Source-linked AI summary
B-CNN: Branch Convolutional Neural Network for Hierarchical Classification
Xinqi Zhu, Michael Bain
TL;DR
Traditional CNN classifiers use a single flat output even when classes differ in difficulty and form hierarchies. The paper introduces B-CNN with coarse-to-fine branch predictions and a shifting-loss training strategy, reporting improvements over baseline CNNs across MNIST, CIFAR-10, and CIFAR-100.
Problem
Traditional CNN classifiers assume equally difficult, exclusive classes despite hierarchical relations and the CNN’s own coarse-to-fine layered representations.
Method
B-CNN adds branch networks for hierarchical predictions, while BT-strategy shifts loss-weight focus from coarse to fine levels during training.
Results
B-CNN models consistently outperform corresponding baseline CNNs on MNIST, CIFAR-10, and CIFAR-100.
Takeaways & Limitations
Hierarchical class priors can guide CNN learning across layers and strengthen classification while retaining a simple CNN-based model.
Takeaways & Limitations
Experiments use fewer than 80 epochs without data augmentation or ensemble methods and prioritize comparison with baselines over state-of-the-art performance.
Abstract
from arXiv · showhide
Convolutional Neural Network (CNN) image classifiers are traditionally designed to have sequential convolutional layers with a single output layer. This is based on the assumption that all target classes should be treated equally and exclusively. However, some classes can be more difficult to distinguish than others, and classes may be organized in a hierarchy of categories. At the same time, a CNN is designed to learn internal representations that abstract from the input data based on its hierarchical layered structure. So it is natural to ask if an inverse of this idea can be applied to learn a model that can predict over a classification hierarchy using multiple output layers in decreasing order of class abstraction. In this paper, we introduce a variant of the traditional CNN model named the Branch Convolutional Neural Network (B-CNN). A B-CNN model outputs multiple predictions ordered from coarse to fine along the concatenated convolutional layers corresponding to the hierarchical structure of the target classes, which can be regarded as a form of prior knowledge on the output. To learn with B-CNNs a novel training strategy, named the Branch Training strategy (BT-strategy), is introduced which balances the strictness of the prior with the freedom to adjust parameters on the output layers to minimize the loss. In this way we show that CNN based models can be forced to learn successively coarse to fine concepts in the internal layers at the output stage, and that hierarchical prior knowledge can be adopted to boost CNN models' classification performance. Our models are evaluated to show that the B-CNN extensions improve over the corresponding baseline CNN on the benchmark datasets MNIST, CIFAR-10 and CIFAR-100.
1. Introduction
Traditional CNN classifiers use a single flat output, despite class hierarchies and CNN layers both supporting coarse-to-fine representations. B-CNN integrates hierarchical class relations through branch outputs and trains them with a tailored strategy.
- Flat CNN classifiers assume all classes are equally difficult to distinguish and treat them exclusively.
- Hierarchical classification groups classes from coarse to fine, restricting errors to subcategories and making predictions more informative.For example, an image can first be classified as fruit before distinguishing apple from orange.
- CNN layers naturally progress from low-level features to high-level concepts, motivating predictions that become finer along the network.Lower layers capture basic shapes, whereas higher layers extract semantic features such as a dog’s face.
- B-CNN embeds hierarchical category relations by adding branch networks that produce hierarchical predictions along the convolutional workflow.The architecture is presented as a special CNN model with multiple output branches.
- The Branch Training strategy activates lower-level parameters earlier and shifts training toward higher-level parameters later.The strategy is intended to reduce vanishing-gradient effects and improve B-CNN performance.
- Experiments on MNIST, CIFAR-10, and CIFAR-100 report empirical benefits for the proposed model and training strategy.
2. Related Work
Related work combines CNNs with hierarchical labels, graphical models, and architectural changes, but differs in how it exploits hierarchy and trains the resulting models. B-CNN is positioned as a direct CNN-based alternative.
- CNN research has improved classification through deeper networks, more filters, activation functions, pooling operations, and initialization strategies.
- Class hierarchies have been predefined by humans or constructed automatically, and have been used to combine models for better performance.
