighter inequality than Cauchy-Schwarz Inequality, leverage it to refine Pearson's

Let = argm 2 CF ( )denote (2). T ( )+ log ( 1/ )+( 1) log ( 1/ ) log ( ) 1 , (5) where ( )isthe RDPof 2. Convergence:IfwerunAcldpwith t = DGpt, whereG2 = max{d1 Substitutingtheboundon4 into Lemma 2 together manipulationgivesproves Theorem 1; see Appendix E.2 fordetails.