[Paper Review] MeliusNet: Can Binary Neural Networks Achieve MobileNet-level Accuracy?

Joseph Bethge, Christian Bartz+3 more|arXiv (Cornell University)|Jan 16, 2020

Advanced Neural Network Applications48 references75 citations

TL;DR

MeliusNet introduces a binary neural network architecture with DenseBlock and ImprovementBlock to increase feature capacity and quality, achieving MobileNet-v1–level accuracy at similar model size and operations on ImageNet.

ABSTRACT

Binary Neural Networks (BNNs) are neural networks which use binary weights and activations instead of the typical 32-bit floating point values. They have reduced model sizes and allow for efficient inference on mobile or embedded devices with limited power and computational resources. However, the binarization of weights and activations leads to feature maps of lower quality and lower capacity and thus a drop in accuracy compared to traditional networks. Previous work has increased the number of channels or used multiple binary bases to alleviate these problems. In this paper, we instead present an architectural approach: MeliusNet. It consists of alternating a DenseBlock, which increases the feature capacity, and our proposed ImprovementBlock, which increases the feature quality. Experiments on the ImageNet dataset demonstrate the superior performance of our MeliusNet over a variety of popular binary architectures with regards to both computation savings and accuracy. Furthermore, with our method we trained BNN models, which for the first time can match the accuracy of the popular compact network MobileNet-v1 in terms of model size, number of operations and accuracy. Our code is published online at https://github.com/hpi-xnor/BMXNet-v2

Motivation & Objective

• Address the accuracy gap between Binary Neural Networks (BNNs) and compact networks on ImageNet.
• Propose an architectural solution that boosts feature capacity and quality specifically for BNNs.
• Reduce the computational burden of the residual 32-bit layers while preserving or improving accuracy.
• Demonstrate that BNNs can reach MobileNet-v1–level performance in terms of model size, FLOPs, and accuracy.

Proposed method

• Introduce MeliusNet, alternating DenseBlocks (increase feature capacity) with ImprovementBlocks (increase feature quality).
• Use 64 new binary channels per DenseBlock, concatenated to the feature map, followed by a 64-channel binary ImprovementBlock with a residual connection to refine these features.
• Redesign the first layers with a grouped-stem approach (replacing a 7x7 conv with three grouped 3x3 convs) to substantially reduce 32-bit operations without increasing model size.
• Avoid binarizing the initial, final, and some downsampling convolutions to preserve accuracy, and omit channel-wise scaling factors based on empirical findings.
• Train models from scratch on ImageNet using a specific STE-based binarization, with Adam/RAdam optimizers and cosine learning rate scheduling, including optional warm-up.

Experimental results

Research questions

• RQ1 Can a binary neural network architecture be designed to match MobileNet-v1 in accuracy at comparable size and compute on ImageNet?
• RQ2 Do architectural innovations tailored for BNNs (capacity-increasing blocks and quality-improving blocks) yield measurable gains over prior binary architectures?
• RQ3 Does a grouped-stem 32-bit layer design reduce operations enough to close the gap with compact networks without sacrificing accuracy?
• RQ4 Are improvements in BNN performance specific to binary architectures, or do they extend to 32-bit networks as well?

Key findings

• MeliusNet consistently outperforms prior binary architectures at similar model sizes and operation counts on ImageNet.
• The grouped stem significantly reduces 32-bit operations while increasing accuracy across tested models.
• A 32-bit version of MeliusNet does not show improvements over DenseNet, indicating the benefits are specific to BNNs.
• MeliusNet configurations can match MobileNet-v1 accuracy for multiple target sizes/operations, demonstrating competitive parity with compact networks.
• The proposed grouped stem combined with the MeliusNet architecture narrows the accuracy gap between BNNs and MobileNet-v1, highlighting potential energy savings on suitable hardware.