- Some methods use tree priors to transfer knowledge between classes or graph-based label semantics, but leave CNN layers as feature extractors.These approaches do not exploit the hierarchical nature within CNN layers.
- HD-CNN exploits CNN hierarchy by sharing a coarse classifier for easy categories and using fine classifiers for difficult categories.The cited discussion identifies a potential requirement for coarse information in HD-CNN.
- Graphical-model approaches may require two-stage training or specialized training methods, while Structured Prediction Energy Networks are not CNN-based.
- BT-strategy addresses vanishing gradients in B-CNNs by shifting the loss weights of coarse-level outputs during training.
3. Model Description
B-CNN adds internal output branches aligned with levels of a hierarchical label tree and combines their losses. Its Branch Training strategy shifts emphasis from coarse to fine predictions during end-to-end training.
- Label hierarchy: A B-CNN label tree contains fine target labels as leaves and groups them into manually constructed or automatically generated coarse categories.Multiple coarse levels can form progressively more general categories.
- Branch architecture: Each B-CNN branch corresponds to a label-tree level and produces a prediction using fully connected and softmax layers.The underlying network can be an arbitrary ConvNet, while experimental branch networks use fully connected layers.
- Branch architecture: The model outputs as many hierarchical predictions as the label tree has levels, with each level contributing to the final loss.
- Loss function: B-CNN’s loss is a weighted sum of prediction losses across coarse and fine levels.The weights determine each branch’s contribution to the final objective.
- Loss function: Setting the final branch’s weight to 1 and all other branch weights to 0 makes a multi-branch B-CNN converge to a traditional CNN.The loss weights are normalized to sum to 1 in the reported representation.
- Branch Training Strategy: BT-strategy shifts the largest loss-weight focus from lower levels to higher levels, guiding coarse feature learning before fine-tuning.This schedule is intended to reduce the effect of vanishing gradients on lower-layer updates.
4. Experiments
The experiments evaluate B-CNN models and the BT-strategy on MNIST, CIFAR-10, and CIFAR-100 against corresponding CNN baselines. Across these benchmarks, B-CNNs show hierarchical learning behavior and consistently outperform their baselines, although pretrained initialization narrows the advantage.
- Experimental setup: Experiments compare B-CNN models with corresponding traditional CNN baselines on MNIST, CIFAR-10, and CIFAR-100.The study uses SGD and limits training to fewer than 80 epochs, emphasizing comparison rather than state-of-the-art performance.
- Hierarchical classification: CIFAR-10 level accuracies are 96.26% for coarse 1, 86.74% for coarse 2, and 82.19% for fine classification.Coarse 1 trains fastest and achieves the highest accuracy, while the fine level is the most difficult task.
- CIFAR-10 comparison: The B-CNN model B reaches 84.41% accuracy on CIFAR-10 versus 82.35% for baseline model B after 60 epochs.The fine-level loss weight increases during training, coinciding with a steeper growth trend toward specific-feature learning.
- CIFAR-100 comparison: On CIFAR-100, B-CNN model B reaches 57.59% versus 51.00% for its baseline, while model C reaches 64.42% versus 62.92%.The B-CNN model can also reach high accuracy faster than the baseline under the limited training schedule.
- Analysis: Across three datasets, B-CNNs consistently outperform corresponding baselines, but pretrained initialization makes the performance gap less pronounced.The experiments attribute the smaller gap for pretrained models to easier baseline fine-tuning and reduced importance of the BT-strategy.
5. Conclusion
B-CNN connects CNNs’ hierarchical representations with structured target-class priors, producing coarse-to-fine predictions. Its Branch Training strategy uses label trees to guide learning and improves performance over traditional CNNs.
- B-CNN connects CNN hierarchical representations with structured priors on target classes.
- B-CNN outputs multiple hierarchical predictions from coarse to fine, making classification more informative and interpretable.
- The Branch Training strategy first learns low-level features before converging toward traditional CNN classification.
- Experiments confirm that B-CNN improves performance over traditional CNN models.
- Future work should investigate structured outputs such as linear chains or graphs.