Pith. sign in

REVIEW 4 major objections 6 minor 227 references

The thesis claims that counterfactual reasoning—changing one meaningful attribute while holding all else fixed—is a single lens for explaining, auditing, and mitigating bias in both image classifiers and text-to-image models, and it builds

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Counterfactual-based methods for concept attribution in classifiers and for dynamic bias evaluation and mitigation in text-to-image models.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A competent compilation thesis whose dynamic bias-axis and intersectional tools are genuinely useful, but the VQA-backed quantitative backbone and the causal framing need more caution before the unified claims are fully secured. the 4 major comments →

arxiv 2508.20881 v1 pith:H3XXXRHX submitted 2025-08-28 cs.CV

Understanding and evaluating computer vision models through the lens of counterfactuals

Pushkar Shukla This is my paper
classification cs.CV
keywords counterfactual reasoninginterpretabilitybias evaluationbias mitigationtext-to-image generationintersectionalityfairnesscomputer vision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Counterfactual reasoning—asking what a model would output if one meaningful attribute changed while everything else stayed fixed—is the thesis's organizing claim: it is not just an explanation technique but a shared mechanism for interpretability, fairness evaluation, causal diagnosis, and bias mitigation in computer vision. The thesis develops five instantiations: CAVLI quantifies how much a classifier's decision depends on a human-defined concept by masking superpixels; ASAC generates adversarial perturbations along protected attributes and fine-tunes classifiers to improve fairness without losing accuracy; TIBET dynamically surfaces prompt-relevant bias axes in text-to-image models and scores them with CAS and MAD; BiasConnect turns those scores into pairwise causal graphs of intersectional bias; InterMit uses those graphs to mitigate multiple biases under user-defined priorities. If these claims hold, bias work in vision—across classifiers and generators—can share one disciplined experimental language: intervene on an attribute, hold the rest fixed, and read the output shift. The thesis reports early evidence for that unification: on CelebA smile classification, equalized-odds difference drops from 0.088 to 0.050 while accuracy rises from 84.29 to 91.91; bias metrics track human rankings in user studies; and the mitigation algorithm removes intersectional bias in fewer steps than a fixed-prompt baseline.

Core claim

The paper's core discovery is that the simple counterfactual operation—take an image or prompt, change one semantically meaningful attribute or identity term, keep everything else fixed—can become a single experimental protocol covering the full life cycle of vision model bias. In discriminative models the protocol works at two scales: CAVLI masks random subsets of superpixels and measures overlap between concept-relevant and decision-relevant regions, producing a Concept Dependency Score that exposes spurious reliance (for example, cows classified by grasslands); ASAC instead perturbs pixels against a protected-attribute classifier and uses a difficulty-ranked curriculum of those adversaria

What carries the argument

The central machinery is the concept-level counterfactual: an alternate input that differs from the original in a small set of semantically meaningful, human-defined concepts while all other attributes are held fixed. Every contribution is an instantiation of this operation. CAVLI implements it by turning superpixels on and off and correlating the resulting concept sensitivities with decision sensitivities, producing the Concept Dependency Score. ASAC implements it adversarially: perturbations computed against a protected-attribute classifier create inputs whose protected attribute flips while visual content is preserved, and a curriculum orders them by how strongly they deceive the target c

Load-bearing premise

The load-bearing premise is that the text answers a vision-language model gives about generated images accurately describe what is actually visible; the thesis reports an 18% error rate in those answers, so if the vision-language reader misreads images, every downstream bias score and causal link built on those answers is distorted.

What would settle it

Take a fixed set of occupation prompts, generate the image sets once, then compute CAS/MAD with two different vision-language readers and with a human-annotated concept set for a subset. If the ordering of bias axes by MAD, or the sign of a BiasConnect cross-axis effect, changes across readers, the counterfactual pipeline is not the source of the measurement—the reader is.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single evaluation protocol could replace disjoint fairness checklists: define bias axes dynamically per prompt, generate counterfactuals, measure concept shift, and aggregate scores across prompts.
  • Vision classifiers can be debiased without external generative counterfactual images; ASAC's adversarial perturbations along protected attributes improve fairness gaps and often raise accuracy, avoiding stereotype artifacts that GAN-based counterfactuals introduce.
  • Intersectional bias can be diagnosed as pairwise causal relationships rather than just subgroup accuracy gaps, so practitioners can anticipate side effects—such as ethnic diversification reducing gender diversity—before intervening.
  • Bias mitigation can become user-controllable and training-free: InterMit accepts user-defined priorities and target distributions and reports fewer mitigation steps than the ITI-GEN baseline with higher mitigation amount and visual quality.
  • The same methodology transfers to any black-box text-to-image model, since TIBET and its successors only require prompts and generated images, not model internals.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • I read the thesis as implying that the quantitative conclusions would need to survive a change of concept reader: if a second, independently calibrated VQA model gives different CAS/MAD rankings on the same image sets, then the measurement is partly a property of the reader rather than the generator.
  • The pairwise causal graphs suggest a natural extension the author does not build: formulate mitigation as constrained optimization over the intersectional sensitivity matrix with user priorities, rather than the multi-step greedy heuristic InterMit uses. This could be tested by comparing InterMit's step selection against a differentiable solver on the same bias axes.
  • The dynamic-axis idea could transfer to video or 3D generation, where 'holding everything fixed' is harder; a risk is that the LLM-proposed axes quietly become a new fixed checklist unless re-validated per prompt.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The thesis argues that counterfactual reasoning provides a unifying methodology for interpretability, bias auditing, and bias mitigation in both vision classifiers and text-to-image (TTI) models. It develops and evaluates four main contributions: CAVLI (a hybrid TCAV/LIME concept-dependency score), ASAC (adversarial attribute-specific counterfactuals with curriculum fine-tuning), TIBET (dynamic bias-axis identification and CAS/MAD bias metrics), and BiasConnect/InterMit (pairwise dependency graphs and training-free mitigation with user-defined priorities). The evaluation spans ImageNet, CelebA, UTK, multiple backbones and TTI models, and includes user studies. The central claim is that counterfactuals are not just diagnostic tools but a principled methodology for fairness, causality, and explainability in vision systems.

