A neural network hidden unit with random fixed projection weights and two trained scale-and-shift parameters reaches accuracies close to fully trained layers on MNIST and FashionMNIST while using fewer trainable weights.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Half-Layered Neural Networks
A neural network hidden unit with random fixed projection weights and two trained scale-and-shift parameters reaches accuracies close to fully trained layers on MNIST and FashionMNIST while using fewer trainable weights.