CIFAR-100 MLP
BatchNorm precedes the activation. The benchmark uses three random seeds.
Preliminary research release
VNorm
in the loop.
A parametric activation function designed to work with Batch Normalization, giving the network a learnable nonlinear response instead of a fixed threshold.
BATCHNORMV-NORM
Normalize first.
Learn the curve.
VNorm is studied as a companion to BatchNorm, not as a claim that one activation wins in every setting.
The public release contains the implementation, configuration, and benchmark code for a CIFAR-100 MLP comparison.
Focused comparison
Four references
VNorm is compared with ReLU, CELU, SiLU, and PReLU.
Transparent release
Preliminary, reproducible, and intentionally clear about its BatchNorm dependence.
From the run log
Results,
in motion.
CIFAR-100 MLP · CUDA · 30 epochs · seeds 1, 7, and 42. Values transcribed from the benchmark run.
Mean validation accuracy
Evolution averaged across the three random seeds.
ReLUCELUSiLUPReLUVNorm
Final test accuracy
Mean accuracy; vertical markers show the range across seeds.
Higher is better · mean values shown at right.
Public release
Inspect it.
Challenge it.
VNorm is part of Carlos M. Patino Machado's research portfolio and is developed under Tetraspace LLC.