Significance. If the technical concerns below are resolved, this would be a substantial contribution. The idea of prompt-dependent dynamic bias axes is a clear improvement over fixed-axis audits, and the thesis explicitly addresses intersectionality rather than treating bias dimensions independently. Strengths include the use of user studies to calibrate LLM/VQA-based measurements, multi-dataset and multi-backbone experiments, comparisons against prior methods (ITI-GEN, Ramaswamy et al.), and unusually honest limitation sections. The thesis also makes falsifiable predictions (e.g., CAS/MAD differences across prompts and models) and can be directly applied to existing black-box TTI systems. The main scientific risk is that the quantitative conclusions—especially those in Chapters 5–7—inherit an unquantified systematic error from the VQA concept extraction, and that the word 'causal' is used without identification assumptions.

major comments (4)
  1. [Section 5.5.3 / Section 5.3.3.1] The VQA concept extraction is the single point of failure for CAS, MAD, BiasConnect, and InterMit. The reported 18% VQA error rate and 13.11% MAD error are treated as if errors were independent and random. If MiniGPT-v2 has systematic biases (for example, the same gender/ethnicity associations as Stable Diffusion), then a low CAS can reflect the VQA model's stereotypes rather than the TTI model's behavior. The user studies validate perceived bias, not the correctness of the extracted concept histograms. The thesis needs a conditional error analysis: on what type of images, and for which bias axes, does the VQA model err? Without ground-truth concept annotations or a negative-control test, the main quantitative claims of Chapters 5–7 are not secure.
  2. [Section 6.5 / Section 6.6] The thesis calls BiasGraph 'causal graphs' and describes the approach as 'pairwise causal discovery', but no SCM is specified and no identification assumptions are stated (consistency, positivity, no unobserved confounding, no interference). The text in Section 2.4.2 itself lists hidden confounding as a central challenge, yet Chapter 6 does not apply any sensitivity analysis. The prompt-level intervention (e.g., changing 'gender' words) may affect multiple latent image-generation factors, and the resulting CAS associations need not identify a causal effect. Either state the assumptions under which the do-operator is justified for these text-to-image interventions, or rename the output as 'conditional association graphs' and add synthetic negative-control experiments where the true causal direction is known.
  3. [Section 3.3 / Algorithm 1] The validation of CAVLI is partly circular. CDS is defined in Algorithm 1 (step 9) as CDSi = gamma_i * CS^k_{C,l}(F, x_i), where CS is the TCAV conceptual sensitivity from Eq. (3.1). The chapter then claims that mean CDS 'resembles' TCAV scores and treats this as evidence that CDS captures concept dependence. Since TCAV sensitivity is a multiplicative factor of CDS, a positive correlation is expected by construction. An independent validation is needed: synthetic images with known concept dependencies, human agreement on the concept heatmaps, or an intervention-based test that removes the concept and measures the decision change.
  4. [Section 4.6.3 / Table 4.7 / Table 4.6] Hyperparameters appear to be selected after inspecting the same test-set metrics. Table 4.7 reports choosing noise magnitudes epsilon = {0.001, 0.01} from the ablation on the smile classifier, and Table 4.6 chooses curriculum direction (ascending vs descending difficulty) based on fairness/accuracy outcomes on the same evaluation. If these choices are not made on a separate validation set, the reported gains (e.g., DEO 0.050 vs baseline 0.088) may be partially due to selection bias. The thesis should either perform nested cross-validation or explicitly state which hyperparameters were chosen a priori and which were post-hoc.
minor comments (6)
  1. [Throughout] There are frequent typos and formatting artifacts: 'Adavantages' (Section 2.4), 'consitency' (Section 2.5.2.1), 'TCA V' / 'CA VLI' spacing, 'TIBET' vs 'TIBET' in figure captions. A careful copyedit is needed.
  2. [Eq. (5.1)–(5.3)] The 'Histogram Intersection-over-Union' is not the standard IoU. Please define it explicitly as sum(min(w_i, w_cf)) / sum(max(w_i, w_cf)), and clarify how zero-frequency concepts are treated in the vocabulary.
  3. [Section 5.3.1] The text says GPT-3, but the experiments use gpt-3.5-turbo. The model version and API date matter for reproducibility; please be precise.
  4. [Figure 5.4 / Appendix A.2] Figure 5.4 mentions 'MAD = 0.49' with no explanation of the normalization. Since the normalization is used in all later comparisons, it should be described in the main text, not only in Appendix A.2.
  5. [Chapter 6 / Chapter 7] The terms 'VLLMs' and 'VLMs', 'TTI' and 'T2I', and 'BiasConnect' vs 'BiasGraph' are used inconsistently. Please harmonize notation.
  6. [Reproducibility] No code release, data release, or detailed hyperparameter table is provided. Given the pipeline includes many LLM/VQA components, a reproducibility statement would strengthen the manuscript considerably.

Circularity Check

1 steps flagged

One partial self-referential validation in CAVLI; TIBET/BiasConnect/InterMit are otherwise externally validated.

specific steps
  1. self definitional [Section 3.3.1 (ImageNet evaluation) and Algorithm 1 (Section 3.2.4)]
    "Calculate Concept Dependency Score: CDSi = γi · CSk C,l(F, xi). ... We propose a hypothesis that if there exists a correlation between the mean CDS scores and global concept methods like TCA V , it indicates that our metric is capable of accurately capturing the dependence between the model decisions and underlying concepts."

