PolyNeXt

Vision models ditch activations for polynomial alternatives

Researchers have developed new activation-free backbone architectures for vision models, utilizing polynomial functions instead of traditional pointwise nonlinearities like ReLU or GELU. These novel modules, integrated …