Abstract:Deep generative neural networks (DGNNs) have achieved realistic and high-quality data generation. In particular, the adversarial training scheme has been applied to many DGNNs and has exhibited powerful performance. Despite of recent advances in generative networks, identifying the image generation mechanism still remains challenging. In this paper, we present an explorative sampling algorithm to analyze generation mechanism of DGNNs. Our method efficiently obtains samples with identical attributes from a quer… Show more
“…For example, in CNNs, decision regions are divided by polyhedral cones (Carlsson 2019) so that the angular difference between feature maps becomes highly related to the Configuration distance. This aligns with the empirical successes of prior work using the Cosine similarity in the feature space (Fong and Vedaldi 2018;Kim et al 2018;Bachman, Hjelm, and Buchwalter 2019;Jeon, Jeong, and Choi 2020). We plan to explore this phenomenon further in our future work.…”
Section: Analysis For the Distance Metricssupporting
Understanding intermediate representations of the concepts learned by deep learning classifiers is indispensable for interpreting general model behaviors. Existing approaches to reveal learned concepts often rely on human supervision, such as pre-defined concept sets or segmentation processes. In this paper, we propose a novel unsupervised method for discovering distributed representations of concepts by selecting a principal subset of neurons. Our empirical findings demonstrate that instances with similar neuron activation states tend to share coherent concepts. Based on the observations, the proposed method selects principal neurons that construct an interpretable region, namely a Relaxed Decision Region (RDR), encompassing instances with coherent concepts in the feature space. It can be utilized to identify unlabeled subclasses within data and to detect the causes of misclassifications. Furthermore, the applicability of our method across various layers discloses distinct distributed representations over the layers, which provides deeper insights into the internal mechanisms of the deep learning model.
“…For example, in CNNs, decision regions are divided by polyhedral cones (Carlsson 2019) so that the angular difference between feature maps becomes highly related to the Configuration distance. This aligns with the empirical successes of prior work using the Cosine similarity in the feature space (Fong and Vedaldi 2018;Kim et al 2018;Bachman, Hjelm, and Buchwalter 2019;Jeon, Jeong, and Choi 2020). We plan to explore this phenomenon further in our future work.…”
Section: Analysis For the Distance Metricssupporting
Understanding intermediate representations of the concepts learned by deep learning classifiers is indispensable for interpreting general model behaviors. Existing approaches to reveal learned concepts often rely on human supervision, such as pre-defined concept sets or segmentation processes. In this paper, we propose a novel unsupervised method for discovering distributed representations of concepts by selecting a principal subset of neurons. Our empirical findings demonstrate that instances with similar neuron activation states tend to share coherent concepts. Based on the observations, the proposed method selects principal neurons that construct an interpretable region, namely a Relaxed Decision Region (RDR), encompassing instances with coherent concepts in the feature space. It can be utilized to identify unlabeled subclasses within data and to detect the causes of misclassifications. Furthermore, the applicability of our method across various layers discloses distinct distributed representations over the layers, which provides deeper insights into the internal mechanisms of the deep learning model.
“…Another work (Shen et al 2020) trains a linear classifier based on artifact-labeled data and removes artifacts by moving the latent code over the trained hyperplane. A sampling method with the trained generative boundaries was suggested to explain shared semantic information in the generator (Jeon, Jeong, and Choi 2020). Classifier-based defective internal featuremap unit identification was devised (Tousi et al 2021).…”
Section: Related Workmentioning
confidence: 99%
“…In this section, we present our main contribution, the concept of local activation and its relation with low visual fidelity for individual generations. From previous research (Bau et al 2019;Jeon, Jeong, and Choi 2020;Tousi et al 2021), we can presume that each internal featuremap unit in the generator handles a specific object (e.g., tree, glasses) for the final generation. In particular, an artifact that has low visual fidelity can also be considered as a type of object.…”
Section: Locally Activated Neurons In Gansmentioning
Despite significant improvements on the image generation performance of Generative Adversarial Networks (GANs), generations with low visual fidelity still have been observed. As widely used metrics for GANs focus more on the overall performance of the model, evaluation on the quality of individual generations or detection of defective generations is challenging. While recent studies try to detect featuremap units that cause artifacts and evaluate individual samples, these approaches require additional resources such as external networks or a number of training data to approximate the real data manifold.
In this work, we propose the concept of local activation, and devise a metric on the local activation to detect artifact generations without additional supervision.
We empirically verify that our approach can detect and correct artifact generations from GANs with various datasets. Finally, we discuss a geometrical analysis to partially reveal the relation between the proposed concept and low visual fidelity.
“…From the previous research [10] that shallow layers handle the abstract generation concepts and deeper layers handle localized information in GANs, we ablate the shallow layers from the first layer to the stopping layer l < L. To prevent the loss of semantic characteristics of a generation as pointed in Section 3.3, we adjust the magnitude of the original featuremaps instead of the simple zero ablation. Line 5 of Algorithm 1 states this soft ablation as,…”
Section: Sequential Correction Of Artifactsmentioning
Generative Adversarial Networks (GANs) have shown satisfactory performance in synthetic image generation by devising complex network structure and adversarial training scheme. Even though GANs are able to synthesize realistic images, there exists a number of generated images with defective visual patterns which are known as artifacts. While most of the recent work tries to fix artifact generations by perturbing latent code, few investigate internal units of a generator to fix them. In this work, we devise a method that automatically identifies the internal units generating various types of artifact images. We further propose the sequential correction algorithm which adjusts the generation flow by modifying the detected artifact units to improve the quality of generation while preserving the original outline. Our method outperforms the baseline method in terms of FID-score and shows satisfactory results with human evaluation.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.