ABSTRACT This paper introduces two novel neural network architectures based on the learned separable transform (LST) for efficient image recognition. First, we present MCNN‐LST, a hybrid architecture that integrates multi‐column convolutional networks with LST blocks to compress multi‐scale features into compact, discriminative embeddings. Second, we propose MCNN‐Eff‐LST, which employs mobile inverted bottleneck convolutions to extract efficient multi‐scale representations and utilizes LST as a learnable alternative to global average pooling. Comprehensive evaluation on Fashion‐MNIST and CIFAR‐10 benchmarks demonstrates the effectiveness of both approaches. The MCNN‐LST architecture achieves 93.69% accuracy on Fashion‐MNIST with only 57.7k parameters (1.8 MFLOPs). The MCNN‐Eff‐LST variant attains 84.40% accuracy on CIFAR‐10 using merely 37.7k parameters (5.5 MFLOPs)—substantially fewer than comparable lightweight models. Ablation studies confirm the contribution of LST blocks: removing them causes accuracy drops of 1.69% and 2.74% for MCNN‐LST and MCNN‐Eff‐LST, respectively. The experiments showed that LST‐based models represent an effective trade‐off between parameter efficiency and recognition performance, making them particularly suitable for edge computing applications.
Vashkevich et al. (Thu,) studied this question.