    CDS is defined as the product of the spatial-overlap correlation γ and the TCAV conceptual sensitivity CS. The validation treats TCAV as an independent reference and takes correlation of mean CDS with TCAV as evidence that CDS captures concept dependence. But because CS is a factor in CDS, the two quantities are coupled by construction; even a CDS that added no concept information beyond TCAV's own sensitivity would correlate with TCAV. The correlation therefore cannot establish the validity of the new overlap term γ. The method still has independent content in γ, so this is a partial, not total, circularity.

full rationale

Aside from the CAVLI validation issue, the thesis is largely self-contained. CAVLI's CDS combines LIME-based perturbation with TCAV; the spatial overlap component is novel. TIBET's CAS/MAD metrics are measured from VQA concept frequencies, with external user studies and ITI-GEN ground-truth comparisons; MAD sensitivity to VQA errors is reported as a limitation, not hidden. BiasConnect is validated against post-mitigation ITI-GEN measurements, an external intervention. InterMit optimizes user-defined target distributions; evaluating against the same objectives is optimization, not circularity. Self-citations to prior papers by the author are normal for a thesis and are not load-bearing: the later chapters use the earlier methods as tools and validate them externally. Overall, one partial self-referential validation prevents a 0-2 score; there is no evidence that the central fairness/causality claims reduce by construction to their inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new natural-kind entities are introduced. The thesis proposes measurement constructs (CDS, CAS, MAD, Intersectional Sensitivity) and algorithms, but these are not physical entities.

free parameters (6)
  • noise magnitudes epsilon (FGSM/PGD) = {0.001, 0.01} for ASACs; {0.03, 0.05} in ablation
    Chosen based on observed fairness/accuracy in Section 4.6.3.2; not justified by a principled criterion.
  • curriculum learning weight alpha = 0.5
    Set as default in Section 4.5.2; no sensitivity analysis.
  • number of superpixels and random subsets in CAVLI = r and n unspecified
    Algorithm 1 leaves r and n unspecified; results may depend on these.
  • number of images per prompt in TIBET = 48
    Section 5.3.2; no analysis of effect of sample size.
  • VQA question templates = hand-designed for common axes
    Table 5.2; the choice of questions affects extracted concepts.
  • number of counterfactuals per axis = K (variable)
    MAD is normalized by K; variation in K could affect comparability.
axioms (5)
  • domain assumption Uniform CAS distribution across counterfactuals implies low bias; skew implies bias
    Section 5.3.4 defines MAD this way; this is a normative assumption about what bias looks like.
  • domain assumption MiniGPT-v2 VQA provides sufficiently accurate concept extraction
    Section 5.3.3 relies on this; the author's own sensitivity analysis (Section 5.5.3) shows 18% error.
  • domain assumption GPT-3.5 can generate relevant bias axes and counterfactuals
    Section 5.3.1; precision/recall in User Study 1 is not perfect.
  • domain assumption Adversarial perturbations of the protected-attribute classifier preserve semantics and only change the protected attribute
    Section 4.4.1; this is the foundation of ASACs but is not empirically verified beyond qualitative examples.
  • domain assumption Pairwise causal effects measured via prompt interventions are unconfounded and can be combined into a causal graph
    Sections 6.4-6.6; no causal identification assumptions (e.g., no hidden confounders) are stated.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding and evaluating computer vision models through the lens of counterfactuals." pith.science (2026). https://pith.science/paper/H3XXXRHX

@misc{pith2026250820881,
  author       = {Pith},
  title        = {Pith review of: Understanding and evaluating computer vision models through the lens of counterfactuals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H3XXXRHX}},
  note         = {Machine review of arXiv:2508.20881}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Counterfactual reasoning -- the practice of asking ``what if'' by varying inputs and observing changes in model behavior -- has become central to interpretable and fair AI. This thesis develops frameworks that use counterfactuals to explain, audit, and mitigate bias in vision classifiers and generative models. By systematically altering semantically meaningful attributes while holding others fixed, these methods uncover spurious correlations, probe causal dependencies, and help build more robust systems. The first part addresses vision classifiers. CAVLI integrates attribution (LIME) with concept-level analysis (TCAV) to quantify how strongly decisions rely on human-interpretable concepts. With localized heatmaps and a Concept Dependency Score, CAVLI shows when models depend on irrelevant cues like backgrounds. Extending this, ASAC introduces adversarial counterfactuals that perturb protected attributes while preserving semantics. Through curriculum learning, ASAC fine-tunes biased models for improved fairness and accuracy while avoiding stereotype-laden artifacts. The second part targets generative Text-to-Image (TTI) models. TIBET provides a scalable pipeline for evaluating prompt-sensitive biases by varying identity-related terms, enabling causal auditing of how race, gender, and age affect image generation. To capture interactions, BiasConnect builds causal graphs diagnosing intersectional biases. Finally, InterMit offers a modular, training-free algorithm that mitigates intersectional bias via causal sensitivity scores and user-defined fairness goals. Together, these contributions show counterfactuals as a unifying lens for interpretability, fairness, and causality in both discriminative and generative models, establishing principled, scalable methods for socially responsible bias evaluation and mitigation.

Figures

Figures reproduced from arXiv: 2508.20881 by Pushkar Shukla.

