acceptodds
Withdrawn from ICLR 2027

TIE: A Training–Inversion–Exclusion Framework for Visually Interpretable and Uncertainty-Guided Out-of-Distribution Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep neural networks often struggle to recognize when an input lies outside their training experience, leading to unreliable and overconfident predictions. Building dependable machine learning systems therefore requires methods that can both estimate predictive uncertainty and detect out-of-distribution (OOD) samples in a unified manner. In this paper, we propose TIE: a Training-Inversion-Exclusion framework for visually interpretable and uncertainty-guided anomaly detection that jointly addresses these challenges through iterative refinement. TIE extends a standard n-class classifier to an (n+1)-class model by introducing a garbage class initialized with Gaussian noise to represent outlier inputs. Within each epoch, TIE performs a closed-loop process of training, inversion, and exclusion, where highly uncertain inverted samples reconstructed from the just-trained classifier are excluded into the garbage class. Over successive iterations, the inverted samples transition from noisy artifacts into visually coherent class prototypes, providing transparent insight into how the model organizes its learned manifolds. During inference, TIE rejects OOD inputs by either directly mapping them to the garbage class or producing low-confidence, uncertain misclassifications within the in-distribution classes that are easily separable, all without relying on external OOD datasets. A comprehensive threshold-based evaluation using multiple OOD metrics and performance measures such as AUROC, AUPR, and FPR@95%TPR demonstrates that TIE offers a unified framework for robust anomaly detection and calibrated uncertainty estimation (UE) achieving near-perfect OOD detection with   0 FPR@95%TPR when trained on MNIST or FashionMNIST and tested against unseen datasets.

Primary area:
interpretability and explainable AI
Keywords:
Anomaly Detection, Interpretability, Inversion

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_atahOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

TIE adds a "garbage" class to an image classifier. During training, a generator is trained against the classifier; the paper calls this "network inversion". The generated samples that the classifier is unsure about are added to the garbage class. At test time, the model rejects anything put in the garbage class, and the rest are checked using a Mahalanobis score. All models are small CNNs trained from scratch on MNIST, FashionMNIST, SVHN, CIFAR-10/100 and TinyImageNet. The paper claims a near-0% FPR on MNIST and FashionMNIST. For CIFAR-10, it reports an average FPR of 7.59, compared to 7.46 for Dream-OOD.

Critical strengths

A simple recipe: it needs no outlier data and no pretrained generator, and it adds no test-time cost.

Critical weaknesses

First and foremost, the experiments are very lacking. It feels like the paper is from ten years ago, when we were still working on classifiers on MNIST and CIFAR-10. Even introduction to computer vision classes don't use MNIST and CIFAR-10 right now and start with ImageNet.The paper basically ignores how OOD detection is done today. Modern top detectors use pretrained models. Zero-shot detectors built on vision-language models (VLMs) do not even need training on the original data:

  • CIFAR-10. Zero-shot CLIP detectors get 94.8 near-OOD and 99.3 far-OOD AUROC (AdaNeg, NeurIPS 2024, OpenOOD benchmark). This is better than a ResNet-18 trained on CIFAR-10, which gets 88.0 / 90.7 with MSP (OpenOOD v1.5).
  • CIFAR-100 vs. CIFAR-10. Pretrained ViTs raise AUROC from 85% to over 96% (Fort et al., NeurIPS 2021).
  • ImageNet-1K. The zero-shot NegLabel reaches 94.21 AUROC (ICLR 2024).

TIE instead trains small CNNs from scratch for 20 epochs. They reach only 85.5% accuracy on CIFAR-10, and the images are never larger than 64×64. TIE does not compare itself with any of these modern methods. Its "≈0 FPR" claim comes from MNIST and FashionMNIST, a setting that is close to saturated: with MNIST as the training data, plain MSP already reaches 98.5 far-OOD AUROC in OpenOOD v1. TIE also requires training the classifier and a generator from scratch, so it is unclear whether it can work with pretrained models.

W2. Limited novelty:

  • The extra OOD class. Handling unknown inputs with "an additional background or garbage class" is an old idea, and Dhamija et al. (NeurIPS 2018) found it "generally insufficient when previously unseen classes are encountered". Further examples:
    • Lee et al. (ICLR 2018, "Training confidence-calibrated classifiers for detecting out-of-distribution samples", App. E) tested an explicit (K+1)-th OOD class and found it often significantly worse than their confidence loss.
    • A classifier with an extra noise class is a standard baseline (Ren et al., NeurIPS 2019).
    • Vernekar et al. (NeurIPS 2019 workshop) already train an (n+1)-class classifier on generated OOD samples, on MNIST and FashionMNIST.
  • Filling that class with generated samples. Training the extra class on samples from a generator that targets the classifier is the (K+1)-class "fake" discriminator of semi-supervised GANs (Salimans et al., NIPS 2016). It is also Nguyen et al.'s fooling-image retraining (CVPR 2015).
  • The new part. Uncertainty-guided exclusion is never tested on its own (no ablation study). It is also missing from the provided notebook, which adds every inverted sample to the garbage class and sets the KL and cosine loss weights to zero.

