The paper proposes the DeSign activation function for binary neural networks (BNNs) to address the loss of precision and fine-grained details caused by common binary activations like the Sign function.
DeSign applies a spatially periodic threshold kernel to the Sign activation, shifting the thresholds for each pixel based on a designed 2D or 3D pattern. This leverages local spatial correlations to better preserve the distribution of values from binary convolutions, compared to using an independent threshold per pixel.
The threshold kernel is designed through an optimization-based methodology that selects the kernel maximizing the preservation of structural information, measured by the expected total variation. The designed kernel is then scaled to align with the batch normalization process.
Experiments on image classification tasks demonstrate that DeSign can boost the accuracy of BNN architectures like VGGsmall and ResNet18, without increasing computational cost. DeSign also mitigates the influence of real-valued batch normalization layers, enhancing baseline BNN accuracy by up to 4.51%. The 3D variant of DeSign, which applies different thresholds per channel, further improves performance.
Overall, the DeSign activation provides an effective way to improve the accuracy of BNNs while maintaining the efficiency of binary operations.
To Another Language
from source content
arxiv.org
Önemli Bilgiler Şuradan Elde Edildi
by Bray... : arxiv.org 05-06-2024
https://arxiv.org/pdf/2405.02220.pdfDaha Derin Sorular