Abstract
URNet (User-Resizable Residual Network) introduces a flexible deep learning architecture that enables runtime adjustment of computational costs without retraining. By incorporating a lightweight Conditional Gating Module (CGM), URNet learns to skip specific residual blocks based on input features and a user-defined resource budget. This approach allows for direct control over the accuracy-efficiency trade-off, making it ideal for environments with fluctuating service demands. Evaluated on ImageNet and CIFAR benchmarks, URNet demonstrates superior performance compared to static compression and existing dynamic networks, maintaining high accuracy even under significant computational constraints.
Introduction
Standard Convolutional Neural Networks (CNNs) are typically designed with a fixed computational cost. However, real-world applications often face fluctuating resource availability and service demands. For instance, a cloud-based inference server may need to process a sudden surge of requests by reducing the per-task precision to maintain overall throughput.
Existing solutions, such as model compression or standard dynamic networks, lack the necessary flexibility. Compression creates a static, smaller model, while dynamic networks decide their own execution paths without allowing for external user control. URNet addresses these limitations by providing a "resizable" architecture that gives users direct control over the trade-off between speed and accuracy via a simple scale parameter.
Methodology
The core innovation of URNet is the Conditional Gating Module (CGM), which is integrated into a standard ResNet backbone.
The Gating Mechanism
A CGM is placed before each residual block. It processes features from the previous layer alongside a user-specified scale parameter $S$ (ranging from 0 to 1). The module outputs a binary decision to either execute or skip the associated residual block. These modules are extremely efficient, adding less than 0.1% computational overhead to the base model.
Training Strategy
To overcome the non-differentiability of binary decisions, URNet employs a probabilistic training approach. During training, the gate operates as a sigmoid function with probability $p$ and a binary step function with probability $1-p$, allowing for standard backpropagation.
To ensure the network adheres to the user-defined budget, the authors introduce a Scale Loss ($L_s$). This loss function penalizes the model if the average number of active blocks deviates from the target scale $S$. The final objective is a weighted sum of the classification loss and the scale loss.
Experiments and Results
URNet was evaluated on CIFAR-10, CIFAR-100, and ImageNet (ILSVRC2012). The results demonstrate that URNet consistently achieves a better accuracy-vs-FLOPs trade-off than other state-of-the-art dynamic methods like BlockDrop.
Key Findings
- Efficiency on ImageNet: Using a ResNet-101 backbone, URNet maintains baseline accuracy (76.4%) while utilizing only 80% of the original computation. At 65% computation, the accuracy drop is limited to approximately 1%.
- Smart Regularization: At scales around 70-80%, URNet occasionally outperforms the baseline ResNet. This suggests the gating mechanism acts as a form of intelligent regularization, similar to stochastic depth.
- Input-Aware Adaptability: Qualitative analysis reveals that the model intelligently skips blocks for "easy" images while activating more blocks for complex or noisy scenes, even when a lower overall scale is requested.
Conclusion and Future Work
URNet provides a robust framework for deploying deep learning models in dynamic environments. By allowing users to "dial" computational costs at runtime, it bridges the gap between static model design and real-world operational requirements. While training for a wide range of scales can slightly impact peak performance, the flexibility and stability of the model across $S \in [0.2, 1.0]$ make it a highly practical solution. Future research may explore hardware-specific optimizations to further translate theoretical FLOP savings into wall-clock speedups.
Summary updated