Figure 1.1
Figure 1.1. Figure 1.1: Counterfactual reasoning reveals hiring bias. In a famous study conducted by Moss [PITH_FULL_IMAGE:figures/full_fig_p021_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. Figure 1.2: This figure illustrates how counterfactual reasoning is employed throughout the thesis to probe and enhance the capabilities of computer vision models. In Chapter 3, we introduce CAVLI, a method for quantifying the influence of visual concepts on model decisions via counterfactuals. Chapter 4 leverages counterfactuals for bias mitigation using ASACs, showing their efficacy in reducing unfair model behavi… view at source ↗
Figure 2.1
Figure 2.1. Figure 2.1: A brief timeline tracing the development of counterfactual reasoning. While the term [PITH_FULL_IMAGE:figures/full_fig_p031_2_1.png] view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: This chapter begins with a central question: When a black-box model predicts a label [PITH_FULL_IMAGE:figures/full_fig_p054_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: Overview of our proposed approach, CAVLI, to estimate the dependence of a concept [PITH_FULL_IMAGE:figures/full_fig_p057_3_2.png] view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: We use concept and decision heatmaps to analyze a classifier’s decisions and their [PITH_FULL_IMAGE:figures/full_fig_p065_3_3.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: What happens when we ask a counterfactual generation engine to modify an image [PITH_FULL_IMAGE:figures/full_fig_p068_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: An example of gender-based counterfactuals generated by StyleGAN2 and our method [PITH_FULL_IMAGE:figures/full_fig_p070_4_2.png] view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: Bias Mitigation Strategy: Our proposed solution for mitigating biases in a model (e.g., smile classifier) M(θ,ρ) involves training sensitive attribute classifier C(θ,ϕ) (shown in the network architecture). We then follow a three-stage pipeline. (1) We generate ASACs that are capable of deceiving C(θ,ϕ) . (2) We define a curriculum assignment strategy that organizes these ASACs based on the degree to whic… view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: Qualitative results showing that our trained model becomes robust to ASACs after [PITH_FULL_IMAGE:figures/full_fig_p083_4_4.png] view at source ↗
Figure 5.1
Figure 5.1. Figure 5.1: In generative modeling, the dimensions along which bias manifests are highly dependent [PITH_FULL_IMAGE:figures/full_fig_p090_5_1.png] view at source ↗
Figure 5.2
Figure 5.2. Figure 5.2: TIBET. Given an input prompt, we query an LLM (GPT-3) to identify axes of biases (Step 1), and generate counterfactual prompts for each axis of bias (Step 2). Here, we show a sample of three counterfactual prompts for the physical appearance bias, and two for the ableism bias. Next, we use a black-box TTI model (Stable Diffusion [3]) to generate images for the initial prompt as well as each counterfactua… view at source ↗
Figure 5.3
Figure 5.3. Figure 5.3: VQA-based Image Comparison. The VQA process in the Image Comparison Module. For each image in the initial set and the counterfactual set, we ask a list of questions to the VQA model, MiniGPT-v2, to extract concepts. The underlined words indicate the axis of bias that is associated with that question. factual prompt in terms of relevant concepts. CASVQA uses a concept-level matching algorithm to compare c… view at source ↗
Figure 5.4
Figure 5.4. Figure 5.4: Metrics: (a) MAD is low when the CAS scores are uniform across all counterfactuals, and high when the CAS scores are skewed. (b) MAD is only dependent on variability in CAS, not on amount of CAS occurring concepts in a given image for a specific bias axis. To calculate this measure we extract concepts from questions specific to a bias axis and sort them in order of frequency over the image set. 5.4 Datas… view at source ↗
Figure 5.5
Figure 5.5. Figure 5.5: Usefulness of TIBET. In this example setting, we show how TIBET can be useful to a user concerned about biases in the images generated by a TTI model. We show how TIBET can analyse biases along human-observable axes of bias, with post-hoc explainablity. 84 [PITH_FULL_IMAGE:figures/full_fig_p104_5_5.png] view at source ↗
Figure 5.6
Figure 5.6. Figure 5.6: Analysis enabled by TIBET. Our approach calculates CAS and MAD scores to measure association with counterfactual prompts and bias degree in generated images. Qualitative metrics like Top-K Concepts and Axis-Aligned Top-K Concepts offer post-hoc model explanations. Additionally, our approach enables comparisons with counterfactual explanations. 5.5 Experiments The section can be broadly divided into three… view at source ↗
Figure 5.7
Figure 5.7. Figure 5.7: Sensitivity Analysis on CAS and MAD for errors in VQA. Per User Study 3 , we estimate an 18% error rate in VQA, leading to 4.73% and 13.11% error in CAS and MAD respectively. Diffusion 2.1 seems to reduce bias among a few of these professions, notably “accountant” and “pharmacist.” The overall trends observed in our bias metrics align with those found in previous studies like T2IAT and DALL-Eval. There a… view at source ↗
Figure 5.8
Figure 5.8. Figure 5.8: Comparing our VQA and CLIP methods for Image Comparison. In this example, we see that humans rank racial bias to be more significant compared to gender bias, which is also observable in the images. We compare our VQA-based method to our CLIP-based method, and observe that the VQA-based method better aligns with human ranking. This is because, in most cases, biases are attributed to specific characteristi… view at source ↗
Figure 5.9
Figure 5.9. Figure 5.9: Bias identification and mitigation. We compute difference in CAS scores for male and female counterfactuals for 11 occupation prompts. (a) and (b) show male and female leaning professions using Stable Diffusion 1.5 and 2.1 respectively. (c) shows how the difference in CAS scores after using ITI-GEN to mitigate gender bias [PITH_FULL_IMAGE:figures/full_fig_p113_5_9.png] view at source ↗
Figure 5.10
Figure 5.10. Figure 5.10: Bias Identification and Mitigation using TIBET and ITI-GEN - Ground Truth. Here, we show ground truth gender differences in the initial set of images before bias mitigation, and after bias mitigation. The reduction in gender bias is in line with what we observe using CAS scores. what we observed based on our CAS scores in [PITH_FULL_IMAGE:figures/full_fig_p113_5_10.png] view at source ↗
Figure 6.1
Figure 6.1. Figure 6.1: This figure demonstrates how a minimal prompt change—from “A [PITH_FULL_IMAGE:figures/full_fig_p117_6_1.png] view at source ↗
Figure 6.2
Figure 6.2. Figure 6.2: Exploring Intersectionality of Biases: Analysing the Top-K concepts shows that pharmacists in Europe and Asia are depicted with different gender distributions. (a) Intersectionality in "a photo of a chef": When we study counterfactuals for "geographical bias", how does the depicted background change? [PITH_FULL_IMAGE:figures/full_fig_p122_6_2.png] view at source ↗
Figure 6.3
Figure 6.3. Figure 6.3: Intersectional Observations using TIBET. We observe that images generated for a chef in Africa may be depicted outdoors (tree) unlike chefs in other regions of the world. Using TIBET, we can study intersectionality by observing counterfactuals along one bias axis, and comparing changes in concepts along another bias axis. This can uncover the interconnectedness between bias axes, showing that modifying o… view at source ↗
Figure 6.4
Figure 6.4. Figure 6.4: An example for which BiasConnect estimates a negative impact of bias mitigation along [PITH_FULL_IMAGE:figures/full_fig_p125_6_4.png] view at source ↗
Figure 6.5
Figure 6.5. Figure 6.5: An overview of BiasConnect. We use a counterfactual-based approach to measure pairwise causality between bias axes. For dependent axes, we measure the causal effect, estimating how bias mitigation on one axis impacts another. 6.4.1.1 Counterfactual Prompts & Image Generation Given an input prompt P and bias axes B = {B1, B2, . . . , Bn}, we generate counterfactual prompts {CF1 i , . . . ,CFj i } for each… view at source ↗
Figure 6.6
Figure 6.6. Figure 6.6: Analyzing bias intersectionality matrices from BiasConnect. (a) Shows how mitigating [PITH_FULL_IMAGE:figures/full_fig_p130_6_6.png] view at source ↗
Figure 6.7
Figure 6.7. Figure 6.7: Sensitivity analysis on BiasConnect. We evaluate the robustness of our approach by analyzing the impact of VQA errors and the effect of the number of images on Intersectional Sensitivity . These high correlations demonstrate that our framework effectively predicts the impact of counterfactual-based interventions on secondary bias dimensions without requiring actual miti￾gation. This capability allows pra… view at source ↗
Figure 6.8
Figure 6.8. Figure 6.8: We compare aggregated causal graphs for four models: Stable Diffusion 1.4, Flux-dev, [PITH_FULL_IMAGE:figures/full_fig_p136_6_8.png] view at source ↗
Figure 6.9
Figure 6.9. Figure 6.9: Additional examples on TIBET dataset (a-b) and Occupation prompt (c) on prompt-level [PITH_FULL_IMAGE:figures/full_fig_p138_6_9.png] view at source ↗
Figure 6.10
Figure 6.10. Figure 6.10: Comparison of real-world and Stable Diffusion 3.5 pairwise causal relationships for [PITH_FULL_IMAGE:figures/full_fig_p139_6_10.png] view at source ↗
Figure 6.11
Figure 6.11. Figure 6.11: The figure illustrates bias interpretations from Bias Connects, combining all pairwise [PITH_FULL_IMAGE:figures/full_fig_p140_6_11.png] view at source ↗
Figure 7.1
Figure 7.1. Figure 7.1: Key challenges in mitigating bias in TTI models. Effective methods must handle how [PITH_FULL_IMAGE:figures/full_fig_p143_7_1.png] view at source ↗
Figure 7.2
Figure 7.2. Figure 7.2: Overview of the InterMit pipeline. Prompts are first analyzed using BiasConnect and [PITH_FULL_IMAGE:figures/full_fig_p147_7_2.png] view at source ↗
Figure 7.3
Figure 7.3. Figure 7.3: Examples of InterMit mitigation. Blue columns are mitigated. (a) and (c) show efficient [PITH_FULL_IMAGE:figures/full_fig_p156_7_3.png] view at source ↗
Figure 7.4
Figure 7.4. Figure 7.4: Modifying D∗ (ideal distribution) in InterMit can have a significant effect on the Intersectional Sensitivity values. Intermit accoounts for this change during the mitigation process and allows user to specify the ideal distribution they want to have. 139 [PITH_FULL_IMAGE:figures/full_fig_p159_7_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

