Skip to main content
. Author manuscript; available in PMC: 2022 May 10.
Published in final edited form as: Adv Neural Inf Process Syst. 2021 Dec;34:4974–4986.

Figure 14:

Figure 14:

STL-digits dataset. Implicit feature modification reduces feature suppression, enhancing the representation of both MNIST and STL10 features simultaneously. All IFM runs use a fixed value ε = 0.1, and loss L+0.5Lε (i.e. weighting parameter α = 0.5) to illustrate robustness to the choice of parameters.