Deeply-supervised Knowledge Synergy (1906.00675v2)

Abstract: Convolutional Neural Networks (CNNs) have become deeper and more complicated compared with the pioneering AlexNet. However, current prevailing training scheme follows the previous way of adding supervision to the last layer of the network only and propagating error information up layer-by-layer. In this paper, we propose Deeply-supervised Knowledge Synergy (DKS), a new method aiming to train CNNs with improved generalization ability for image classification tasks without introducing extra computational cost during inference. Inspired by the deeply-supervised learning scheme, we first append auxiliary supervision branches on top of certain intermediate network layers. While properly using auxiliary supervision can improve model accuracy to some degree, we go one step further to explore the possibility of utilizing the probabilistic knowledge dynamically learnt by the classifiers connected to the backbone network as a new regularization to improve the training. A novel synergy loss, which considers pairwise knowledge matching among all supervision branches, is presented. Intriguingly, it enables dense pairwise knowledge matching operations in both top-down and bottom-up directions at each training iteration, resembling a dynamic synergy process for the same task. We evaluate DKS on image classification datasets using state-of-the-art CNN architectures, and show that the models trained with it are consistently better than the corresponding counterparts. For instance, on the ImageNet classification benchmark, our ResNet-152 model outperforms the baseline model with a 1.47% margin in Top-1 accuracy. Code is available at https://github.com/sundw2014/DKS.

• The paper introduces a DKS method that uses auxiliary supervision branches and synergy loss to enable dynamic knowledge sharing among CNN classifiers.
• The methodology leverages intermediate layer support, resulting in enhanced generalization and a 1.47% Top-1 accuracy improvement on ImageNet.
• The approach enhances CNN training without additional inference costs, offering a scalable upgrade for state-of-the-art architectures.

Deeply-supervised Knowledge Synergy: Enhancing CNN Training

The paper "Deeply-supervised Knowledge Synergy" by Sun et al. introduces an innovative methodology to enhance the training of deep Convolutional Neural Networks (CNNs) for image classification without additional computational costs during inference. In response to the need for improved generalization capabilities in deeper and more intricate CNN architectures, this research presents the Deeply-supervised Knowledge Synergy (DKS) framework. The core idea involves utilizing auxiliary supervision and a novel synergy loss to dynamically facilitate knowledge sharing among classifiers in a network, leading to more robust training outcomes.

Overview

The prevailing approach to training CNNs predominantly involves adding supervision only to the final network layer. While effective, this method risks underutilizing the potential of intermediate layers, especially in deeper networks. Sun et al. propose an advanced training scheme by incorporating auxiliary supervision branches on intermediate layers, drawing inspiration from the deeply-supervised learning paradigm. However, they innovate further by introducing the concept of a synergy loss, which ensures pairwise knowledge matching across all supervision branches, thus enhancing the learning process in both top-down and bottom-up directions.

Methodology

The proposed methodology of DKS involves several key components:

1. Auxiliary Supervision Branches: By appending auxiliary branches at certain intermediate layers, the model gains additional supervision throughout its architecture. These branches are designed to be complex, containing building blocks akin to those in the backbone network, thus ensuring consistency in feature extraction and classification tasks.
2. Synergy Loss: The primary innovation lies in the synergy loss, which orchestrates pairwise knowledge sharing among all classifiers, including those in auxiliary branches. Unlike traditional methods where only the final layer's classifier guides the training, the synergy loss enables dynamic interaction amongst various points in the network, promoting a balanced learning experience across the network's depth.
3. Dense Pairwise Matching: The synergy mechanism is employed via dense pairwise knowledge matching operations. This approach ensures that each classifier influences and is influenced by others, fostering a cohesive learning environment that reduces overfitting and enhances generalization.

Results

Experiments conducted on major image classification datasets such as ImageNet and CIFAR-100 using state-of-the-art architectures like ResNet, DenseNet, and MobileNet demonstrate the significant improvements that DKS offers. Notably, on the ImageNet dataset, DKS showed a notable improvement of 1.47% in Top-1 accuracy for the ResNet-152 model compared to its baseline counterpart. These results underscore the efficacy of the proposed methodology in enhancing model performance while maintaining computational efficiency.

Implications and Future Directions

The introduction of DKS offers both practical and theoretical advancements in CNN training methodologies. Practically, it provides a mechanism to leverage knowledge from intermediate layers without additional inference costs, crucial for deploying CNNs in resource-constrained environments. Theoretically, the framework challenges traditional hierarchical learning paradigms by promoting holistic layer interactions.

Looking forward, the concepts explored in DKS could be extended to other domains where multi-layer interactions and training regularization are beneficial. Furthermore, potential research directions could involve exploring the integration of DKS with other forms of regularization or applying it to other neural architectures beyond CNNs, such as transformers or graph neural networks.

In summary, the Deeply-supervised Knowledge Synergy offers a compelling enhancement to the training of neural networks, underscored by its innovative use of synergy loss and auxiliary branches for dynamic inter-layer knowledge transfer, marking a notable contribution to the field of deep learning.