227 extracted references · 51 canonical work pages · 10 internal anchors

  1. [1]

    Moss-Racusin, John F

    Corinne A. Moss-Racusin, John F. Dovidio, Victoria L. Brescoll, Mark J. Graham, and Jo Handelsman. Science faculty’s subtle gender biases favor male students. Proceedings of the National Academy of Sciences, 109(41):16474–16479, 2012

    2012

  2. [2]

    Counterfactuals

    David Lewis. Counterfactuals. Harvard University Press, 1973

    1973

  3. [3]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684–10695, 2022

    2022

  4. [4]

    Towards accuracy-fairness paradox: Adversarial example-based data augmentation for visual debiasing

    Yi Zhang and Jitao Sang. Towards accuracy-fairness paradox: Adversarial example-based data augmentation for visual debiasing. In Proceedings of the 28th ACM International Conference on Multimedia, pages 4346–4354, 2020

    2020

  5. [5]

    Ramaswamy, Sunnie SY

    Vikram V . Ramaswamy, Sunnie SY . Kim, and Olga Russakovsky. Fair attribute classification through latent space de-biasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9301–9310, 2021

    2021

  6. [6]

    Image Counterfactual Sensitivity Analysis for Detecting Unintended Bias

    Emily Denton, Ben Hutchinson, Margaret Mitchell, Timnit Gebru, and Andrew Zaldivar. Image counterfactual sensitivity analysis for detecting unintended bias. arXiv preprint arXiv:1906.06439, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  7. [7]

    Towards fairness in visual recognition: Effective strategies for bias mitigation

    Zeyu Wang, Klint Qinami, Ioannis Christos Karakozis, Kyle Genova, Prem Nair, Kenji Hata, and Olga Russakovsky. Towards fairness in visual recognition: Effective strategies for bias mitigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8919–8928, 2020

    2020

  8. [8]