None of these prior works is cited.

Questions for authors
  1. How does TIE compare with zero-shot VLM detectors (such as MCM or NegLabel) and with pretrained models on the OpenOOD CIFAR-10 and ImageNet-1K benchmarks? Can TIE be used on top of a pretrained model?
  2. What percentage of normal test images does the garbage class wrongly reject?
Reviewer_QaMgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes TIE, a Training–Inversion–Exclusion framework for OOD detection and uncertainty estimation without external OOD training data. The method augments an n-class classifier with a garbage class initialized by Gaussian noise. After each training epoch, a conditional generator inverts the current classifier; inverted samples whose softmax-based uncertainty exceeds a dynamic threshold estimated from ID training uncertainties are added to the garbage class, yielding an iterative training–inversion–exclusion loop. At inference time, OOD inputs are first rejected by the garbage class, while the remaining samples are handled by a second threshold-based detector, with Mahalanobis distance used for the reported overall FPR. The paper presents qualitative visualizations of how inverted samples and latent features evolve, and evaluates the method on MNIST, FashionMNIST, SVHN, CIFAR-10, CIFAR-100, and TinyImageNet-200 using custom CNNs, including comparisons with a static-garbage baseline and with VOS, NPOS, and DREAM-OOD. Overall, however, I find the contribution insufficiently novel and the current empirical evidence insufficiently reliable/presented for acceptance.

Critical strengths
  1. The training–inversion–exclusion loop is conceptually coherent and easy to interpret. It gives the model a way to construct its own negative/reject samples without requiring an external OOD dataset, and the visual evolution of the inverted samples provides an intuitive diagnostic of how the learned class manifolds and rejection region change during training.
  2. The paper makes a useful effort to verify that the iterative component matters rather than relying only on the added garbage class. The comparison against the static-garbage “No TIE” baseline shows a large difference in direct OOD rejection, and the paper also reports convergence/sensitivity analyses and a direct comparison with representative synthetic-outlier methods (VOS, NPOS, and DREAM-OOD).
Critical weaknesses
  1. My main concern is that the methodological novelty is not sufficiently established or isolated. TIE combines several established ingredients: classifier/network inversion, uncertainty/confidence-based filtering, synthetic outlier generation for decision-boundary regularization, and an explicit reject/garbage class. The inversion loss itself is adopted from prior network-inversion work. More importantly, a closely matching public work titled “Network Inversion for Uncertainty-Aware Out-of-Distribution Detection” (arXiv 2025; AAAI 2026 Student Abstract) already appears to describe the defining recipe of the present submission: an (n+1)-class classifier, Gaussian-noise garbage initialization, per-epoch inversion, exclusion of uncertain reconstructions into the garbage class, iterative retraining, and garbage-class OOD rejection. The submission does not cite or distinguish itself from this prior public version. The relationship should be clarified and the genuinely new contribution of the present manuscript should be made explicit. Even setting this overlap aside, the experiments do not isolate the value of the full feedback loop against natural simpler alternatives such as one-shot inversion negatives, inversion samples selected without the iterative exclusion/retraining loop, or a matched n+1 classifier trained with generic synthetic negatives. This matters because the reported CIFAR-10 comparison is competitive but not clearly better than DREAM-OOD on average (7.59% vs. 7.46% average FPR).
  2. The empirical evidence and presentation quality are not reliable enough to support the paper’s main claims. The proposed UE score simplifies algebraically to K/(K−1)·(1−Σ_i p_i²), i.e., a normalized Gini-impurity/confidence score from a single softmax vector, yet the paper presents this as uncertainty estimation without standard calibration or selective-prediction evaluation (e.g., ECE, NLL, Brier score, risk–coverage) or quantitative comparison with the UQ methods discussed in related work. The two-tier “overall FPR” also multiplies the fraction of OOD samples surviving the garbage stage by FPR@95%TPR computed only on the remaining OOD subset. This is not automatically the standard end-to-end FPR@95%TPR of the complete detector because the overall ID TPR also depends on ID samples rejected at stage one. The near-zero values in Table 3 therefore need end-to-end ROC/FPR95 verification on the full ID/OOD populations. Beyond these protocol issues, a substantial portion of the visual evidence is difficult to assess because multiple figures are noticeably blurred/low-resolution, and several appear heavily synthetic or AI-stylized rather than conventionally exported scientific plots/diagrams. Regardless of their provenance, this materially reduces readability and confidence in the presentation of the evidence. Finally, most experiments rely on custom CNNs and cross-dataset far-OOD pairs rather than standard ResNet/WRN backbones and modern near-/far-OOD benchmarks.
Questions for authors
  1. Please clarify the relationship between this submission and the prior public work “Network Inversion for Uncertainty-Aware Out-of-Distribution Detection” (arXiv 2025 / AAAI 2026 Student Abstract), and state precisely what methodological, theoretical, or empirical contribution is new in the present paper. Could you also add matched ablations against simpler synthetic-negative variants (e.g., one-shot inversion, no uncertainty filtering, no iterative feedback, and generic generated negatives)?
  2. Can the authors evaluate the complete two-stage detector end-to-end using standard AUROC/AUPR/FPR@95%TPR on the full ID and OOD test sets, with thresholds chosen for 95% overall ID TPR, and separately report the fraction of ID samples rejected by the garbage stage? Please also provide standard uncertainty/calibration evaluations and clarify the provenance and generation pipeline of the main figures, especially those that appear blurred or heavily stylized, given the AI-use statement that generative AI was used only for language editing. High-resolution/vector originals should be supplied.
Reviewer_KmWVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds a garbage class to a classifier to help it reject OOD inputs. In each epoch, it trains the classifier, uses a generator to produce images through network inversion, and adds uncertain generated images to the garbage class. The uncertainty threshold is calculated from ID training examples. At test time, the model first rejects garbage-class predictions, then applies a Mahalanobis detector to the remaining inputs. The experiments cover six image datasets and show how the generated images change during training.

