30:T162e,html`

Interpretability is increasingly becoming an important problem for the adoption of machine learning models in a variety of applications. For discriminative models, the ability for humans to interpret the rationale behind a model decision, e.g., why a prediction was made in a classification task, can help instill trust in users when applying the model downstream ${citations.get('choo2018visual')} ${citations.get('dovsilovic2018explainable')}. The manner in which humans consume interpretations of discriminative models is typically passive: the model makes a prediction, and then we provide some explanation for its prediction. In contrast, for generative models ${citations.get('jebara2012machine')} ${citations.get('goodfellow2014generative')}, the problem of interpretability is more complex, as end-users tend to be more active in using these models, e.g., generative models are maturing as a mechanism for design, ideation, and content creation within a number of domains ${citations.get('karras2019style')} ${citations.get('zhang2017stackgan')} ${citations.get('ledig2017photorealistic')} ${citations.get('perarnau2016invertible')} ${citations.get('mariani2018bagan')} ${citations.get('frid2018gan')} ${citations.get('tan2017artgan')} ${citations.get('mogren2016c')} ${citations.get('zhang2017stackgan')}. For these use cases, the ability to understand the latent semantics within a generative model is of high importance as it can benefit how one ultimately controls the model for the target applications.

Identifying distinctive semantics, i.e., disentanglement, in the generator is the predominant approach to make sense of a generative model. Disentanglement is concerned with the discovery of directions, found within the latent space of a generative model, where a single direction tends to capture a semantic, human-nameable concept, and only that concept. Such disentangled directions are ideal for controlling generative models, especically considering that a generator's latent space is usually a dense, high-dimensional vector space that is otherwise challenging to interpret. Many semantic discovery approaches have been proposed, e.g., ensuring individual components of the vector space capture distinct concepts ${citations.get('higgins2016beta')}, supervised direction finding given concept-annotated datasets ${citations.get('shen2020interfacegan')} ${citations.get('yang2021semantic')}, as well as unsupervised direction finding methods ${citations.get('voynov2020unsupervised')} ${citations.get('spingarn2020gan')} ${citations.get('harkonen2020ganspace')} ${citations.get('shen2020closed')}.

Despite the rapid progress, a fundamental issue with existing works is in evaluation: it is challenging to quantify just how disentangled a given direction is, particularly for modern models ${citations.get('brock2018large')} ${citations.get('karras2017progressive')} ${citations.get('karras2019style')} ${citations.get('goetschalckx2019ganalyze')}. Conventionally, summary statistics are proposed based on a predefined characterization of disentanglement, e.g., Eastwood & Williams ${citations.get('eastwood2018framework')} propose completeness, informativeness, and distinctiveness as desirable properties, alongside heuristic measures for each. On the other extreme, individual draws from the latent space can be shown to users for validation ${citations.get('higgins2016beta')}. Though we see value in both, we believe that a middle ground exists for improving our understanding of latent space directions, where interactive visualization can support the user in evaluating a given direction.

To this end, we propose a visual analytics approach for assessing disentanglement within generative models. Our visualization design is centered on walks in the latent space of a generative model. Specifically, a walk is defined by taking a single code (a high-dimensional vector) in the latent space, and producing a sequence of samples that lie along a given direction (another vector). At a high level, our goal is to help users understand the quality of a given direction, through visually analyzing a collection of walks, the result of taking a collection of latent space codes and producing a walk for each code. Our approach relies on a set of pre-trained image classifiers that produce continuously valued scores, indicating the extent to which an image contains a particular attribute or concept that can be easily recognized. For a single walk, and its corresponding generated images, we can obtain a sequence of scores for each attribute.

This detailed attribute-based information forms the basis for our visualization design and various analyses that we aim to support. Specifically, if there exists a consistent tren across walks for a given attribute, e.g., the attribute scores monotonically increase/decrease, we believe the direction is representative of the attribute. On the other hand, there may exist attribute correlation with respect to the trends, e.g., for face images, attributes representative of smiling and cheekbone are likely to be correlated - when one increases, the other is likely to increase as well. Identifying when correlation exists, and whether or not the correlation is sensible, is critical for determining the quality of a direction, above and beyond merely identifying distinct factors of variation. Furthermore, not all codes in the latent space need to have an equivalent response to the direction, giving rise to different attribute trends. This particular matter is exemplified in Fig. 1.

`31:T59c,html`

Determining groups of walks that have similar trends thus becomes essential to obtain a more comprehensive understanding of the direction. Our visualization design is intended to help the user probe a given direction in support of these types of analyses, through a set of linked and coordinated views that depict attribute scores, their trends, and more general similarity of walks.

We conduct use cases through visually exploring StyleGAN ${citations.get('karras2019style')} using a generative model of high-resolution faces ${citations.get('karras2017progressive')}, alongside facial attribute classifiers ${citations.get('shen2020interfacegan')}. Our interface allows for the analysis of arbitrary directions, where we investigate directions produced from supervised methods ${citations.get('shen2020interfacegan')}, unsupervised methods ${citations.get('shen2020closed')}, as well as simple, vector arithmetic of latent codes. We contribute two main use cases, enabling a deeper understanding of directions in generative models:

We show directions computed from unsupervised and supervised methods that capture, at face value, the same predominant semantic property, nevertheless differ significantly in other properties.
We demonstrate an exploration of bias in directions, wherein continuous semantics that are correlated across walks highlight a learned bias in the model

`32:T1b78,html`

2. Related Work

There are several areas of research that are relevant to the proposed approach. Major bodies of related work revolve around semantic discovery in latent space that mostly originate from the machine learning community, as well as a variety of visual analytic approaches for exploring latent spaces that arise from the visualization community.

2.1 Supervised semantic discovery

The ability to identify and explore high-level concepts in latent representation is crucial for making sense of neural networks. Often, we can identify a vector direction in the latent space that corresponds to clear semantics. From the early discovery that semantics can be obtained through vector arithmetic in word embedding spaces ${citations.get('mikolov2013distributed')}, to later work that confirmed meaningful concepts can also be encoded as a linear direction in the latent space of GANs ${citations.get('radford2015unsupervised')} and CNNs ${citations.get('kim2018interpretability')}, a variety of approaches have been proposed to uncover these vector direction through supervised and unsupervised means. Supervised methods typically rely on external models or known transformations to identify meaningful directions. In TCAV ${citations.get('kim2018interpretability')}, example images of the concept of interests are gathered in order to train a simple linear model to identify the concept direction in the latent space. The "interface" work ${citations.get('shen2020interfacegan')} employs an external model that assesses attribute scores of an image. The semantic directions are derived from SVM classifiers that were trained on the GAN-synthesized random images using labels generated by a pre-trained classifier. Ganalyze ${citations.get('goetschalckx2019ganalyze')} employs an external model that assesses the memorability of the generated image. They find directions such that for a given starting image, a manipulated image's memorability score results in a large change. However, they do not consider a wider variety of different semantics. The GAN steerability work ${citations.get('jahanian2019steerability')} finds a direction by first sampling an image using a randomly sampled latent code, applying an edit such as zoom or rotation to the generated image, and recording a latent code that produces an image that minimizes the loss between itself and the previous image. However, the method can be computationally heavy and only considers coarse image manipulations.

2.2 Unsupervised semantic discovery

Despite the effectiveness of supervised methods, they are limited by the type and variety of concepts that can be identified, and more often than not we may not have labels to utilized the supervised method to begin with. Voynov et al. introduce an unsupervised approach ${citations.get('voynov2020unsupervised')}, relying on the intuition that meaningful concepts are often more disentangled with each other and therefore more predictable. The method trains a special component called "reconstructor" that takes two images generated by latent codes along a vector direction (a column in a matrix that is part of the trainable parameter). and predicts which vector/concept induces the difference between the images and the magnitude of the change. GANSpace ${citations.get('harkonen2020ganspace')} identifies meaningful directions in a GAN's latent space by applying PCA to latent codes in subsequent layers, subsequently mapping them back into the original latent space. Shen et al. ${citations.get('shen2020closed')} propose a closed-form approach, where decomposition of a given model's weight matrix is used for identifying directions that result in large changes in the model, and consequently, in the output images. Among both these supervised and unsupervised approaches, many introduce metrics for evaluating the quality of obtaining directions that are based on ground truth obtained from supervised methods, the optimization objectives, and human evaluation. However, all these approaches focus on producing a single number instead of providing more granularity in their evaluation, i.e., where and how did some of these methods fail, which can be crucial to uncover the cause of the failure and possible avenues for improvement. The proposed work aims to address this gap in existing works.

