Switchable Activation Networks

ArXi:2603.06601v1 Announce Type: new Deep neural networks, and recently large-scale generative models such as large language models (LLMs) and large vision-action models (LVAs), achieve remarkable performance across diverse domains, yet their prohibitive computational cost hinders deployment in resource-constrained environments. Existing efficiency techniques offer only partial remedies: dropout improves regularization during