Indirectly Parameterized Concrete Autoencoders
This work addresses a specific bottleneck in neural network-based feature selection for high-dimensional or costly data, offering an incremental improvement over the state-of-the-art CAE method.
The paper tackled the instability and duplicate selection issues in Concrete Autoencoders (CAEs) for feature selection by proposing Indirectly Parameterized CAEs (IP-CAEs), which learn an embedding to map to Gumbel-Softmax parameters, resulting in significant improvements in generalization and training time across multiple datasets.
Feature selection is a crucial task in settings where data is high-dimensional or acquiring the full set of features is costly. Recent developments in neural network-based embedded feature selection show promising results across a wide range of applications. Concrete Autoencoders (CAEs), considered state-of-the-art in embedded feature selection, may struggle to achieve stable joint optimization, hurting their training time and generalization. In this work, we identify that this instability is correlated with the CAE learning duplicate selections. To remedy this, we propose a simple and effective improvement: Indirectly Parameterized CAEs (IP-CAEs). IP-CAEs learn an embedding and a mapping from it to the Gumbel-Softmax distributions' parameters. Despite being simple to implement, IP-CAE exhibits significant and consistent improvements over CAE in both generalization and training time across several datasets for reconstruction and classification. Unlike CAE, IP-CAE effectively leverages non-linear relationships and does not require retraining the jointly optimized decoder. Furthermore, our approach is, in principle, generalizable to Gumbel-Softmax distributions beyond feature selection.