PaperlayerView exact revision on arXiv

Source-linked AI summary

Reliable Visual Question Answering: Abstain Rather Than Answer Incorrectly

Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, Marcus Rohrbach

arXiv:2204.13631v3cs.CV

TL;DR

VQA models can mislead users when they answer uncertain questions, but abstention has received little attention in multimodal research. This paper adds abstention to VQA, evaluates selective models, and proposes Effective Reliability; at 1% risk, softmax-based abstention leaves a state-of-the-art model answering fewer than 7.5% of questions.

  • Problem

    VQA models are generally required to answer every question, although incorrect answers can mislead users and abstention has not received direct attention in VQA research.

    Source: §1 · pp. 1–2

  • Method

    The paper formalizes selective VQA, evaluates coverage and risk, explores calibration and a multimodal selector, and introduces a cost-based Effective Reliability metric.

    Source: §1 · p. 3, §6 · p. 14

  • Results

    At 1% risk, softmax-based abstention lets a state-of-the-art VQA model answer less than 7.5% of questions, while a multimodal selection function consistently improves coverage, particularly at low risk.

    Source: §1 · p. 3, §6 · p. 14

  • Takeaways & Limitations

    Effective Reliability offers a single model-comparison measure that reflects risk and coverage while allowing users to specify the cost of an incorrect answer.

    Source: §1 · p. 3

    Higher VQA accuracy does not guarantee better coverage at low risk, so improving the risk–coverage trade-off also requires better abstention policies.

    Source: §5.2 · p. 10

Abstract

from arXiv · show

Machine learning has advanced dramatically, narrowing the accuracy gap to humans in multimodal tasks like visual question answering (VQA). However, while humans can say "I don't know" when they are uncertain (i.e., abstain from answering a question), such ability has been largely neglected in multimodal research, despite the importance of this problem to the usage of VQA in real settings. In this work, we promote a problem formulation for reliable VQA, where we prefer abstention over providing an incorrect answer. We first enable abstention capabilities for several VQA models, and analyze both their coverage, the portion of questions answered, and risk, the error on that portion. For that, we explore several abstention approaches. We find that although the best performing models achieve over 70% accuracy on the VQA v2 dataset, introducing the option to abstain by directly using a model's softmax scores limits them to answering less than 7.5% of the questions to achieve a low risk of error (i.e., 1%). This motivates us to utilize a multimodal selection function to directly estimate the correctness of the predicted answers, which we show can increase the coverage by, for example, 2.3x from 6.8% to 15.6% at 1% risk. While it is important to analyze both coverage and risk, these metrics have a trade-off which makes comparing VQA models challenging. To address this, we also propose an Effective Reliability metric for VQA that places a larger cost on incorrect answers compared to abstentions. This new problem formulation, metric, and analysis for VQA provide the groundwork for building effective and reliable VQA models that have the self-awareness to abstain if and only if they don't know the answer.

1 Introduction

VQA models can mislead users when they must answer despite uncertainty, yet abstention has received little attention in multimodal VQA. The paper formalizes abstention for VQA, evaluates approaches using risk and coverage, and introduces Effective Reliability.

1 Introduction

  • VQA outputs may guide users’ decisions, and incorrect answers can mislead them; this risk is especially consequential for users with visual impairments who may lack a way to verify answers.The paper gives assistance with daily tasks as an example of VQA’s intended utility.Source: §1 · pp. 1–2
  • The authors identify a gap in VQA research: models generally answer every question, and prior selective-prediction work has not focused on multimodal settings.They distinguish uncertainty specific to a model from questions that are intrinsically difficult or unanswerable.Source: §1 · p. 2
  • The paper frames reliable VQA as selective prediction, where a model either answers or abstains, and evaluates the answered portion through coverage and risk.Coverage is the portion answered; risk is the error on that portion.Source: §1 · pp. 2–3
  • Effective Reliability summarizes the risk–coverage trade-off by rewarding correct answers while assigning a cost to incorrect answers and accounting for abstentions.The paper presents it as a scalar metric for comparing models.Source: §1 · p. 3
  • At 1% risk, softmax-based abstention limits a state-of-the-art VQA model to answering less than 7.5% of questions, while a learned multimodal selector more than doubles coverage.The paper also reports that the selector consistently improves coverage across VQA models and risk levels, particularly at low risk.Source: Figure 1 · p. 2, §1 · p. 3

2 Related Work

Related work addresses VQA uncertainty through several distinct lenses, including intrinsic question difficulty, calibration, and selective prediction. This paper focuses on model-specific uncertainty and learned selection in multimodal VQA.

