Breaking01 / AI
Reusing FAct improves image classification accuracy in tests
FAct ranked first in 11 of 16 test settings when training settings were tuned separately for each activation function.

Researchers reported on 20 Nov 2026 that reusing FAct raised image classification accuracy by 1.46 percentage points in a new model with the same architecture.
FAct is an activation function, which transforms signals in a neural network. The researchers trained it alongside a model, then fixed its parameters before applying it to the new model.
According to the paper, FAct with fixed parameters achieved the highest mean accuracy in 15 of 16 test settings when all activation functions shared the same training settings. It ranked first in 11 of 16 settings when training settings were tuned separately for each function.
The paper used a vision transformer with the same architecture and the existing GELU activation function as the baseline. Both models were trained from scratch on ImageNet-1K, an image classification dataset.
The abstract did not disclose how using FAct with fixed parameters affected model training time.