2.3 Disentanglement in latent space

Most semantic discovery methods approach the problem from a post hoc perspective, however, we can also directly build a latent space in which each dimension has a distinctive concept. Many approaches have been proposed, such as beta-VAE ${citations.get('higgins2016beta')} and InfoGAN ${citations.get('chen2016infogan')}, for obtaining a disentangled latent space. However, defining and estimating whether different factors are disentangled is a non-trivial task. In ${citations.get('higgins2018towards')}, Higgins et al. aim to work toward a formal definition. Moreover, the inherent bias in the data may also contribute to challenges in understanding disentanglement, and the fairness of disentangled representations is questioned ${citations.get('locatello2019fairness')}. One analysis goal of the proposed method is to provide a visual analytic centric alternative for examining and understanding the disentanglement between the discovered semantic directions, whether it is from a supervised or unsupervised method (or even a disentangled latent dimension).

2.4 Visual exploration of latent space and GAN

Besides methods focusing on purely computational methods, visualization approaches provide powerful and flexible alternatives for exploring concepts in latent spaces. In the latent space cartography ${citations.get('liu2019latent')} work, Liu et al. introduced a visualization system that maps and compares meaningful semantic dimensions within latent spaces. For natural language processing, the word embedding latent space are explored by several visualization works that focus on various aspects, from examining semantics encoded in vector directions ${citations.get('liu2017visual')} to comparing the embeddings themselves ${citations.get('heimerl2020embcomp')}. Several visual analytics approaches specifically focused on visualizing GAN and its training process. The GANlab ${citations.get('kahng2018gan')} provides a playground-like environment for understanding how GANs work. By leveraging a 2D function estimation problem, the system allows user to see how GANs evolve during training and explore how different hyperparameters affect the output. GANViz${citations.get('arnout2019visual')} aims to shed light on the complex training dynamic among different components of a GAN, to help domain experts evaluate and potentially improve their models. Our work contributes towards similar, higher-level aims of understanding GANs, but is instead focused on assessing directions in GAN latent spaces.

`33:Tc33,html`

The central goal of our work is to support users in obtaining a deeper understanding of disentanglement, provided a direction within the latent space of a generative model. We want to help users assess how good the direction is, as an avenue for generating meaningful factors of variation, provided an arbitrary code in the latent space. Forming an accepted, understood definition of disentanglement remains challenging; nevertheless, a number of works have introduced properties of disentangled representations ${citations.get('eastwood2018framework')} ${citations.get('higgins2018towards')} ${citations.get('locatello2019fairness')}, and corresponding heuristics for measuring these properties. We are inspired by these prior works, but instead approach evaluating disentanglement from an exploratory perspective.

More specifically, we would like the user to obtain a better understanding of disentanglement through assessing the following properties:

[O1] Distinctiveness : how many factors of variation exist for a given direction. A greater variation tends to indicate the direction is not distinct.

[O2] Consistency : how predictable the variation is. A direction is considered consistent if the same factors of variation are present across most codes in the latent space.

[O3] Informativeness : the extent to which an end-user can recognize, e.g. easily name, what the direction represents.

Limiting our analysis to only the generator, the representations that it learns, and its image output, poses significant challenges for addressing the above objectives. This is due to the complexity of the data, once provided a direction, e.g. for a walk we have a sequence of latent codes and their corresponding generator representations/images. To support our analysis, we rely on pretrained attribute classifiers, ones that score an image for a prescribed, semantic concept, e.g. various facial attributes. The use of attributes enables us to simplify a walk in the latent space to a sequence of attribute scores, similar to prior work ${citations.get('shen2020interfacegan')}. However, in our work we aim to use attribute classifiers, alongside the generator representations, to obtain a more comprehensive understanding of disentanglement. Given this information, our visualization design is driven by the following set of user tasks aimed at satisfying the above objectives:

[T1] Distinguish salient attributes, e.g. those that have similar trends across codes, from less salient attributes. A small number of salient attributes suggests a distinct direction.[O1]

[T2] Identify trends in attributes with respect to walks, e.g. for a given walk, which attributes monotonically increase/decrease in score. Such trends signal consistency in directions. [O2]

[T3] Analyze correlations between attributes. A set of attributes might change in consistent ways, but may be unrelated to one another, indicative of uninformative directions. [O3]

`34:T127d,html`

We focus our analysis on the StyleGAN ${citations.get('karras2019style')} model, due to its ubiquity within the machine learning community, and its capability to produce realistic, high-resolution images. StyleGAN synthesizes images by first drawing a high-dimensional vector, whose components are independently sampled from a (truncated) normal distribution, and feeding this vector through a multi-layer perceptron (MLP), yielding what is colloquially known as the ${tex`\mathcal{W}`} space ${citations.get('karras2019style')}. Unless otherwise stated, this latent space is the focus of our analysis, where we denote a sample in the latent space as a code, a ${tex`d`}-dimensional vector (c.f. Fig. 2(a)). The code is fed through a series of convolutional layers (c.f. Fig. 2(b)), whose activations are modulated by style-based features, to ultimately produce an image (c.f. Fig. 2(c)).

We are further provided a direction in the latent space, denoted as a unit-norm vector ${tex`\mathbf{a} \in \mathcal{W}`}. Given a code ${tex`\mathbf{z} \in \mathcal{W}`}, we sample a sequence of ${tex`m`} equal-spaced steps along the direction, centered at a direction-neutral transformation of the code, denoted ${tex`\mathbf{\tilde{z}} \in \mathcal{W}`}, found by projecting the original code onto the orthogonal complement of the direction: ${tex.block`\mathbf{\tilde{z}} = (\mathcal{I} - \mathbf{a}\mathbf{a}^{\top})\mathbf{z}`}

This provides us a sequence of codes ${tex`\mathbf{z}(t) \in \mathcal{W}`} where t indexes over steps. We denote this sequence as a walk in the latent space, and we may obtain a sequence of images by feeding each code through the rest of the StyleGAN model. Ideally, the further that a code lies along a direction ${tex`\mathbf{a}`}, relative to the neutral code ${tex`\mathbf{\tilde{z}} \in \mathcal{W}`}, then a single, recognizable concept will become more prominent in the image. The converse should also hold for codes that point in the opposite direction, e.g. if a direction captures the facial property "closed eyes", then such codes result in faces with eyes wide open.

We are interested in understanding a collection of walks, formed by sampling codes in the latent space, and for each, generating a sequence of samples as described above. To aid in the analysis of walks, we assume the existence of a set of pretrained classifiers optimized to predict the presence of attributes - semantic, recognizable concepts that describe an image. We use the raw outputs of these pretrained models, rather than their binarized classification decisions, to obtain continuously-valued scores and thus a means of quantifying the presence/absence of an attribute. Thus, for each code along a given walk, we take its synthesized image and run the attribute classifier to obtain a score for this image (c.f. Fig. 2(d)).

Although informative, a sequence of raw scores for each attribute is not the most concise way of describing a direction's relationship with respect to a given attribute, particularly for analyzing a large collection of walks, and comparing across attributes. Thus, we summarize a sequence of attribute scores by fitting a linear regression model to a given walk's sequence of attribute scores. Our assumption is that attribute scores will largely either monotonically increase, decrease, or remain constant, as the walk progresses along a direction, indicating a salient change (or lack of change) in the attribute. Specifically, the slope of the resulting model suggests the type of trend, while the ${tex`R^2`} coefficient indicates a goodness of fit (confidence) - we use both of these values to summarize a sequence of attribute scores.

Last, in order to ensure that our analysis is not biased by the predefined collection of attributes, we would also like to have a more general notion of similarity between walks, in order to better understand their distribution. To this end, we extract representations from the generator as a proxy for similarity - indeed, most (un)supervised walk finding methods are based on features of the generator, and so we expect this to encode semantically-meaningful information. Specifically, given a walk, for each of its codes we take the convolutional output at a given layer in the generator, typically a layer close to the latent space. For each output we perform max spatial pooling to obtain a vector, followed by concatenating all of the walk's vectors to obtain a feature representation for the walk. Last, we perform UMap ${citations.get('mcinnes2018umap')} to obtain a 2D projection for the set of walks, e.g. each 2D position will correspond to a single walk.

`35:Ta49,html`

The individual views are linked and coordinated through a set of interactions that the user may perform (c.f. Fig. 6).