    Balasubramanian, and Amit Sharma

    Saloni Dash, Vineeth N. Balasubramanian, and Amit Sharma. Evaluating and mitigating bias in image classifiers: A causal perspective using counterfactuals. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 915–924, 2022

    2022

  9. [9]

    Towards causal bench- marking of biasin face analysis algorithms

    Guha Balakrishnan, Yuanjun Xiong, Wei Xia, and Pietro Perona. Towards causal bench- marking of biasin face analysis algorithms. In Deep Learning-Based Face Analytics, pages 327–359. Springer, 2021

    2021

  10. [10]

    Meaningfully debugging model mistakes using conceptual counterfactual explanations

    Abubakar Abid, Mert Yuksekgonul, and James Zou. Meaningfully debugging model mistakes using conceptual counterfactual explanations. In Proceedings of the International Conference on Machine Learning, pages 66–88. PMLR, 2022. 150

    2022

  11. [11]

    Causalm: Causal model explanation through counterfactual language models

    Amir Feder, Nadav Oved, Uri Shalit, and Roi Reichart. Causalm: Causal model explanation through counterfactual language models. Computational Linguistics, 47(2):333–386, 2021

    2021

  12. [12]

    Cavli-using image associations to produce local concept-based explanations

    Pushkar Shukla, Sushil Bharati, and Matthew Turk. Cavli-using image associations to produce local concept-based explanations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3749–3754, 2023

    2023

  13. [13]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient- based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 618–626, 2017

    2017

  14. [14]

    Tibet: Identifying and evaluating biases in text-to-image generative models

    Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. Tibet: Identifying and evaluating biases in text-to-image generative models. Proceedings of the European Conference on Computer Vision, 2024

    2024

  15. [15]

    Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory, and antiracist politics

    Kimberle Crenshaw. Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory, and antiracist politics. In University of Chicago Legal Forum, pages 139–167. 1989

    1989

  16. [16]

    Utilizing Adversarial Examples for Bias Mitigation and Accuracy Enhancement

    Pushkar Shukla, Dhruv Srikanth, Lee Cohen, and Matthew Turk. Utilizing adversarial examples for bias mitigation and accuracy enhancement. arXiv preprint arXiv:2404.11819, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Balasubramanian, Leonid Sigal, and Matthew Turk

    Pushkar Shukla, Aditya Chinchure, Emily Diana, Alexander Tolbert, Kartik Hosanagar, Vineeth N. Balasubramanian, Leonid Sigal, and Matthew Turk. Mitigate one, skew another? tackling intersectional biases in text-to-image models. arXiv preprint arXiv:2505.17280, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Counterfactual visual explanations

    Yash Goyal, Ziyan Wu, Jonas Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In Proceedings of the International Conference on Machine Learning, 2019

    2019

  19. [19]

    Counterfactual explanations without opening the black box: Automated decisions and the gdpr

    Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 2017

    2017

  20. [20]

    Estimating causal effects of treatments in randomized and nonrandomized studies

    Donald B Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5):688–701, 1974

    1974

  21. [21]

    Causality: Models, Reasoning and Inference

    Judea Pearl. Causality: Models, Reasoning and Inference. Cambridge University Press, 2000

    2000

  22. [22]

    Mothilal, Amit Sharma, and Chenhao Tan

    Ramaravind K. Mothilal, Amit Sharma, and Chenhao Tan. Explaining machine learning classifiers through diverse counterfactual explanations. In Proceedings of the 2020 conference on Fairness, Accountability, and Transparency, pages 607–617, 2020

    2020

  23. [23]

    Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel S. Weld. Polyjuice: Gen- erating counterfactuals for explaining, evaluating, and improving models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th 151 International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pa...

    2021

  24. [24]

    Counterfactual fairness

    Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. Counterfactual fairness. Advances in Neural Information Processing Systems, 30, 2017

    2017

  25. [25]

    Counterfactual sensitivity and robustness

    Timothy Christensen and Benjamin Connault. Counterfactual sensitivity and robustness. Econometrica, 91(1):263–298, 2023

    2023

  26. [26]

    Divyansh Kaushik, Eduard Hovy, and Zachary C. Lipton. Learning the difference that makes a difference with counterfactually-augmented data. arXiv preprint arXiv:1909.12434, 2019

    Pith/arXiv arXiv 1909

  27. [27]

    Counterfactuals

    Jonathan Bennett and Kai von Fintel. Counterfactuals. https://plato.stanford. edu/entries/counterfactuals/, 2021. Stanford Encyclopedia of Philosophy, Accessed: 2025-06-12

    2021

  28. [28]

    Argumentation and counterfactual reasoning in parmenides and melissus

    Andr´e Laks. Argumentation and counterfactual reasoning in parmenides and melissus. Archai: Revista de Estudos sobre as Origens do Pensamento Ocidental , 9:45–56, 2012. URL https://www.scielo.br/j/archai/a/Cv7GBGFGwMZDTfQfmWK99Pr/ ?lang=en. Accessed: 2025-06-12

    2012

  29. [29]

    Philosophy in Classical India: The Proper Work of Reason

    Jonardon Ganeri. Philosophy in Classical India: The Proper Work of Reason . Oxford University Press, 2001

    2001

  30. [30]

    The Character of Logic in India

    Bimal Krishna Matilal and Kisor Kumar Chakrabarti. The Character of Logic in India. State University of New York Press, 1998

    1998

  31. [31]

    Georges B.J. Dreyfus. Recognizing Reality: Dharmak ¯ırti’s Philosophy and Its Tibetan Interpretations. SUNY Press, 1997

    1997

  32. [32]

    An Enquiry Concerning Human Understanding

    David Hume. An Enquiry Concerning Human Understanding. A. Millar, 1748. Originally published 1748; many modern editions available

  33. [33]

    Essays on the Active Powers of the Human Mind

    Thomas Reid. Essays on the Active Powers of the Human Mind. Bell & Bradfute, 1788

