Standard Convolutional Neural Networks (CNNs) are typically designed as a linear hierarchy where information flows from low-level edge detectors to high-level semantic features. In this paradigm, only the final feature vector is used for classification. However, research suggests that the layers immediately preceding the output contain rich contextual information that is often discarded. This paper proposes a "multilevel context" approach that integrates features from multiple high-level blocks to improve scale and translation invariance.
The Multilevel Context Approach
The core hypothesis is that features close to the higher layers provide better invariance to transformations than the top layer alone. By "short-circuiting" the network to include outputs from the penultimate layers, the model gains a more robust representation of the input.
To implement this, the authors introduced modifications to two standard architectures:
- AlexNet++: Instead of feeding only
conv5into the first fully connected layer (fc6), the output ofconv4is also included. Theconv4features are max-pooled to match spatial dimensions and normalized using a tanh function to ensure both layers contribute equally to the final representation. - GoogLeNet++: A similar connection is added from an earlier Inception block (
inception4e) directly to the final concatenation layer before the classifier.
Addressing Fine-Grained Recognition
One of the primary motivations for using multilevel context is the difficulty of distinguishing between highly similar classes. In ImageNet, for instance, the difference between a "Tabby cat," a "Tiger cat," and an "Egyptian cat" can be subtle.
The authors argue that while the top layer might focus on specific discriminative parts, the preceding layers provide the broader context necessary to handle variations in scale, occlusion, and orientation that are common in these complex categories.
Experimental Results and Efficiency
The proposed models were evaluated on the ILSVRC ImageNet dataset. The results demonstrated a consistent improvement in classification accuracy:
- Accuracy Gains: Both AlexNet++ and GoogLeNet++ reduced classification error by approximately 1-2% compared to their baseline counterparts. GoogLeNet++ achieved a Top-1 accuracy of 69.42% using a single center crop, outperforming the original's 68.89%.
- Inference Speedup: A significant finding was that the "++" models with a single-crop input could match the performance of standard models using a 144-crop ensemble. This effectively provides a 144x speedup during inference for the same level of accuracy.
- Computational Cost: The architectural changes are remarkably efficient. Training GoogLeNet++ was only 0.4% slower than the original, making it a low-cost upgrade for significant performance gains.
Conclusion
By rethinking the connectivity of high-level layers, this work demonstrates that multilevel context is a powerful tool for improving object recognition. The approach is orthogonal to traditional data augmentation techniques and provides a path toward highly efficient, high-accuracy inference. Future work may extend these principles to temporal context in recurrent architectures.
Summary updated