2 Related Work

  • The authors report that, to their knowledge, existing VQA models lack built-in abstention, and analyze abstention in several prominent models.They distinguish this work from research with non-standard VQA problem statements.Source: §2 · pp. 3–4
  • Prior VQA work detects intrinsically difficult or unanswerable questions, whereas this paper focuses on uncertainty specific to a model.The authors also find that some questions a model abstains on are ambiguous or unanswerable.Source: §2 · p. 4
  • The paper explores vector scaling for selective prediction and learned selection functions for multimodal VQA, building on related calibration and selective-prediction work.It highlights that multimodal VQA selection must account for interactions between modalities and potentially multiple acceptable answers.Source: §2 · p. 4

3 Visual Question Answering with Abstention

The paper extends VQA to let a model answer or abstain, then evaluates selective models through coverage, risk, and a cost-based measure. Its metrics account for VQA’s partial correctness and the trade-off between answering more questions and making more errors.

3.1 Problem Definition

  • A selective VQA model maps each input to either an answer or abstention, with a VQA function producing answers and a selection function deciding whether to answer.The selection function can threshold a confidence estimate of answer correctness.Source: §3.1 · p. 5

3.2 Evaluation Metrics

  • Coverage measures the share of questions answered, while risk measures error among those answered; lowering the selection threshold increases both.The paper evaluates coverage across risk levels and also summarizes the risk–coverage curve.Source: §3.2 · p. 6
  • VQA risk must account for partial correctness because multiple human answers may be acceptable and predictions can receive partial credit.The paper follows the VQA accuracy convention under which agreement with one, two, or three annotators earns partial credit, while agreement with at least four is fully correct.Source: §3.2 · p. 6
  • Effective Reliability rewards at least partially correct answers, penalizes wholly incorrect answers according to a cost, and assigns zero to abstention.The cost is deployment-specific, and the metric is intended to jointly reflect correctness and effectiveness.Source: §3.2 · p. 7
  • Effective Reliability equals VQA accuracy when a model abstains if and only if it is incorrect, and it provides a single score meaningfully correlated with risk–coverage.The user-defined error cost also makes it suitable when specifying an error penalty is more intuitive than setting a target risk.Source: §3.2 · p. 7

4 Selection Functions

The paper compares softmax confidence, calibration, and a learned multimodal selection function for deciding when VQA models should abstain. The selector estimates answer correctness using representations of the image, question, joint features, and answer.

4 Selection Functions

  • MaxProb uses the predicted class’s maximum softmax probability for abstention, while vector scaling calibrates logits before applying MaxProb.Temperature scaling is excluded because it does not change confidence rankings or the risk–coverage curve.Source: §4 · p. 8
  • The Selector is trained to predict answer correctness from image, question, multimodal, and answer representations, with VQA accuracy as the regression target.This setup accounts for partial correctness rather than treating every answer as simply correct or incorrect.Source: §4 · p. 8

5 Experiments

The experiments evaluate VQA abstention methods across models, finding that a learned Selector improves low-risk coverage over MaxProb and that Effective Reliability compares selective models under different error costs. Ablations and examples show the value of multimodal selection and accuracy regression, while results remain short of best-possible coverage.

5.1 Data and Models

  • The study trains VQA models on VQA v2 training data and evaluates selective models on a held-out 106k-example validation split.The split also includes 86k examples for VQA-model validation and Selector training, and 22k for Selector validation; evaluated models range from 70.24% to 74.17% test-std accuracy.Source: §5.1 · pp. 8–9

5.2 Benchmarking Risk and Coverage

  • 2.3×: CLIP-ViL with Selector reaches C@1% 2.3× that of CLIP-ViL with MaxProb, with the largest gains at lower risk tolerances.Selector consistently improves coverage and AUC over MaxProb; Calibration can also improve coverage, but less and less consistently, especially at low risk.Source: §5.2 · pp. 9–10, Table 1 · p. 9
  • Higher VQA accuracy does not guarantee better low-risk coverage, and Selector’s C@1% remains more than 50% below each model’s best possible result.Validation-selected thresholds generalize reasonably well to test-time, but their differing actual risks prevent direct coverage comparisons; Effective Reliability instead compares models at a predefined error cost.Source: §5.2 · pp. 10–11

