Width-Adaptive Convolutional Autoencoder with Channels’ Relevance Weighting Mechanism

In this paper, we propose a novel Width-Adaptive Convolutional Autoencoder (WACAE) that automatically learns the optimal network width. The proposed approach assigns a relevance weight to each channel in the encoder’s hidden layers and leverages these weights to guide architectural adaptation. Based on the learned relevance, the model incrementally introduces new channels when needed and prunes irrelevant ones to achieve an optimal configuration. The WACAE simultaneously trains the network and learns its width in an unsupervised manner. Moreover, a novel cost function is devised to optimize channel relevance weights concurrently with model hyperparameters. Unlike conventional static or widening strategies, the proposed method adaptively enhances feature expressiveness within a single encoder–decoder framework. The model is evaluated on standard benchmark datasets (MNIST and CIFAR-10) and two real-world medical datasets (Brain Tumor MRI and Kvasir-Capsule). Experimental results demonstrate its effectiveness compared to state-of-the-art methods based on empirical tuning and network-width scaling. Furthermore, the proposed inner-product-based relevance weighting mechanism reduces model complexity while achieving high classification accuracy.