The user may brush over the regression scatterplot view, in order to select attributes of walks that have a range of slopes and goodness-of-fits that they may want to investigate. Upon brushing, walks in the UMap projection that belong to the user's selection are highlighted, e.g. walks that contain a more general similarity. In the attribute summarization view, line marks corresponding to the selected items (walk and attribute) are highlighted, and an attribute's histogram is updated with a superimposition based on that attribute's set of walks in the selection. As an example use, brushing points that have high ${tex`R^2`}, and observing trends in the attribute scores in the summarization view, can help the user assess the informativeness of the provided direction [T3].

The user may also brush in the UMap projection, prompting linked updates to the regression and summarization views in accordance with selected walks, across all attributes. This interaction is useful for determining whether groups of similar walks are consistent in their attribute score [T2].

In hovering over the attribute text labels in the summarization view, linked updates are made to the regression scatterplot by highlighting all walks that correspond to the attribute, while in the UMap projection we update the appearance of all circles through the selected attribute's set of slopes and ${tex`R^2`} coefficients. Specifically, slope magnitude is encoded by size, the circle's color is updated based on the slope sign (consistent with the color scheme for the histogram), and the ${tex`R^2`} coefficient is encoded via stroke thickness of the circle. Linked hovering is useful as another view on consistency, e.g. verifying groups of walks that are in close proximity in the UMap projection also have related, and confident, trends in a given attribute [T2].

Last, clicking on a point in the UMap projection will bring up the sequence of synthesized images that correspond to the selected walk. Based on findings made through using the interface, inspecting individual images of walks can support the user in verifying hypotheses they might have formed. For instance, in conjunction with hovering the user may discover a group of walks in the UMap projection that have similar trends across several attributes, and subsequently, investigate the image sequences for these walks to verify the distinctiveness of the direction [T1].

`36:T7f5,html`

In this section we show how our visualization can be used to perform a variety of analyses for better understanding directions in GANs. Our current interface supports two GAN models:

StyleGAN optimized on the FFHQ dataset ${citations.get('karras2017progressive')}, using the CelebA dataset ${citations.get('liu2015deep')} for attribute analysis.
StyleGAN optimized on the outdoor images in the Places 2 dataset ${citations.get('zhou2017places')}, using the Transient Attributes ${citations.get('Laffont14')} dataset for attribute analysis. Artificial objects and humans were filtered with Yolo v4 ${citations.get('bochkovskiy2020yolov4')} to focus on the natural scenery.

Furthermore, our interface supports three different types of direction finding methods:

Vector arithmetic: we manually construct directions by identifying a pair of latent codes whose images differ by a semantic concept, and take their difference. For robustness, we average three latent codes with the attribute and another three without the attribute.
Unsupervised direction finding: we use the approach of SEFA ${citations.get('shen2020closed')} to find directions directly from the StyleGAN style layer mappings, e.g. directions that result in large change in the linear maps. This provides us with directions that depend on (a) the layer under investigation and (b) their overall influence on the style layer.
Supervised direction finding: we use the approach of Shen et al. ${citations.get('shen2020interfacegan')} to find directions in the latent space that are discriminative of predefined attributes, measured in the corresponding synthesized images. The attributes we use for analysis are also used for supervised direction finding.

In our analyses we compare directions found for a given supervised method, comparison between different direction-finding methods, and assess potential bias that exists within a given direction.

`37:T5d1,html`

We first assess directions with respect to the main properties of disentanglement: distinctiveness, consistency, and informativeness.

We first show how to verify whether a direction is distinct and consistent. We expect a direction to be distinct if it its factors of variation can be described by a small set of attributes, and for it to be consistent if most attribute scores monotonically increase/decrease across all codes. To help us understand these properties, we take two directions found through the supervised direction-finding method that are based on "Smiling" and "No Beard" attributes.

Fig. 7 depicts a comparison between these two directions. Within the Regression Plot, we brush all walks that have a high ${tex`R^2`} coefficient, in order to limit our analysis to regression fits that are confident. Through this interaction, we observe that the "Smiling" direction can be described by a small number of attributes, whereas the "No Beard" direction results in a larger number of attributes (see annotated red rectangles), indicating that the "Smiling" direction is more distinct. Furthermore, consistency of the "Smiling" direction can be verified via the linked highlight of the histogram plots, indicating a high concentration of large slope values across the most confident regression fits - this is further indicated by the group of points in the upper-right quadrant of the Regression plot, namely those fits that are confidently increasing. `38:T492,html`

Outside of verifying directions found through supervised methods, we can also use our interface to assess directions found through simpler schemes, namely vector arithmetic. For our scene-based GAN, we have built a direction based on flatness in natural formations, e.g. from mountains to open plains. Directions built from vector arithmetic are more susceptible to larger variation due to the more localized, hand-crafted approach, hence here we wish to see what types of "flatness" exist in the generative model.

Fig. 9 shows how one can use our interface to discover different variations in flatness. In the histogram plots we can observe trends in attributes regarding "ice" and "moist" - indeed, we find that there is a tendency for land formations to result in rivers/lakes, some iced over. Highlighting the "ice" attribute, we can see groups of walks in the UMap projection that cluster together regarding a consistent increase in ice; we highlight a single walk's image sequence indicated by the arrow in the scatterplot. In general, we find this grouping of walks to consist of mountains that change to icy, flat, landscapes.

`39:T6ec,html`

6.2 Comparing Direction Finding Methods

Here we study the differences that may exist between direction-finding methods, given the same predominant variation found in both. Namely, we consider the FFHQ GAN, and compare the "Smiling" direction found through the supervised and unsupervised method. For the unsupervised method, we manually selected the direction via manual inspection of directions found at different layers.

Fig. 10(a) shows our interface for the direction found via the unsupervised method. We observe consistent trends in "High Cheekbones", "Smiling", and "Bags Under Eyes" for confident regression fits (see: brush in the Regression Plot), further supporting our identification of this direction as "Smiling". Additionally, we find that "High Cheekbones" and "Bags Under Eyes" are attributes that have confident fits in the supervised direction as well. Hence, we use both of these attributes as a basis for comparison (Fig. 10 - right). Specifically, we highlight these attributes for both directions, and show the resulting scatterplots.

From here, we can draw two conclusions. First, the supervised method finds a more salient direction for smiling, given the correlated attributes have more consistent trends (Fig. 10(c)) in comparison with the unsupervised method (Fig. 10(b)), e.g. walks in the UMap projection tend to have a larger, red circle with bold stroke, indicative of confident and increasing trends. Secondly, although the unsupervised direction is less salient, the grouping in the UMap projection is helpful for identifying what portions of the latent space give us a salient change in smiling, e.g. the group of points in the center-right portion of Fig. 10(b) correspond to children smiling.

`3a:T527,html`

Since GANs learn to generate images that follow the data distribution of the provided dataset, they are prone to biases present in the dataset. Our interface can be used to study these biases - how semantics are represented in the latent space. We study bias using two supervised directions - age and gender - using interactive examples, supplied with default interactions.

Bias can be studied by examining consistent attributes for a direction. Therefore, we use the same interaction used in assessing a direction - brushing the regression plot in a well-fit region. Once we identify consistent attributes, we can reason whether these attributes can be explained.

Fig. 11 allows one to interactively explore bias for the "Young" direction. We can observe several attributes that are unrelated to youth, e.g. "Wearing Lipstick", "Male". Indeed, we find that as we walk along the young direction, many walks end up with female gender. This suggests a gender bias in the dataset: more older men than older women. Further, we also find that older people end up with glasses along this direction. A direction that exclusively varies along age should not be entangled with glasses, further suggesting a bias in the dataset wherein older people tend to wear eyeglasses for corrective vision.

`3b:T9a7,html`

In this work we have shown how our visualization can enable users to understand the latent space of generative models from different angles. It can be used to discover directions and determine direction saliency, in turn assessing the distinctiveness of a direction. It can also be used to determine if multiple attributes are consistently related, and if there exists correlation in the related attributes. Last, informativeness of a direction can easily be assessed through interactions that let users quickly reason about trends across attributes. The visualization can also be used to compare different directions, permitting the understanding of how different walking methods treat similar directions.

Beyond studying well-defined disentaglement properties, we showed that our visualization can be used to compare different walking methods, for which researchers studying direction finding methods may benefit from. Studying bias in the latent space is another use-case of our visualization, which benefits data scientists and general GAN community. Finally, we take a step further into the generator representation by linking projected convolutional activations with namable attributes.