5.3 Effective Reliability

  • At Φ100, selective models operate at R ≈0.6–1.3% and answer about 5–11% of questions, versus R ≈18–22% and 74–82% coverage at Φ1.Selector has the highest Effective Reliability across models and cost levels; its gain over MaxProb for CLIP-ViL rises from 0.70% at Φ1 to 5.18% at Φ100.Source: §5.3 · pp. 11–12, Table 2 · p. 11
  • Annotation noise slightly changes high-cost Effective Reliability scores, but preserves rankings between models and selection functions.Human evaluation found that the vast majority of predictions penalized as incorrect were properly marked incorrect.Source: §5.3 · p. 12
  • Without abstention, models underperform selective counterparts at c = 1, and their Φc becomes firmly negative as error costs rise.Selective models retain positive Φc even at high cost.Source: Table 2 · p. 11, §5.3 · p. 12

5.4 Selection Function Ablations

  • Multimodal representations outperform unimodal inputs for low-risk coverage, while regressing answer accuracy improves over classifying correctness, especially at low risk.At C@1% and Φ100, unimodal selectors have coverage ≤0.02% and Φ100 < 1%; accuracy regression can rank partially correct answers, unlike correctness classification.Source: §5.4 · pp. 12–13, Table 3 · p. 13

5.5 Qualitative Analysis

  • CLIP-ViL examples show Selector abstaining on an incorrect commonsense answer and a false-premise question, while answering a synonymous answer that MaxProb rejects.The authors suggest that Selector can account for answer similarities that split MaxProb’s softmax confidence across synonyms.Source: §5.5 · pp. 13–14, Figure 3 · p. 14

6 Conclusion

The conclusion frames VQA as selective prediction and reports that a multimodal selector improves low-risk coverage over softmax confidence. It also introduces Effective Reliability to reflect the cost of incorrect answers and abstentions.

6 Conclusion

  • At 1% risk, softmax confidence lets a state-of-the-art VQA model answer <7.5% of questions; multimodal selection raises coverage from 6.8% to 15.6%.Source: §6 · p. 14
  • The paper frames VQA as selective prediction, evaluating coverage and risk and introducing Effective Reliability to account for abstention and the cost of incorrect answers.Source: §6 · p. 14
  • Effective Reliability falls from 70% for perfect abstention to 8% for the best abstention model, despite 70% standard VQA accuracy on the same data.Source: §6 · p. 14

Appendix to Reliable Visual Question Answering: Abstain Rather Than Answer Incorrectly

A Changelog

  • Version 2 added code, model statistics, and further analyses; version 3 corrected evaluation references to use the original VQA v2 annotations.Source: A Changelog · p. 21

B Selector Design Ablations

Selector design ablations find multimodal representations most effective overall, while comparisons across architectures suggest feature choices and training objectives matter more than added architectural complexity.

B Selector Design Ablations

  • The ablations vary question, multimodal, answer, and image representations, including visual features that incorporate question information.The tested Selector uses a 2-layer MLP, and the architecture comparisons use the same features.Source: B Selector Design Ablations · pp. 21–22
  • A 1-layer Selector slightly raises Φ100 but lowers C@1% by about 2.5%; a Transformer performs comparably to the 2-layer Selector.Source: Table 4 · p. 22, B Selector Design Ablations · p. 22
  • Multimodal inputs are most effective overall, with top performers combining answer representations and multimodal representations.In isolation, multimodal representations outperform unimodal representations on C@1%, C@5%, Φ10, and Φ100; answer representations lead on C@10% and C@20%.Source: B Selector Design Ablations · p. 22

C Comparing to Data Augmentation

Using the Dev set to train the Selector is generally more beneficial for reliability than using it to augment MaxProb training, with MaxProb overtaking near full coverage as risk tolerance rises.

C Comparing to Data Augmentation

D Manual Evaluation of Label Noise

Manual label-noise checks found that correcting unfairly penalized answers raises Effective Reliability estimates modestly without changing selection-function rankings; most penalized cases were genuinely incorrect.

D Manual Evaluation of Label Noise

  • The annotated label-noise categories include incomplete ground truth, semantic matches, incomplete predictions, and singular/plural mismatches that remain correct.Source: D Manual Evaluation of Label Noise · pp. 23–24
  • The manual check annotates image-question-answer triples with VQA accuracy 0 and recalculates Φ′100 by treating unfairly penalized answers as abstentions or fully correct.These alternatives provide lower and upper bounds because the correctness of non-matching answers is uncertain.Source: D Manual Evaluation of Label Noise · p. 24
  • About 93% of questions contributing to Φ100 penalties were properly marked incorrect, and accounting for noise did not change selection-function rankings.With CLIP-ViL, treating unfairly penalized cases as abstentions increased Φ′100 by 0.37% for MaxProb, 0.47% for Calibration, and 0.57% for Selector.Source: D Manual Evaluation of Label Noise · p. 24

E Analysis of Selector Decisions

The analysis compares which question difficulties Selector and MaxProb answer and tests whether label noise changes Effective Reliability rankings. Selector answers more questions at every difficulty level; the reported label noise does not change selection-function rankings.