Critical strengths
  1. The method does not need an external OOD dataset. In Table 1, it detects many more OOD examples than the baseline trained with a fixed set of Gaussian-noise garbage samples. This suggests that updating the garbage samples during training is useful.

  2. The generated images shown across epochs make the training process easier to understand. The experiments also cover six datasets, including CIFAR-100 and TinyImageNet, and Table 4 includes comparisons with VOS, NPOS, and DREAM-OOD.

Critical weaknesses
  1. It is unclear whether the reported overall FPR corresponds to keeping 95% of ID examples. Section 6.3 evaluates OOD examples missed by the garbage class. Section 6.4 then multiplies their FPR by the fraction of OOD examples that reach this second stage. The multiplication is correct at a fixed threshold, but ID examples can also be rejected at the first stage. A second-stage TPR of 95% therefore does not by itself give a final ID TPR of 95%. Please report both FPR and ID TPR for the complete detector. There is also a calculation that needs checking: for MNIST→FashionMNIST, Tables 1 and 2 give 89.5% garbage rejection and 0.83% FPR on the remaining OOD examples. Equation (1) gives 0.08715%, but Table 3 reports 0.00%.

  2. The experiments leave several questions about where the improvement comes from. The static-garbage baseline removes both inversion and adaptive exclusion, so it does not show how much each part helps. Table 4 also needs to state whether the methods use the same backbone, preprocessing, training budget, and threshold selection. The differences in average FPR are small enough that these choices could affect the comparison. Du et al., VOS: Learning What You Don't Know by Virtual Outlier Synthesis (2022), and Tao et al., Non-Parametric Outlier Synthesis (2023), are already compared in Table 4. It would help to explain what generating images through inversion adds to their methods, which generate outliers in feature space. Two other relevant approaches are joint classifier/generator training in Lee et al., Training Confidence-calibrated Classifiers for Detecting Out-of-Distribution Samples (2018), and density regularization with transformed ID examples and noise in Huang et al., Density-driven Regularization for Out-of-distribution Detection (2022). Please discuss how the generated samples and training losses differ. Any added experiments should use the same backbone and data access, with comparable training budgets. Finally, UE measures how concentrated the predicted probabilities are. Higher confidence on generated images does not show that the probabilities are better calibrated. The appendix gives the algorithm but does not add calibration measurements or separate tests of the method's components.

Questions for authors
  1. What are the measured FPR and ID TPR after both rejection stages? When choosing the second-stage threshold for 95% TPR, do you use all ID examples or only those that pass the first stage? Please also explain how the threshold is chosen without tuning on test OOD data, and check the MNIST→FashionMNIST calculation.

  2. Were the Table 4 baselines rerun with the same experimental setup? Please give their backbones and training data. Can you show separate results for inversion, uncertainty-based sample selection, and a fixed versus changing exclusion threshold?

  3. Do the uncertainty estimates improve on held-out data? NLL, Brier scores, and reliability plots would help check the calibration claim. A risk-coverage curve would separately show whether the uncertainty score helps identify predictions the model should abstain from.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.