  34. [34]

    Critique of Pure Reason

    Immanuel Kant. Critique of Pure Reason. Hartknoch, 1781. English translations by Norman Kemp Smith and others available

  35. [35]

    A System of Logic, Ratiocinative and Inductive

    John Stuart Mill. A System of Logic, Ratiocinative and Inductive. Parker, 1843. See Book III for the Method of Difference

  36. [36]

    The Emotions and the Will

    Alexander Bain. The Emotions and the Will. Longman, Green, Longman, and Roberts, 1859

  37. [37]

    Causality: Models, Reasoning, and Inference

    Judea Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, 2 edition, 2009

    2009

  38. [38]

    Efficient search for diverse coherent explanations

    Chris Russell. Efficient search for diverse coherent explanations. Proceedings of the 2020 FAT* Conference, 2019. 152

    2020

  39. [39]

    Multi-objective counter- factual explanations

    Susanne Dandl, Christoph Molnar, Martin Binder, and Bernd Bischl. Multi-objective counter- factual explanations. Machine Learning, 110(4):961–989, 2021

    2021

  40. [40]

    Face: Feasible and actionable counterfactual explanations

    Rafael Poyiadzi, Kacper Sokol, Raul Santos-Rodriguez, Tijl De Bie, and Peter Flach. Face: Feasible and actionable counterfactual explanations. Proceedings of the AAAI Conference on Artificial Intelligence, 34(03):1399–1407, 2020

    2020

  41. [41]

    Model-agnostic counter- factual explanations for consequential decisions

    Amir-Hossein Karimi, Gilles Barthe, Borja Balle, and Isabel Valera. Model-agnostic counter- factual explanations for consequential decisions. In International Conference on Artificial Intelligence and Statistics, pages 895–905. PMLR, 2020

    2020

  42. [42]

    Focus: Flexible optimizable counterfactual explanations for tree ensembles

    Ana-Marija Lucic, Hinda Haned, and Maarten de Rijke. Focus: Flexible optimizable counterfactual explanations for tree ensembles. In Proceedings of the 2022 Conference on Fairness, Accountability, and Transparency, 2022

    2022

  43. [43]

    Hidden assumptions in counterfactual explanations and fairness

    Solon Barocas, Moritz Hardt, and Arvind Narayanan. Hidden assumptions in counterfactual explanations and fairness. Communications of the ACM, 63(9):62–71, 2020

    2020

  44. [44]

    Testing DNN Image Classifiers for Confusion & Bias Errors

    Amit Sharma, Peter Henderson, Shubham Ghosh, Prateek Agarwal, and Matei Zaharia. Certifai: A common framework to provide explanations and recourse through counterfactuals. arXiv preprint arXiv:1905.07831, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  45. [45]

    Good counterfactuals and where to find them: A case-based technique for generating counterfactuals for explainable ai (xai)

    Mark T Keane and Barry Smyth. Good counterfactuals and where to find them: A case-based technique for generating counterfactuals for explainable ai (xai). Proceedings of the AAAI Conference on Artificial Intelligence, 34(03):1364–1371, 2020

    2020

  46. [46]

    Path-specific counterfactual fairness

    Silvia Chiappa. Path-specific counterfactual fairness. In Proceedings of the AAAI conference on Artificial Intelligence, volume 33, pages 7801–7808, 2019

    2019

  47. [47]

    Fairness through causal awareness: Learning causal latent representations for fair prediction

    David Madras, Elliot Creager, Toniann Pitassi, and Richard Zemel. Fairness through causal awareness: Learning causal latent representations for fair prediction. Proceedings of the 35th Conference on Uncertainty in Artificial Intelligence, 2019

    2019

  48. [48]

    Chi, and Alex Beutel

    Sahaj Garg, Vincent Perot, Nicole Limtiaco, Ankur Taly, Ed H. Chi, and Alex Beutel. Counterfactual fairness in text classification through robustness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 219–226, 2019

    2019

  49. [49]

    Conditional counterfactuals for fairness audits in nlp

    Maarten Hall and Hanna Wallach. Conditional counterfactuals for fairness audits in nlp. Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2020

    2020

  50. [50]

    Fliptest: Fairness testing via statistical and semantic counterfactual generation

    Emily Black, Alireza Farhadi, and Osbert Bastani. Fliptest: Fairness testing via statistical and semantic counterfactual generation. Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021

    2021

  51. [51]

    Constructing and measuring visual counterfactuals for grounded bias analysis

    Aparna Balagopalan, Vajira Thambawita, Sigurdur Olafsson, Jesse Hicks, Egil Arnesen, Thomas Goodwin, Eric Breck, and James Wexler. Constructing and measuring visual counterfactuals for grounded bias analysis. In CVPR Workshop on Responsible Vision, 2022. 153

    2022

  52. [52]

    Nonparametric Multiple-Output Center-Outward Quantile Regression

    Jieyu Zhao, Tianlu Wang, Daniel Khashabi, Sameer Singh, and Dan Roth. Evaluating bias and fairness in pretrained multimodal models. arXiv preprint arXiv:2204.11756, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  53. [53]

    A survey on bias and fairness in machine learning

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM Computing Surveys (CSUR), 54(6): 1–35, 2021

    2021

  54. [54]

    Counterfactual fairness through semantic data augmentation

    Tien Wu, Allison Woodruff, Jilin Chen, and Munmun De Choudhury. Counterfactual fairness through semantic data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2021

    2021

  55. [55]

    Actionable recourse in linear classification

    Berk Ustun, Alexander Spangher, and Yang Liu. Actionable recourse in linear classification. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 10–19, 2019

    2019

  56. [56]

    Explaining and harnessing adver- sarial examples

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adver- sarial examples. Proceedings of the International Conference on Learning Representations, 2015

    2015

  57. [57]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019

    Pith/arXiv arXiv 1903

  58. [58]