E Analysis of Selector Decisions

  • The analysis groups questions as unimodal, simple-multimodal, or difficult-multimodal, based on whether answering requires the image and how difficult the multimodal question is.The thresholds for both selection functions are chosen on validation to maximize Φ100.Source: E Analysis of Selector Decisions · p. 24
  • Although Table 6 reports a small amount of label noise, correcting its treatment does not change the selection-function ranking by Effective Reliability.The table contrasts the original Φ100 score with Φ′100, which treats some label-error cases as abstentions or correct answers.Source: Table 6 · p. 25
  • Selector answers 1.1× more unimodal, 1.4× more simple-multimodal, and 2.4× more difficult-multimodal questions than MaxProb.The comparison uses test-split decisions with thresholds chosen to maximize Φ100 on validation.Source: E Analysis of Selector Decisions · pp. 24–25

F More Qualitative Analysis

The appendix illustrates differences between Selector and MaxProb decisions, examines how validation-selected thresholds transfer to test, and reports calibration results. Vector scaling improves calibration for all models, while its effect on risk–coverage is described as apparent rather than definitive.

F More Qualitative Analysis

  • Selector sometimes answers correctly when MaxProb abstains, and sometimes abstains on errors that MaxProb answers, but it also answers some incorrect samples.Examples use CLIP-ViL selection functions optimized for Φ100 on validation; some MaxProb abstentions may reflect confidence split across semantically agreeing answers.Source: F More Qualitative Analysis · pp. 25–26, Figure 6 · p. 27

G Threshold Generalization

  • For VisualBERT, validation-selected thresholds yield test-risk differences of at most 0.26%, suggesting reasonable generalization, though not a direct same-risk coverage comparison.Effective Reliability permits a clear model comparison using thresholds chosen from validation.Source: G Threshold Generalization · p. 26, Table 7 · p. 28

H Effect of Model Calibration

  • Vector scaling improves calibration for all models, and the authors state that this can help the risk–coverage trade-off when calibration changes relative confidence rankings.Calibration is evaluated with ECE, using the most frequent human-annotated answer as ground truth for VQA.Source: H Effect of Model Calibration · pp. 26–28

I Additional Dataset Split Details

The experiments use image-disjoint VQA v2 splits, several established VQA architectures, and separately trained calibration and selection functions. The appendix details how data, model representations, and selection-function inputs are organized.

I Additional Dataset Split Details

  • VQA v2 validation data are split for VQA-model training and validation, selection-function training and validation, and selective-model testing, with no images shared across splits.The held-out test set is not seen during training or validation of either component and is used only for evaluation.Source: I Additional Dataset Split Details · p. 28, Table 9 · p. 29

J.1 VQA Models

  • The experiments include Pythia, ViLBERT, VisualBERT, and CLIP-ViL, with the described pretrained or from-scratch training setups and model-specific visual features.Pythia and CLIP-ViL use VQA v2 training data from scratch; ViLBERT and VisualBERT use pretrained encoder weights and are fine-tuned on VQA v2.Source: §J.1 · pp. 29–30, Table 10 · p. 30

J.2 Selection Functions

  • Calibration maps VQA answer logits to calibrated logits, while Selector estimates confidence from the answer, question, image, and multimodal representations.The VQA-model weights are frozen while each selection function is trained.Source: §J.2 · pp. 30–31

K Extended Results

The appendix reports variation across repeated runs for two models, proves a condition under which Effective Reliability equals VQA Accuracy, and relates the framework to conformal prediction and risk control.

K Extended Results

  • Pythia and CLIP-ViL results include means and standard deviations over 10 random seeds, whereas ViLBERT and VisualBERT results are single runs.The authors cite difficulty reproducing the latter models’ pretrained and fine-tuned performance as the reason for using existing MMF checkpoints.Source: K Extended Results · p. 31

L Proof of Lemma 1

  • When a model abstains if and only if it is incorrect, Effective Reliability equals VQA Accuracy.The proof shows the equality under the condition that abstention occurs exactly when accuracy is zero.Source: L Proof of Lemma 1 · p. 32

M Relation to Conformal Prediction

  • Conformal prediction offers probabilistic guarantees for output sets, but defining correct answer sets in VQA is difficult; risk-control guarantees are presented as complementary future work.The passage notes that distinct answers may both be true, while other sets may be contradictory.Source: M Relation to Conformal Prediction · p. 32

Related papers

Loading 2204.13631v3…