Abstract

Classification problems solved with deep neural networks (DNNs) typically rely on a closed world paradigm, and optimize over a single objective (e.g., minimization of the cross-entropy loss). This setup dismisses all kinds of supporting signals that can be used to reinforce the existence or absence of particular patterns. The increasing need for models that are interpretable by design makes the inclusion of said contextual signals a crucial necessity. To this end, we introduce the notion of Self-Supervised Autogenous Learning (SSAL). A SSAL objective is realized through one or more additional targets that are derived from the original supervised classification task, following architectural principles found in multi-task learning. SSAL branches impose low-level priors into the optimization process (e.g., grouping). The ability of using SSAL branches during inference, allow models to converge faster, focusing on a richer set of class-relevant features. We equip state-of-the-art DNNs with SSAL objectives and report consistent improvements for all of them on CIFAR100 and Imagenet. We show that SSAL models outperform similar state-of-the-art methods focused on contextual loss functions, auxiliary branches and hierarchical priors.

deep learning self-supervised autogenous learning with auxiliary grouping classifiers

Paper (accepted for publication at ICPR 2020)

          Cite as: Self-Supervised Autogenous Learning \cite{palacio2020contextual}

          @InProceedings{palacio2020contextual,
            author = {Sebastian Palacio and Philipp Engler and Joern Hees and Andreas Dengel},
            title = {Contextual Classification Using Self-Supervised Auxiliary Models for Deep Neural Networks},
            booktitle = {International Conference on Pattern Recognition (ICPR)},
            month = {January},
            year = {2021}
          }
        

State of the art on Cifar100 and Imagenet

state of the art classification on imagenet and cifar100

Structured Prediction and Better Interpretation with Class Activation Maps

more structured interpretation of predictions with ssal auxiliary branches using class activation maps

Supplementary Material

PDF with supplementary material.

Additional experiments and sample predictions

Video summary

Example code.

Repository for SSAL networks (with pre-trained models)

Blog Post

Checkout our recent blog post about the paper!

Acknowledgments

This work was supported by the BMBF projects ExplAINN (01IS19074), DeFuseNN (01IW17002) and the NVIDIA AI Lab program.