One limitation of our research is that the quality of the latent space study hinges on the quality of attribute classifiers. Studying various aspects of a direction would be difficult without semantic attribute-based classifiers. We plan on studying how to use activations in the generator and discriminator to free the dependency on attributes. The second limitation comes from the fact that our visualization is primarily built to understand walks given a single direction. Although not impossible, it is more difficult to compare multiple directions at the same time. The ability to compare different directions is an important step in assessing disentanglement more broadly, e.g. ensuring a set of directions are distinct, yet completely describe the generative model.

Moving forward, we plan to redesign our visualization to enable a comparison between multiple directions, as opposed to multiple attributes. This lets researchers understand a single latent code in-depth to reveal whether the latent code might fail to respond to directions, or has a strong property that cannot be altered via walking along directions. We believe such an interface can become useful for the purposes of image editing and content creation.

`3c:T44d3,bibtex = [ ` @article{choo2018visual, title={Visual analytics for explainable deep learning}, author={Choo, Jaegul and Liu, Shixia}, journal={IEEE computer graphics and applications}, volume={38}, number={4}, pages={84--92}, year={2018}, publisher={IEEE} }`, ` @inproceedings{dovsilovic2018explainable, title={Explainable artificial intelligence: A survey}, author={Do{\v{s}}ilovi{\'c}, Filip Karlo and Br{\v{c}}i{\'c}, Mario and Hlupi{\'c}, Nikica}, booktitle={2018 41st International convention on information and communication technology, electronics and microelectronics (MIPRO)}, pages={0210--0215}, year={2018}, organization={IEEE} }`, ` @inproceedings{tan2017artgan, title={ArtGAN: Artwork synthesis with conditional categorical GANs}, author={Tan, Wei Ren and Chan, Chee Seng and Aguirre, Hern{\'a}n E and Tanaka, Kiyoshi}, booktitle={2017 IEEE International Conference on Image Processing (ICIP)}, pages={3760--3764}, year={2017}, organization={IEEE} }`, // Art ` @article{mogren2016c, title={C-RNN-GAN: Continuous recurrent neural networks with adversarial training}, author={Mogren, Olof}, journal={arXiv preprint arXiv:1611.09904}, year={2016} }`, // Music // ` // @article{xiao2017dna, // title={DNA-GAN: Learning disentangled representations from multi-attribute images}, // author={Xiao, Taihong and Hong, Jiapeng and Ma, Jinwen}, // journal={arXiv preprint arXiv:1711.05415}, // year={2017} // }`, // Disentanglement Learning ` @article{higgins2016beta, title={beta-vae: Learning basic visual concepts with a constrained variational framework}, author={Higgins, Irina and Matthey, Loic and Pal, Arka and Burgess, Christopher and Glorot, Xavier and Botvinick, Matthew and Mohamed, Shakir and Lerchner, Alexander}, year={2016} }`, // Mostly employ qualitative analysis, but talk about quantitative analysis. // ` // @article{wu2020stylespace, // title={StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation}, // author={Wu, Zongze and Lischinski, Dani and Shechtman, Eli}, // journal={arXiv preprint arXiv:2011.12799}, // year={2020} // } // `, // Supervised direction ` @inproceedings{eastwood2018framework, title={A framework for the quantitative evaluation of disentangled representations}, author={Eastwood, Cian and Williams, Christopher KI}, booktitle={International Conference on Learning Representations}, year={2018} } `, // Disentanglement, completeness // ` // @article{desjardins2012disentangling, // title={Disentangling factors of variation via generative entangling}, // author={Desjardins, Guillaume and Courville, Aaron and Bengio, Yoshua}, // journal={arXiv preprint arXiv:1210.5474}, // year={2012} // } // `, // Disentanglement // ` // @article{zhou2020evaluating, // title={Evaluating the disentanglement of deep generative models through manifold topology}, // author={Zhou, Sharon and Zelikman, Eric and Lu, Fred and Ng, Andrew Y and Carlsson, Gunnar and Ermon, Stefano}, // journal={arXiv preprint arXiv:2006.03680}, // year={2020} // } // `, ` @inproceedings{chen2016infogan, title={Infogan: Interpretable representation learning by information maximizing generative adversarial nets}, author={Chen, Xi and Duan, Yan and Houthooft, Rein and Schulman, John and Sutskever, Ilya and Abbeel, Pieter}, booktitle={Proceedings of the 30th International Conference on Neural Information Processing Systems}, pages={2180--2188}, year={2016} } `, ` @inproceedings{liu2015deep, title={Deep learning face attributes in the wild}, author={Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou}, booktitle={Proceedings of the IEEE international conference on computer vision}, pages={3730--3738}, year={2015} } `, ` @article{goodfellow2014generative, title={Generative adversarial networks}, author={Goodfellow, Ian J and Pouget-Abadie, Jean and Mirza, Mehdi and Xu, Bing and Warde-Farley, David and Ozair, Sherjil and Courville, Aaron and Bengio, Yoshua}, journal={arXiv preprint arXiv:1406.2661}, year={2014} } `, ` @article{radford2015unsupervised, title={Unsupervised representation learning with deep convolutional generative adversarial networks}, author={Radford, Alec and Metz, Luke and Chintala, Soumith}, journal={arXiv preprint arXiv:1511.06434}, year={2015} } `, // DCGAN ` @article{brock2018large, title={Large scale GAN training for high fidelity natural image synthesis}, author={Brock, Andrew and Donahue, Jeff and Simonyan, Karen}, journal={arXiv preprint arXiv:1809.11096}, year={2018} } `, // Big GAN ` @article{karras2017progressive, title={Progressive growing of gans for improved quality, stability, and variation}, author={Karras, Tero and Aila, Timo and Laine, Samuli and Lehtinen, Jaakko}, journal={arXiv preprint arXiv:1710.10196}, year={2017} } `, // Progressive GAN ` @inproceedings{karras2019style, title={A style-based generator architecture for generative adversarial networks}, author={Karras, Tero and Laine, Samuli and Aila, Timo}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={4401--4410}, year={2019} } `, // Style GAN ` @inproceedings{goetschalckx2019ganalyze, title={Ganalyze: Toward visual definitions of cognitive image properties}, author={Goetschalckx, Lore and Andonian, Alex and Oliva, Aude and Isola, Phillip}, booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision}, pages={5744--5753}, year={2019} } `, // Supervised approach, modifies memorability which is a higher-level concept ` @article{jahanian2019steerability, title={On the" steerability" of generative adversarial networks}, author={Jahanian, Ali and Chai, Lucy and Isola, Phillip}, journal={arXiv preprint arXiv:1907.07171}, year={2019} } `, // Supervised approach, tries to evaluate walks by quantifying them ` @article{shen2020interfacegan, title={Interfacegan: Interpreting the disentangled face representation learned by gans}, author={Shen, Yujun and Yang, Ceyuan and Tang, Xiaoou and Zhou, Bolei}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, year={2020}, publisher={IEEE} } `, // Supervised approach, build 500K synthetic dataset, 10K for training, SVM. // ` // @inproceedings{yin2017towards, // title={Towards large-pose face frontalization in the wild}, // author={Yin, Xi and Yu, Xiang and Sohn, Kihyuk and Liu, Xiaoming and Chandraker, Manmohan}, // booktitle={Proceedings of the IEEE international conference on computer vision}, // pages={3990--3999}, // year={2017} // } // `, // Supervised, for face pose. Face pose for single identity can be used to build specific datasets. ` @article{yang2021semantic, title={Semantic hierarchy emerges in deep generative representations for scene synthesis}, author={Yang, Ceyuan and Shen, Yujun and Zhou, Bolei}, journal={International Journal of Computer Vision}, volume={129}, number={5}, pages={1451--1466}, year={2021}, publisher={Springer} } `, // Supervised, argues about hierarchically changing attributes in different layers ` @inproceedings{voynov2020unsupervised, title={Unsupervised discovery of interpretable directions in the gan latent space}, author={Voynov, Andrey and Babenko, Artem}, booktitle={International Conference on Machine Learning}, pages={9786--9796}, year={2020}, organization={PMLR} } `, // Unsupervised ` @article{spingarn2020gan, title={GAN Steerability without optimization}, author={Spingarn-Eliezer, Nurit and Banner, Ron and Michaeli, Tomer}, journal={arXiv preprint arXiv:2012.05328}, year={2020} } `, // Unsupervised ` @article{harkonen2020ganspace, title={Ganspace: Discovering interpretable gan controls}, author={H{\"a}rk{\"o}nen, Erik and Hertzmann, Aaron and Lehtinen, Jaakko