Null-Sampling for Interpretable and Fair Representations

Kehrenberg, Thomas; Bartlett, Myles; Thomas, Oliver; Quadrianto, Novi

doi:10.1007/978-3-030-58574-7_34

Cited by 13 publications

(8 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Interpreting the representations We now show that the bijectivity of FNF enables interpretability analyses, an active area of research in fair representation learning [71][72][73]. To that end, we consider the Crime dataset where for a community x: (i) race (non-white vs. white majority) is the sensitive attribute a, (ii) the percentage of whites (highly correlated with, but not entirely predictive of the sensitive attribute) is in the feature set, and (iii) the label y strongly correlates with the sensitive attribute.…”

Section: Comparison With Adversarial Trainingmentioning

confidence: 99%

Fair Normalizing Flows

Balunović¹,

Ruoss²,

Vechev³

2021

Preprint

View full text Add to dashboard Cite

Fair representation learning is an attractive approach that promises fairness of downstream predictors by encoding sensitive data. Unfortunately, recent work has shown that strong adversarial predictors can still exhibit unfairness by recovering sensitive attributes from these representations. In this work, we present Fair Normalizing Flows (FNF), a new approach offering more rigorous fairness guarantees for learned representations. Specifically, we consider a practical setting where we can estimate the probability density for sensitive groups. The key idea is to model the encoder as a normalizing flow trained to minimize the statistical distance between the latent representations of different groups. The main advantage of FNF is that its exact likelihood computation allows us to obtain guarantees on the maximum unfairness of any potentially adversarial downstream predictor. We experimentally demonstrate the effectiveness of FNF in enforcing various group fairness notions, as well as other attractive properties such as interpretability and transfer learning, on a variety of challenging real-world datasets.Preprint. Under review.

show abstract

Section: Comparison With Adversarial Trainingmentioning

confidence: 99%

Fair Normalizing Flows

Balunović¹,

Ruoss²,

Vechev³

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Fair representation learning A wide range of methods have been proposed to learn fair representations of user data. Most of these works consider group fairness and employ techniques such as adversarial learning [17,38,49,53], disentanglement [12,51,66], duality [70], low-rank matrix factorization [60], and distribution alignment [3,52,85]. Individually fair representation learning has recently gained attention, with similarity metrics based on logical formulas [65], Wasserstein distance [22,46], fairness graphs [47], and weighted ℓ p -norms [84].…”

Section: Related Workmentioning

confidence: 99%

Latent Space Smoothing for Individually Fair Representations

Peychev¹,

Ruoss²,

Balunović³

et al. 2021

Preprint

View full text Add to dashboard Cite

Fair representation learning encodes user data to ensure fairness and utility, regardless of the downstream application. However, learning individually fair representations, i.e., guaranteeing that similar individuals are treated similarly, remains challenging in high-dimensional settings such as computer vision. In this work, we introduce LASSI, the first representation learning method for certifying individual fairness of high-dimensional data. Our key insight is to leverage recent advances in generative modeling to capture the set of similar individuals in the generative latent space. This allows learning individually fair representations where similar individuals are mapped close together, by using adversarial training to minimize the distance between their representations. Finally, we employ randomized smoothing to provably map similar individuals close together, in turn ensuring that local robustness verification of the downstream application results in end-to-end fairness certification. Our experimental evaluation on challenging real-world image data demonstrates that our method increases certified individual fairness by up to 60%, without significantly affecting task utility.

show abstract

“…Omniglot [38] OM [41] Handwritten chars Few-shot MiniImagenet [70] MI [44] ImageNet subset Few-shot Labeled Faces in the Wild [27] LFW [8] Faces Diversity -Faces UCF101 [64] UCF [28] Action videos Diversity -Actions Imagenet-R [23] IR [52] ImageNet art Diversity -Art Imagenet-Sketch [71] IS [71] Imagenet sketches Diversity -Sketch Indoor Scene Recognition [51] ISR [56] Indoor location Diversity -Indoor CIFAR10 [37] C10 [14] 10 classes Diversity Imagenet-A [24] IA [52] Difficult images Robustness Colorectal Histology [30] CH [16] Medical Images AI for good Multi-label CelebA Attributes [43] CAA [31] 40 attribute Label overlap UTK Faces [75] UTK [29] Gender, age, race Label overlap Yale Faces [15] YF [32] 11 face labels Label overlap Common Objects in Context [40] COCO [59] 90 objects Object Detection iMaterialist (Fashion) [18] IM [18] Fashion/apparel Fine-grained We randomly sample classes and data points from each dataset 100 times and average the results.…”

Section: Sota Content Selection Reason Singlementioning

confidence: 99%

Personalizing Pre-trained Models

Khan¹,

Srivatsa²,

Rane³

et al. 2021

Preprint

View full text Add to dashboard Cite

Self-supervised or weakly supervised models trained on large-scale datasets have shown sample-efficient transfer to diverse datasets in few-shot settings. We consider how upstream pretrained models can be leveraged for downstream few-shot, multilabel, and continual learning tasks. Our model CLIPPER (CLIP PERsonalized) uses image representations from CLIP, a large-scale image representation learning model trained using weak natural language supervision. We developed a technique, called Multi-label Weight Imprinting (MWI), for multi-label, continual, and few-shot learning, and CLIPPER uses MWI with image representations from CLIP. We evaluated CLIPPER on 10 single-label and 5 multi-label datasets. Our model shows robust and competitive performance, and we set new benchmarks for few-shot, multi-label, and continual learning. Our lightweight technique is also compute-efficient and enables privacy-preserving applications as the data is not sent to the upstream model for fine-tuning. Thus, we enable few-shot, multilabel, and continual learning in compute-efficient and privacy-preserving settings.

show abstract

Null-Sampling for Interpretable and Fair Representations

Cited by 13 publications

References 8 publications

Fair Normalizing Flows

Fair Normalizing Flows

Latent Space Smoothing for Individually Fair Representations

Personalizing Pre-trained Models

Contact Info

Product

Resources

About