Semantic-Aware Auto-Encoders for Self-supervised Representation Learning

The resurgence of unsupervised learning can be attributed to the remarkable progress of self-supervised learning, which includes generative <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathcal{G})$</tex> and discriminative <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathcal{D})$</tex> models. In computer vision, the mainstream self-supervised learning algorithms are <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{D}$</tex> models. However, designing a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{D}$</tex> model could be over-complicated; also, some studies hinted that a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{D}$</tex> model might not be as general and interpretable as a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{G}$</tex> model. In this paper, we switch from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{D}$</tex> models to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{G}$</tex> models using the classical auto-encoder <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(AE)$</tex> . Note that a vanilla <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{G}$</tex> model was far less efficient than a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{D}$</tex> model in self-supervised computer vision tasks, as it wastes model capability on overfitting semantic-agnostic high-frequency details. Inspired by perceptual learning that could use cross-view learning to perceive concepts and semantics <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> Following [26], we refer to semantics as visual concepts, e.g., a semantic-ware model indicates the model can perceive visual concepts, and the learned features are efficient in object recognition, detection, etc., we propose a novel <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$AE$</tex> that could learn semantic-aware representation via cross-view image reconstruction. We use one view of an image as the input and another view of the same image as the reconstruction target. This kind of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$AE$</tex> has rarely been studied before, and the optimization is very difficult. To enhance learning ability and find a feasible solution, we propose a semantic aligner that uses geometric transformation knowledge to align the hidden code of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$AE$</tex> to help optimization. These techniques significantly improve the representation learning ability of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$AE$</tex> and make selfsupervised learning with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathcal{G}$</tex> models possible. Extensive experiments on many large-scale benchmarks (e.g., ImageNet, COCO 2017, and SYSU-30k) demonstrate the effectiveness of our methods. Code is available at https://github.com/wanggrun/Semantic-Aware-AE.