and Paris, Sylvain}, journal={arXiv preprint arXiv:2004.02546}, year={2020} } `, // Unsupervised ` @article{shen2020closed, title={Closed-form factorization of latent semantics in gans}, author={Shen, Yujun and Zhou, Bolei}, journal={arXiv preprint arXiv:2007.06600}, year={2020} }`, // Unsupervised SeFA // ` // @misc{cherepkov2021navigating, // title={Navigating the GAN Parameter Space for Semantic Image Editing}, // author={Anton Cherepkov and Andrey Voynov and Artem Babenko}, // year={2021}, // eprint={2011.13786}, // archivePrefix={arXiv}, // primaryClass={cs.LG} // } // `, // Do not modify latent codes, but modifies GAN parameters. => New trend ` @article{bau2018gan, title={Gan dissection: Visualizing and understanding generative adversarial networks}, author={Bau, David and Zhu, Jun-Yan and Strobelt, Hendrik and Zhou, Bolei and Tenenbaum, Joshua B and Freeman, William T and Torralba, Antonio}, journal={arXiv preprint arXiv:1811.10597}, year={2018} }`, // GAN Dissection. Shows that intermediate neurons convey meanful attribute manipulations. ` @article{frid2018gan, title={GAN-based synthetic medical image augmentation for increased CNN performance in liver lesion classification}, author={Frid-Adar, Maayan and Diamant, Idit and Klang, Eyal and Amitai, Michal and Goldberger, Jacob and Greenspan, Hayit}, journal={Neurocomputing}, volume={321}, pages={321--331}, year={2018}, publisher={Elsevier} } `, // Medical imaging GAN data augmentation ` @article{mariani2018bagan, title={Bagan: Data augmentation with balancing gan}, author={Mariani, Giovanni and Scheidegger, Florian and Istrate, Roxana and Bekas, Costas and Malossi, Cristiano}, journal={arXiv preprint arXiv:1803.09655}, year={2018} } `, // Data augmentation ` @misc{zhang2017stackgan, title={StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks}, author={Han Zhang and Tao Xu and Hongsheng Li and Shaoting Zhang and Xiaogang Wang and Xiaolei Huang and Dimitris Metaxas}, year={2017}, eprint={1612.03242}, archivePrefix={arXiv}, primaryClass={cs.CV} } `, // Content creation - text to image. ` @misc{ledig2017photorealistic, title={Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network}, author={Christian Ledig and Lucas Theis and Ferenc Huszar and Jose Caballero and Andrew Cunningham and Alejandro Acosta and Andrew Aitken and Alykhan Tejani and Johannes Totz and Zehan Wang and Wenzhe Shi}, year={2017}, eprint={1609.04802}, archivePrefix={arXiv}, primaryClass={cs.CV} } `, // Super resolution ` @misc{perarnau2016invertible, title={Invertible Conditional GANs for image editing}, author={Guim Perarnau and Joost van de Weijer and Bogdan Raducanu and Jose M. Álvarez}, year={2016}, eprint={1611.06355}, archivePrefix={arXiv}, primaryClass={cs.CV} } `, // Image editing ` @article{mikolov2013distributed, title={Distributed representations of words and phrases and their compositionality}, author={Mikolov, Tomas and Sutskever, Ilya and Chen, Kai and Corrado, Greg and Dean, Jeffrey}, journal={arXiv preprint arXiv:1310.4546}, year={2013} }`, // Vector arithmetic, NLP // ` // @misc{he2015deep, // title={Deep Residual Learning for Image Recognition}, // author={Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun}, // year={2015}, // eprint={1512.03385}, // archivePrefix={arXiv}, // primaryClass={cs.CV} // } // `, // ResNet // ` // @incollection{NEURIPS2019_9015, // title = {PyTorch: An Imperative Style, High-Performance Deep Learning Library}, // author = {Paszke, Adam and Gross, Sam and Massa, Francisco and Lerer, Adam and Bradbury, James and Chanan, Gregory and Killeen, Trevor and Lin, Zeming and Gimelshein, Natalia and Antiga, Luca and Desmaison, Alban and Kopf, Andreas and Yang, Edward and DeVito, Zachary and Raison, Martin and Tejani, Alykhan and Chilamkurthy, Sasank and Steiner, Benoit and Fang, Lu and Bai, Junjie and Chintala, Soumith}, // booktitle = {Advances in Neural Information Processing Systems 32}, // editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett}, // pages = {8024--8035}, // year = {2019}, // publisher = {Curran Associates, Inc.}, // url = {http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf} // } // `, // PyTorch // ` // @misc{kingma2017adam, // title={Adam: A Method for Stochastic Optimization}, // author={Diederik P. Kingma and Jimmy Ba}, // year={2017}, // eprint={1412.6980}, // archivePrefix={arXiv}, // primaryClass={cs.LG} // } // `, // Adam Optimizer ` @article{liu2017visual, title={Visual exploration of semantic relationships in neural word embeddings}, author={Liu, Shusen and Bremer, Peer-Timo and Thiagarajan, Jayaraman J and Srikumar, Vivek and Wang, Bei and Livnat, Yarden and Pascucci, Valerio}, journal={IEEE transactions on visualization and computer graphics}, volume={24}, number={1}, pages={553--562}, year={2017}, publisher={IEEE} }`, ` @article{heimerl2020embcomp, title={embcomp: Visual interactive comparison of vector embeddings}, author={Heimerl, Florian and Kralj, Christoph and Moller, Torsten and Gleicher, Michael}, journal={IEEE Transactions on Visualization and Computer Graphics}, year={2020}, publisher={IEEE} } `, // NLP Attention visualization ` @article{kahng2018gan, title={Gan lab: Understanding complex deep generative models using interactive visual experimentation}, author={Kahng, Minsuk and Thorat, Nikhil and Chau, Duen Horng and Vi{\'e}gas, Fernanda B and Wattenberg, Martin}, journal={IEEE transactions on visualization and computer graphics}, volume={25}, number={1}, pages={310--320}, year={2018}, publisher={IEEE} }`, // GAN Viz -- training process ` @article{arnout2019visual, title={Visual Evaluation of Generative Adversarial Networks for Time Series Data}, author={Arnout, Hiba and Kehrer, Johannes and Bronner, Johanna and Runkler, Thomas}, journal={arXiv preprint arXiv:2001.00062}, year={2019} }`, // GAN direction ` @article{mcinnes2018umap, title={Umap: Uniform manifold approximation and projection for dimension reduction}, author={McInnes, Leland and Healy, John and Melville, James}, journal={arXiv preprint arXiv:1802.03426}, year={2018} }`, // Umap ` @article{locatello2019fairness, title={On the fairness of disentangled representations}, author={Locatello, Francesco and Abbati, Gabriele and Rainforth, Tom and Bauer, Stefan and Sch{\"o}lkopf, Bernhard and Bachem, Olivier}, journal={arXiv preprint arXiv:1905.13662}, year={2019} } `, ` @article{higgins2018towards, title={Towards a definition of disentangled representations}, author={Higgins, Irina and Amos, David and Pfau, David and Racaniere, Sebastien and Matthey, Loic and Rezende, Danilo and Lerchner, Alexander}, journal={arXiv preprint arXiv:1812.02230}, year={2018} } `, ` @article {Laffont14, title = {Transient Attributes for High-Level Understanding and Editing of Outdoor Scenes}, author = {Pierre-Yves Laffont and Zhile Ren and Xiaofeng Tao and Chao Qian and James Hays}, journal = {ACM Transactions on Graphics (proceedings of SIGGRAPH)}, volume = {33}, number = {4}, year = {2014} }`, // Transient Attributes ` @article{zhou2017places, title={Places: A 10 million Image Database for Scene Recognition}, author={Zhou, Bolei and Lapedriza, Agata and Khosla, Aditya and Oliva, Aude and Torralba, Antonio}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, year={2017}, publisher={IEEE} }`, // Places ` @misc{bochkovskiy2020yolov4, title={YOLOv4: Optimal Speed and Accuracy of Object Detection}, author={Alexey Bochkovskiy and Chien-Yao Wang and Hong-Yuan Mark Liao}, year={2020}, eprint={2004.10934}, archivePrefix={arXiv}, primaryClass={cs.CV} }`, // Yolo v4 ` @inproceedings{liu2019latent, title={Latent space cartography: Visual analysis of vector space embeddings}, author={Liu, Yang and Jun, Eunice and Li, Qisheng and Heer, Jeffrey}, booktitle={Computer Graphics Forum}, volume={38}, number={3}, pages={67--78}, year={2019}, organization={Wiley Online Library} }`, //latent space cartography // Book ` @book{jebara2012machine, title={Machine learning: discriminative and generative}, author={Jebara, Tony}, volume={755}, year={2012}, publisher={Springer Science \& Business Media} }`, // ` // @inproceedings{wang2020scanviz, // title={SCANViz: Interpreting the symbol-concept association captured by deep neural networks through visual analytics}, // author={Wang, Junpeng and Zhang, Wei and Yang, Hao}, // booktitle={2020 IEEE Pacific Visualization Symposium (PacificVis)}, // pages={51--60}, // year={2020}, // organization={IEEE} // } // `, //SCANViz ` @inproceedings{kim2018interpretability, title={Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav)}, author={Kim, Been and Wattenberg, Martin and Gilmer, Justin and Cai, Carrie and Wexler, James and Viegas, Fernanda and others}, booktitle={International conference on machine