    Persistent homology with non-contractible preimages

    Robin Morris, Brandon Carter, and William Herlands. Reassessing the robustness of visual counterfactual explanations. arXiv preprint arXiv:2105.08130, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  59. [59]

    Counterfactual data augmen- tation for mitigating gender stereotypes in languages with rich morphology

    Ran Zmigrod, Ryan Cotterell, Jacob Andreas, and Frank Keller. Counterfactual data augmen- tation for mitigating gender stereotypes in languages with rich morphology. In Proceedings of Association of Computational Linguists, 2019

    2019

  60. [60]

    Aguirre, and Thomas Sch ¨on

    Antˆonio H Ribeiro, Koen Tiels, Luis A. Aguirre, and Thomas Sch ¨on. Beyond exploding and vanishing gradients: analysing rnn training using attractors and smoothness. In International Conference on Artificial Intelligence and Statistics, pages 2370–2380. PMLR, 2020

    2020

  61. [61]

    Estimating individual treatment effect: generalization bounds and algorithms

    Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. In Proceedings of the International conference on machine Learning, pages 3076–3085. PMLR, 2017

    2017

  62. [62]

    Learning representations for counterfactual inference

    Fredrik Johansson, Uri Shalit, and David Sontag. Learning representations for counterfactual inference. In Proceedings of the International conference on machine Learning , pages 3020–3029. PMLR, 2016

    2016

  63. [63]

    Causal effect inference with deep latent-variable models

    Christos Louizos, Uri Shalit, Joris M Mooij, David Sontag, Richard Zemel, and Max Welling. Causal effect inference with deep latent-variable models. Advances in Neural Information Processing Systems, 30, 2017

    2017

  64. [64]

    Ganite: Estimation of individual- ized treatment effects using generative adversarial nets

    Jinsung Yoon, James Jordon, and Mihaela Van Der Schaar. Ganite: Estimation of individual- ized treatment effects using generative adversarial nets. In Proceedings of the International Conference on Learning Representations, 2018. 154

    2018

  65. [65]

    Learn: Controllable counterfactual genera- tion via latent disentanglement

    Xinyi Lu, Piji Qi, and William Yang Wang. Learn: Controllable counterfactual genera- tion via latent disentanglement. Findings North American Chapter of the Association for Computational Linguistics, 2022

    2022

  66. [66]

    Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge

    Laura Rieger, Chandresh Singh, W James Murdoch, and Bin Yu. Interpretations are useful: Penalizing explanations to align neural networks with prior knowledge. In Proceedings of the International Conference on Machine Learning, 2020

    2020

  67. [67]

    REIN: A Comprehensive Benchmark Framework for Data Cleaning Methods in ML Pipelines

    Zixuan Zhang, Lei Huang, Hongyu Zhang, and Yulan He. A survey on counterfactual data generation for text. arXiv preprint arXiv:2302.04702, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  68. [68]

    Explanation by progressive exaggeration

    Sahil Singla and Soheil Feizi. Explanation by progressive exaggeration. In Proceedings of the International Conference on Machine Learning, 2019

    2019

  69. [69]

    Explaining image classifiers using statistical fault localization

    Chun-Hao Chang, Cedric Anders, Andreas Muller, and Been Kim. Explaining image classifiers using statistical fault localization. In Advances in Neural Information Processing Systems, 2019

    2019

  70. [70]

    Gan dissection: Visualizing and understanding generative adversarial networks

    David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Josh Tenenbaum, Bill Freeman, and Antonio Torralba. Gan dissection: Visualizing and understanding generative adversarial networks. Proceedings of the International Conference on Learning Representations, 2019

    2019

  71. [71]

    Explaining black-box classifiers using post-hoc explanations: A unified review

    Kedar Dhamdhere, Mukund Sundararajan, and Qiqi Yan. Explaining black-box classifiers using post-hoc explanations: A unified review. arXiv preprint arXiv:1909.13584, 2019

    Pith/arXiv arXiv 1909

  72. [72]

    On robustness of causal claims in text classification

    Shyam Upadhyay, Kai-Wei Chang, and Dan Roth. On robustness of causal claims in text classification. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2021

    2021

  73. [73]

    Experience-Embedded Visual Foresight

    Dibya Mahajan, S Karthik, and Abhishek Mishra. Preserving semantic relations for zero-shot learning. arXiv preprint arXiv:1911.05071, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  74. [74]

    Jon Gauthier, Tianyi Zhang, Bernd Bohnet, and Alexander M. Rush. Counterfactual data augmentation for robust nlp. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2021

    2021

  75. [75]

    Donghyun Kim, Siddhartha Dalmia, and Alan W. Black. Counterfactual reasoning in multimodal language models. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2022

    2022

  76. [76]

    Anchors: High-precision model- agnostic explanations

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Anchors: High-precision model- agnostic explanations. In AAAI Conference on Artificial Intelligence, 2018

    2018

  77. [77]

    Grounding visual explanations

    Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In Proceedings of the European Conference on Computer Vision, 2018

    2018

  78. [78]

    Algorithmic re- course: From counterfactual explanations to interventions

    Amir-Hossein Karimi, Gilles Barthe, Bernhard Sch¨olkopf, and Isabel Valera. Algorithmic re- course: From counterfactual explanations to interventions. In Advances in Neural Information Processing Systems, 2020. 155

    2020

  79. [79]

    Karthik, and Vivek Srikumar

    Dibya Mahajan, Aditya Singh, S. Karthik, and Vivek Srikumar. Generating interpretable coun- terfactual explanations by structural interventions. Proceedings of International Conference on Machine Learning, 2023

    2023

  80. [80]

    Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkata- subramanian

    Michael Feldman, Sorelle A. Friedler, John Moeller, Carlos Scheidegger, and Suresh Venkata- subramanian. Certifying and removing disparate impact. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2015

    2015

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.