[Paper Review] Scale-Invariant Convolutional Neural Networks

Yichong Xu, Tianjun Xiao+3 more|arXiv (Cornell University)|Nov 24, 2014

Advanced Neural Network ApplicationsComputer Science16 references105 citations

TL;DR

This paper proposes Scale-Invariant Convolutional Neural Networks (SiCNN), a multi-column CNN architecture that achieves scale invariance by sharing filter parameters across columns via scale transformations, enabling robust performance across object scales without increasing model size or requiring extensive data augmentation. SiCNN reduces classification error by 2.1% on CIFAR-10 (14.22%) compared to baseline CNN, and when combined with maxout, achieves 11.33% error rate with minimal additional training cost.

ABSTRACT

Even though convolutional neural networks (CNN) has achieved near-human performance in various computer vision tasks, its ability to tolerate scale variations is limited. The popular practise is making the model bigger first, and then train it with data augmentation using extensive scale-jittering. In this paper, we propose a scaleinvariant convolutional neural network (SiCNN), a modeldesigned to incorporate multi-scale feature exaction and classification into the network structure. SiCNN uses a multi-column architecture, with each column focusing on a particular scale. Unlike previous multi-column strategies, these columns share the same set of filter parameters by a scale transformation among them. This design deals with scale variation without blowing up the model size. Experimental results show that SiCNN detects features at various scales, and the classification result exhibits strong robustness against object scale variations.

Motivation & Objective

• To address the limited scale invariance in standard CNNs, which struggle to recognize objects across varying input scales.
• To develop a model that maintains robustness to scale variations without increasing the number of free parameters or relying on data augmentation like scale-jittering.
• To enable multi-scale feature learning within a single architecture by specializing each column on a specific scale while enforcing parameter sharing.
• To explore incremental training strategies that reduce training cost while preserving performance gains from scale invariance.
• To demonstrate that SiCNN complements existing techniques like maxout and dropout, enhancing overall accuracy on benchmark datasets.

Proposed method

• SiCNN employs a multi-column CNN architecture, where each column processes the input at a different scale using scale-transformed filters.
• Filters across columns are parameter-shared via a scale transformation, ensuring that the same filter is applied at different scales without duplicating parameters.
• The network uses standard convolutional and max-pooling layers in each column, with final features concatenated before classification.
• A key innovation is the use of scale transformation to map filters from one column to another, maintaining parameter efficiency while enabling scale-specific feature detection.
• Incremental training is applied: first train a single-column CNN, then initialize multi-column SiCNN from its filters, and finally refine the full model with minimal additional training.
• The model uses a single softmax classifier, and feature fusion is achieved via concatenation across columns.

Experimental results

Research questions

• RQ1 Can a CNN architecture be designed to achieve scale invariance without increasing model size or requiring data augmentation?
• RQ2 Does parameter sharing across columns via scale transformation lead to improved robustness against scale variations in object recognition?
• RQ3 Can incremental training reduce the computational cost of training SiCNN while maintaining high performance?
• RQ4 How does SiCNN compare to state-of-the-art models like maxout and network-in-network on standard benchmarks?
• RQ5 Can SiCNN be effectively combined with other optimization techniques such as dropout or maxout to further improve accuracy?

Key findings

• SiCNN achieves a test error rate of 14.22% on the unaugmented CIFAR-10 dataset, representing a 2.1% absolute improvement over the baseline CNN.
• The model demonstrates strong robustness to scale variations, with the column specialized for larger scales activating more strongly on larger input patterns.
• Incremental training reduces training cost by nearly half (to 3.5× baseline cost) while maintaining comparable performance to training from scratch (14.71% error).
• When combined with maxout, SiCNN reduces the error rate from 11.68% to 11.33% using a single 2-column maxout-SiCNN model.
• A 6-column SiCNN with independent filters overfits severely (test error ~19%), confirming the necessity of parameter sharing for generalization.
• The method is complementary to existing techniques like dropout and maxout, and shows promise for larger datasets like ImageNet.