learning}, pages={2668--2677}, year={2018}, organization={PMLR} }` //TCAV ]3d:[{"id":"f8b331c6dc749aec08ff4b89eed32ea794e73f2f25933a79acac24837d83af86c04654578de52851a4c5ac6fcdbc68c96b308311d8eebc4d7c5a5177a049e5ac","href":"https://static.observableusercontent.com/files/f8b331c6dc749aec08ff4b89eed32ea794e73f2f25933a79acac24837d83af86c04654578de52851a4c5ac6fcdbc68c96b308311d8eebc4d7c5a5177a049e5ac","name":"Hypotheses@1.svg","type":"image/svg+xml","lastModified":1625778076750.456,"size":62208},{"id":"b75d606a4ff29c39fe4d2767e5127808566d95ca1793752ceb5715527874c06040402805e95c9f5a9921dc5d6105922782193845c6bbf099098420f5b448d986","href":"https://static.observableusercontent.com/files/b75d606a4ff29c39fe4d2767e5127808566d95ca1793752ceb5715527874c06040402805e95c9f5a9921dc5d6105922782193845c6bbf099098420f5b448d986","name":"overview_final.svg","type":"image/svg+xml","lastModified":1627528340947.742,"size":198185},{"id":"3fb94c81c1b99228c938fb4b62487713e9d172566ea5554be30abb3eb210ed0754c604456585cf30c6da6c45325fbef8b6ac1ced5cb0079a93e54dcb15f51739","href":"https://static.observableusercontent.com/files/3fb94c81c1b99228c938fb4b62487713e9d172566ea5554be30abb3eb210ed0754c604456585cf30c6da6c45325fbef8b6ac1ced5cb0079a93e54dcb15f51739","name":"Interface Capture@1.svg","type":"image/svg+xml","lastModified":1627529065219.912,"size":2687250},{"id":"0ffbf44cd92c4ff0556a7c3c6d7483b3a4180d58645f3873b52a4b580b83ebc911e8869077a6cdd9c23b77866a35ce5123e8215095cbf54b290124b2f05654c4","href":"https://static.observableusercontent.com/files/0ffbf44cd92c4ff0556a7c3c6d7483b3a4180d58645f3873b52a4b580b83ebc911e8869077a6cdd9c23b77866a35ce5123e8215095cbf54b290124b2f05654c4","name":"scatterplot_generic@1.svg","type":"image/svg+xml","lastModified":1627530081570.287,"size":1169720},{"id":"daf43faea39ee381435e3d39b5d4c215f69ea0e801d25586c49f26fd4fc9ed061bc75bbd6439ef978e981352156495a21634d36c7a986a059be0b2e3c8734061","href":"https://static.observableusercontent.com/files/daf43faea39ee381435e3d39b5d4c215f69ea0e801d25586c49f26fd4fc9ed061bc75bbd6439ef978e981352156495a21634d36c7a986a059be0b2e3c8734061","name":"summarization_generic.svg","type":"image/svg+xml","lastModified":1627531888860.869,"size":316873},{"id":"1056dca376e337627dcfdcdcceee263471346d02ba491cb0718fb95bc113b9f96e1cb7e2f06ad041dde850c62f466ce53e162cc573820b14585292c4aeaf3041","href":"https://static.observableusercontent.com/files/1056dca376e337627dcfdcdcceee263471346d02ba491cb0718fb95bc113b9f96e1cb7e2f06ad041dde850c62f466ce53e162cc573820b14585292c4aeaf3041","name":"scat_summ_interaction.svg","type":"image/svg+xml","lastModified":1627536006594.187,"size":15121819},{"id":"fa569b5fc17ac59782ea6c35eb96b480749f7c799dbc663004452150185c6c36a1193fc3458bbd14614188614ea595884cc67cab5c9991d6f43b08b05dace4b3","href":"https://static.observableusercontent.com/files/fa569b5fc17ac59782ea6c35eb96b480749f7c799dbc663004452150185c6c36a1193fc3458bbd14614188614ea595884cc67cab5c9991d6f43b08b05dace4b3","name":"exp_1_assess_dir_figure.png","type":"image/png","lastModified":1628126077914.758,"size":896490},{"id":"6b8c47c4f4d7007504da189833da1e7d486b98ac12f814f7b0e5677cdd09dec8a38f3ef4142d2af79566bba681cb175091565bd627456af59430335f6952bc84","href":"https://static.observableusercontent.com/files/6b8c47c4f4d7007504da189833da1e7d486b98ac12f814f7b0e5677cdd09dec8a38f3ef4142d2af79566bba681cb175091565bd627456af59430335f6952bc84","name":"Interface Capture@2.svg","type":"image/svg+xml","lastModified":1627651806947.173,"size":2529784},{"id":"e574f1ebca069421cf55de978f363bb2aa17c85759754c9063a169a1e075d95f6067bf76e62b1cff097758ef0089e456ea14cc0f5a2537934f57200a024234a0","href":"https://static.observableusercontent.com/files/e574f1ebca069421cf55de978f363bb2aa17c85759754c9063a169a1e075d95f6067bf76e62b1cff097758ef0089e456ea14cc0f5a2537934f57200a024234a0","name":"Comparison_Figure_1@2.svg","type":"image/svg+xml","lastModified":1627512779026.803,"size":256369},{"id":"648667c5d3c1b7077c1a0a4ad878675d800045d739434dbed15ebdb3ab9a45c31d28ece0bb2546895e66ed252f70abce2a313304f12dedb2bb364a02675775b9","href":"https://static.observableusercontent.com/files/648667c5d3c1b7077c1a0a4ad878675d800045d739434dbed15ebdb3ab9a45c31d28ece0bb2546895e66ed252f70abce2a313304f12dedb2bb364a02675775b9","name":"vector_arithmetic_annotated.png","type":"image/png","lastModified":1628284104521.664,"size":998233},{"id":"85767e8ab8d898f03a18161578742f5a8410776423e4be08683b9028032f84e8c4dbaa71adba01e89dd43ebf711f053cd64ab55cb0745fc8de6097c40c842d45","href":"https://static.observableusercontent.com/files/85767e8ab8d898f03a18161578742f5a8410776423e4be08683b9028032f84e8c4dbaa71adba01e89dd43ebf711f053cd64ab55cb0745fc8de6097c40c842d45","name":"exp_2_figure@6.svg","type":"image/svg+xml","lastModified":1628284248421.651,"size":2347223},{"id":"607c28b4b2000180b3ca2084f801577202eeb15181dc8e80398f303eac80e60b75e1e324d89db3ae3443c2287913f1d74876e12a21b639c70b26501f5303eaa3","href":"https://static.observableusercontent.com/files/607c28b4b2000180b3ca2084f801577202eeb15181dc8e80398f303eac80e60b75e1e324d89db3ae3443c2287913f1d74876e12a21b639c70b26501f5303eaa3","name":"scat_summ_interaction@1.svg","type":"image/svg+xml","lastModified":1627672422053.757,"size":4330355},{"id":"c6be5b326b19bd791d7e1b25f47700d8e33f4642297a8fd934885465c55d92d57e2162f5b0eeb528510df87cd21d27d667a8bd53aebcb5063d7e021fbcb6bced","href":"https://static.observableusercontent.com/files/c6be5b326b19bd791d7e1b25f47700d8e33f4642297a8fd934885465c55d92d57e2162f5b0eeb528510df87cd21d27d667a8bd53aebcb5063d7e021fbcb6bced","name":"Visualization Introduction.png","type":"image/png","lastModified":1627332409667.654,"size":695276},{"id":"11d52df70d5c37b41e5d04ce0c4c629fdd687d363f2802e0a90e64db9764bc2e04d381c317cccbc8a96fdddd2234b06cdb2b6184320108e66a8f4a692cb14213","href":"https://static.observableusercontent.com/files/11d52df70d5c37b41e5d04ce0c4c629fdd687d363f2802e0a90e64db9764bc2e04d381c317cccbc8a96fdddd2234b06cdb2b6184320108e66a8f4a692cb14213","name":"exp_2_figure@3.svg","type":"image/svg+xml","lastModified":1628175500696.762,"size":4425987},{"id":"eeeebe6d76be5045259c27924d894ad716a59db4ff908197b73e13a38aabe75b550bfdea162f484103d096157499d5360a091187f1ba6dff487c2467e0827bc1","href":"https://static.observableusercontent.com/files/eeeebe6d76be5045259c27924d894ad716a59db4ff908197b73e13a38aabe75b550bfdea162f484103d096157499d5360a091187f1ba6dff487c2467e0827bc1","name":"Data Preprocess.svg","type":"image/svg+xml","lastModified":1624913602372.098,"size":63801},{"id":"3dc0fb9acb2630733697b511f265ac0a2ccb2700a272a4244d9c35c70270efaa3a8cf75ea6f0446e8771f4a1a8116d68a7dcffdecacd86cfa7965bfb538b4368","href":"https://static.observableusercontent.com/files/3dc0fb9acb2630733697b511f265ac0a2ccb2700a272a4244d9c35c70270efaa3a8cf75ea6f0446e8771f4a1a8116d68a7dcffdecacd86cfa7965bfb538b4368","name":"exp_4_figure@1.svg","type":"image/svg+xml","lastModified":1628183304340.97,"size":2708169},{"id":"084305b2ffc58642d592289fe26ab597f4f80fcb7570284f3e4992c726fb6dff1367fab15745644f5aae13a9fb1b8f6bf1d39638e558faf3446a0e9876edf621","href":"https://static.observableusercontent.com/files/084305b2ffc58642d592289fe26ab597f4f80fcb7570284f3e4992c726fb6dff1367fab15745644f5aae13a9fb1b8f6bf1d39638e558faf3446a0e9876edf621","name":"exp_1_informativeness@1.svg","type":"image/svg+xml","lastModified":1628185731961.337,"size":2295461},{"id":"8674743d33a0c96920d9985d6926aa1f7127c3a87cd1811f8f880d94e81577dd4b2576289bc4920a0689769df200c4be96cf30692eee54f09335e193a83340ed","href":"https://static.observableusercontent.com/files/8674743d33a0c96920d9985d6926aa1f7127c3a87cd1811f8f880d94e81577dd4b2576289bc4920a0689769df200c4be96cf30692eee54f09335e193a83340ed","name":"overview_final_2.svg","type":"image/svg+xml","lastModified":1627614526927.987,"size":198734},{"id":"cbeea6f6504e83a619e2a881a1da032bb7b67c337525df015853995c75923d4a1e9ab6daf7b4f927cb2a6dfcf028fc579e4e31d84d5a1a0cb9d713d2dd503050","href":"https://static.observableusercontent.com/files/cbeea6f6504e83a619e2a881a1da032bb7b67c337525df015853995c75923d4a1e9ab6daf7b4f927cb2a6dfcf028fc579e4e31d84d5a1a0cb9d713d2dd503050","name":"scatterplot_generic@2.svg","type":"image/svg+xml","lastModified":1627614606492.813,"size":1169719}] 2e:{"title":"Interactively Assessing Disentanglement in GANs","cells":[{"id":0,"value":"md`# Interactively Assessing Disentanglement in GANs\n\nSangwon Jeong, Vanderbilt University\n\nShusen Liu, Lawrence Livermore National Laboratory\n\nMatthew Berger, Vanderbilt University\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":5403,"value":"### Our paper was published in Eurovis 2022. Link to access the paper.","pinned":false,"hidden":false,"mode":"md"},{"id":5408,"value":"https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.14524","pinned":false,"hidden":false,"mode":"md"},{"id":888,"value":"html`

