What Do Deep Saliency Models Learn about Visual Attention? The supplementary materials provide additional results to complement our analyses in the main paper, and elaborate on the implementation details of our visualization method. In the main paper, we visualize the weights of different semantic categories (e.g., action, social, and scene) for saliency prediction in various scenarios. Here we provide complementary results on detailed semantics, which are used to derive the results shown in the main paper (see the listed sections below). In particular, Figure 1 shows the weights of detailed semantics for DINet [1] trained on different datasets (Section 4.2 of the main paper).

Let ϑ Rd1 and µ Rd2 be the dual variables corresponding to the d1 equality constraints and the d2 inequality constraints respectively. Let X? be an optimal solution to (SDP) and let X?FW be an optimal solution to (SDP-LSE). For ease of notation, let u= A(1)(X) b(1) andv = b(2) A(2)(X), (1) and define (bu,bv), (uFW,vFW) and (u?,v?) by substituting bX, XFW and X? respectively in (1). Upper bound on the objective. Rearranging the terms, using the duality of the `1 and ` norms, and the fact that µ? 0, gives hC, bX i hC,X?i+

Our presentation is organized for following reasons: In Section 2.3, we present the228 details of generative models, evaluated datasets, and analysis approaches (including our visualization229 tool, histogram matching attack, and human evaluation). They are independent of each other, thus230 we discuss them in parallel in the main paper. In Section 3.1, we investigate the feature extractors231 by first identifying their attention on visual semantics, followed by investigating their robustness to232 the histogram matching attack. Finally, we filter extractors that define similar representation spaces.233 These studies are gradually deepening, thus they are organized in a progressive manner.

A good metric, which promises a reliable comparison between solutions, is essential for any well-defined task. Unlike most vision tasks that have per-sample groundtruth, image synthesis tasks target generating unseen data and hence are usually evaluated through a distributional distance between one set of real samples and another set of generated samples. This study presents an empirical investigation into the evaluation of synthesis performance, with generative adversarial networks (GANs) as a representative of generative models. In particular, we make indepth analyses of various factors, including how to represent a data point in the representation space, how to calculate a fair distance using selected samples, and how many instances to use from each set. Extensive experiments conducted on multiple datasets and settings reveal several important findings. Firstly, a group of models that include both CNN-based and ViT-based architectures serve as reliable and robust feature extractors for measurement evaluation. Secondly, Centered Kernel Alignment (CKA) provides a better comparison across various extractors and hierarchical layers in one model. Finally, CKA is more sampleefficient and enjoys better agreement with human judgment in characterizing the similarity between two internal data correlations. These findings contribute to the development of a new measurement system, which enables a consistent and reliable re-evaluation of current state-of-the-art generative models. 1

For FC-individual, we train each testing assay separately with a two-layer fully-connected base learner. For FC-All, a two-layer fully connected model is trained on samples from both support setandquery setofsource assays andfrom thesupport setofthetargetassay. C.1 DrugActivityPredictionData For drug activity prediction, here we summarized the number of assays belonging to each target family: GPCR (685), Ion channel (215), Kinase (665), NHR (123), Binding (2523), Phenotypic (2299), Functional (1689), Proteinase (289),ADME (55).

The compositionality of the model improves the capacity of generalization tovarious andevenout-of-distribution tasks.