1. Introduction

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":772,"value":"$30","pinned":false,"hidden":false,"mode":"ojs"},{"id":3436,"value":"image_1 = await new Promise((resolve, reject) => {\n const img = html``\n img.onload = () => resolve(img)\n img.onerror = reject\n img.src = codecomparisonidenticalcode\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":1036,"value":"preset_fig_caption('1: Example Walks', \n 'Above walks are generated from three different latent codes, but along a \"youth\" direction. Note the different changes along this direction: the first walk does not change significantly in any other attributes but age, but the second walk does change in smiling. The third example eyeglasses disappears along with age.'\n )","pinned":false,"hidden":false,"mode":"ojs"},{"id":3429,"value":"$31","pinned":false,"hidden":false,"mode":"ojs"},{"id":647,"value":"$32","pinned":false,"hidden":false,"mode":"ojs"},{"id":3164,"value":"html`

3. Objectives & Tasks

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3227,"value":"$33","pinned":false,"hidden":false,"mode":"ojs"},{"id":2953,"value":"html`

4. GAN Latent Space Analysis

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":1969,"value":"html`\n

\n\nIn this section we discuss the generator model that we study in our work, as well as corresponding data we extract from the generator for visual analysis.\n\n

\n`\n","pinned":false,"hidden":false,"mode":"ojs"},{"id":1997,"value":"image = await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = pipelinefigure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":2017,"value":"preset_fig_caption(\n'2: Data Generation Process Overview',\n 'Demonstration of a data generation pipeline for a single walk with three codes. a sequence of latent codes are input to the generator. We extract convolutional activations as a generator representation of the code. Next, generated images are assessed by external attribute classifiers, which is then used to explain the walk in terms of classified attributes.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4254,"value":"$34","pinned":false,"hidden":false,"mode":"ojs"},{"id":2762,"value":"html`\n\n

5. Visualization Design

\n\n

\nOur visualization is designed for users to interactively explore a provided direction, and assess disentanglement, built around the tasks and data described above. Our design is comprised of four main views (c.f. Fig. 3): (1) a detailed view of linear regression models , (2) a depiction of attribute scores and summary of attribute trends, (3) a general similarity view on walks, and (4) detailed inspection of individual walks.\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3020,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = contentintroductionsrc\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":3900,"value":"preset_fig_caption(\n'3: Overview of our visual interface',\n ''\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4597,"value":"md`

5.1 Visual Encodings

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3045,"value":"html`\n\n

\nRegression Plot:\n\nWithin this scatterplot, the points correspond to all combinations of walks and their attribute-specific linear regression models, where we encode the model's slope on the x-axis, and its ${tex`R^2`} coefficient on the y-axis. This allows the user to see, at-a-glance, the quality of a selected direction (c.f. Fig. 4). For instance, if all points are concentrated around the centerline, then this suggests the direction is not reflective of any of the provided attributes. On the other hand, points that spread out near the upper left/right of the plot suggest a subset of walks that consistently decrease/increase along certain attributes.\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3908,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = scatter_generic_figure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":3926,"value":"preset_fig_caption(\n'4: Regression Plot',\n 'A scatterplot of regression models enables an at-a-glance quality assessment for a given direction.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4841,"value":"html`

\nAttribute View:\nIn this view we show, for each attribute, its set of scores over all walks as superimposed line plots - 1 line mark per code. Next to the line plots, we show a summary of the model slopes as a histogram, positive slopes being colored as red bars and negative slopes blue.\nAttributes sorted in descending order of ${tex`R^2`}, so the most relevant attributes for a given direction can be quickly determined. Please see Fig. 5 for an illustration.\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3090,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = summarization_generic_figure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":3973,"value":"preset_fig_caption(\n'5: Attribute Summary View',\n 'We allow the user to see for a given attribute, a summary of slopes over linear regression fits as a histogram, alongside detailed attribute scores for all walks.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":2949,"value":"html`\n\n

\nUmap Projection: \nWe show the UMap-projected generator representation as a scatterplot of walks. Its role is to act as a window for selecting walks in terms of similarity in their activations. This representation of the walks is decoupled from the information that regression provides, and enables cross-examination of walks in two different representation.\n\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":3136,"value":"html`\n\n

\nGenerated Images:\nLast, we allow for detailed inspection by showing a walk's corresponding sequence of synthesized images. This detailed inspection is essential to confirm potential findings made in the remaining views that depict quantitative information regarding the generator, and attribute-based, analyses. Showing the sequence of images can reveal whether a walk indeed captures expected changes with respect to an identified attribute.\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4058,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = scat_summ_interation_figure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":4063,"value":"preset_fig_caption(\n'6: Interaction between Scatter Plot and Summarization View',\n \"The interaction between the two views are bi-directional. Hovering over attribute labels highlight the corresponding attribute regressions in the scatter plot. Brushing over a region in the scatter plot highlights the corresponding walks' attribute scores and number of instances present across histogram bins.\"\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4601,"value":"md`

5.2 Interactions

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4619,"value":"$35","pinned":false,"hidden":false,"mode":"ojs"},{"id":236,"value":"html`

6. Experimental Results

\n\n\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4322,"value":"$36","pinned":false,"hidden":false,"mode":"ojs"},{"id":4309,"value":"html`\n

6.1 Direction Assessment

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4565,"value":"$37","pinned":false,"hidden":false,"mode":"ojs"},{"id":4447,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = step_1_distinctive_figure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":4450,"value":"preset_fig_caption(\n'7: Assessing Directions: Distinctiveness & Consistency',\n '(a) is a smiling direction and the (b) is no beard direction. An area that contains regressions with good fit is brushed for each direction.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4312,"value":"html`\n\n

\nFig. 8 further verifies the \"Smiling\" direction for informativeness, or whether the set of attributes related to the direction reflect the dominant factor of variation. Histograms highlighted for \"High Cheekbones\", \"Bags Under Eyes\", and \"Narrow Eyes\" indicate a correlation amongst these attributes, and these attributes indeed tend to change as one smiles. Further, inspecting the general walk scatterplot (UMap projection) indicates that most walks can be described in terms of these attributes.\n

\n\n

\nSelecting an example highlights the particular Smiling variation; however, we note a slight change in age for this walk. Indeed, the \"Young\" attribute shows a small correlation with the brushed set of examples, indicating that this direction can mix in unrelated concepts for certain walks.\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4670,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = step_1_informativeness\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":4674,"value":"preset_fig_caption(\n'8: Assessing Directions: Informativeness',\n 'We find that a \"Smiling\" direction results in correlation of attributes that can be easily understood. However, the direction may not be informative for a small set of codes, e.g. shown here as the \"Young\" attribute can also consistently decrease for certain portions of the latent space.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":5222,"value":"$38","pinned":false,"hidden":false,"mode":"ojs"},{"id":5219,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = exp_4_figure\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":5228,"value":"preset_fig_caption(\n'9: Vector arithmetic in a scene-based GAN',\n 'We show the types of variations produced when using a direction corresponding to \"flatness\", built from vector arithmetic. We see a tendency to produce lakes/rivers, some iced river, as further indicated by the \"ice\" and \"moist\" attributes.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":5116,"value":"$39","pinned":false,"hidden":false,"mode":"ojs"},{"id":4701,"value":"await new Promise((resolve, reject) => {\n const img = html``;\n img.onload = () => resolve(img);\n img.onerror = reject;\n img.src = exp_2_same_dir_diff_method\n})","pinned":false,"hidden":false,"mode":"ojs"},{"id":4704,"value":"preset_fig_caption(\n'10: Comparison of a smiling direction between walking methods',\n '(a) shows the basic procedure for assessing a direction. (b)-1 and (c)-1 is attribute hover of the cheekbone attribute for each unsupervised direction and supervised direction. (b)-2 and (c)-2 is attribute hove of the smiling attribute for unsupervised and supervised direction respectively. (d) are walks where bags under eyes attribute diminishes with an increased smiling.'\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4719,"value":"html`\n

\n\nThe supervised direction is not without its limitations, however. In Fig. 10(d) we can observe a small amount of walks which contain a decrease with respect to \"Bags Under Eyes\". Upon closer inspection, we find these walks introduce glasses as the presence of smiling increases. These types of walks are not present in the unsupervised direction.\n\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4636,"value":"html`\n

6.3 Bias in the Latent Space

\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4743,"value":"$3a","pinned":false,"hidden":false,"mode":"ojs"},{"id":4746,"value":"html`\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4754,"value":"preset_fig_caption(\n'11: Supervised walking - young direction',\n ''\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4750,"value":"html`\n\n

\n\nIn Fig. 12 we show the Gender direction, along with a default highlighted attribute for \"Smiling\". As in the above example, we find an age bias here, and a tendency for younger females along this direction. Moreover, for the default highlighted attribute \"Smiling\", we find a tendency for men to smile less than women, further suggesting a gender bias in the dataset along the smiling attribute.\n\n

\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4779,"value":"html`\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4790,"value":"preset_fig_caption(\n'12: Supervised walking - gender direction',\n ''\n)","pinned":false,"hidden":false,"mode":"ojs"},{"id":4616,"value":"html`\n\n

7. Interface

\n\n

\nFor more exploratory analysis we have provided an interface to permit an investigation of directions found through different GANs, as well as different direction-finding methods.\n

\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":1654,"value":"html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":1704,"value":"html`

8. Discussion

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":4017,"value":"$3b","pinned":false,"hidden":false,"mode":"ojs"},{"id":524,"value":"html`

Bibliography

`","pinned":false,"hidden":false,"mode":"ojs"},{"id":695,"value":"viewof citations = genBibliography(bibtex, {reftype: 'indexed'})","pinned":false,"hidden":false,"mode":"ojs"},{"id":699,"value":"$3c","pinned":false,"hidden":false,"mode":"ojs"},{"id":89,"value":"md`# CSS`","pinned":false,"hidden":false,"mode":"ojs"},{"id":116,"value":"html`\n\n`","pinned":false,"hidden":false,"mode":"ojs"},{"id":118,"value":"html``","pinned":false,"hidden":false,"mode":"ojs"},{"id":688,"value":"md`# Imports`","pinned":false,"hidden":false,"mode":"ojs"},{"id":690,"value":"Cite = require('https://bundle.run/citation-js@0.5.0-alpha.10')","pinned":false,"hidden":false,"mode":"ojs"},{"id":693,"value":"function genBibliography(bibliography, opt = {}) {\n let { reftype, style, output } = { output: 'html', reftype: 'indexed', style: 'citation-apa', ...opt}\n const cite = new Cite()\n for (const b of bibliography) {\n cite.add(b);\n }\n const authorObj = cite.get().sort((a,b) => a.author[0].family > b.author[0].family ? 1 : -1);\n const authormap = [];\n for (const [i, o] of Object.entries(authorObj)) {\n authormap.push([o.id, reftype == \"indexed\" ? `\\[${+i+1}\\]` : `(${o.author[0].family})`])\n }\n const bibcontainer = cite.get({type: output, style: style})\n if (reftype == \"indexed\") {\n bibcontainer.querySelectorAll('.csl-entry').forEach((e,i) => {\n e.innerHTML = `[${i + 1}] ${e.innerHTML}`;\n })\n }\n const result = html`${bibcontainer}`;\n result.value = new Map(authormap);\n return result;\n}","pinned":false,"hidden":false,"mode":"ojs"},{"id":2011,"value":"datapreprocesssvgsrc = FileAttachment(\"Data Preprocess.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":2362,"value":"hypothesessvgsrc = FileAttachment(\"Hypotheses@1.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":3023,"value":"contentintroductionsrc = FileAttachment(\"Interface Capture@2.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":3437,"value":"codecomparisonidenticalcode = FileAttachment(\"Comparison_Figure_1@2.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":3679,"value":"pipelinefigure = FileAttachment('overview_final_2.svg').url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":3905,"value":"scatter_generic_figure = FileAttachment(\"scatterplot_generic@2.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":3969,"value":"summarization_generic_figure = FileAttachment(\"summarization_generic.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":4052,"value":"scat_summ_interation_figure = FileAttachment(\"scat_summ_interaction@1.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":4444,"value":"step_1_distinctive_figure = FileAttachment('exp_1_assess_dir_figure.png').url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":4667,"value":"step_1_informativeness = FileAttachment('exp_1_informativeness@1.svg').url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":4699,"value":"exp_2_same_dir_diff_method = FileAttachment(\"exp_2_figure@6.svg\").url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":4938,"value":"exp_4_figure = FileAttachment('vector_arithmetic_annotated.png').url()","pinned":false,"hidden":false,"mode":"ojs"},{"id":2019,"value":"function preset_fig_caption(fig_num, description) {\n return html`\n Figure ${fig_num} \n \n \n

\n ${description}\n

\n`\n}","pinned":false,"hidden":false,"mode":"ojs"}],"files":"$W3d","stdlib":"1","titling":"auto"}