Invariant Transformation and Resampling based Epistemic-Uncertainty Reduction
| Authors | Sha Hu |
| Year | 2026 |
| Field | AI / Agents |
| arXiv | 2602.23315 |
| Download | |
| Categories | cs.AI |
Abstract
An artificial intelligence (AI) model can be viewed as a function that maps inputs to outputs in high-dimensional spaces. Once designed and well trained, the AI model is applied for inference. However, even optimized AI models can produce inference errors due to aleatoric and epistemic uncertainties. Interestingly, we observed that when inferring multiple samples based on invariant transformations of an input, inference errors can show partial independences due to epistemic uncertainty. Leveraging this insight, we propose a "resampling" based inferencing that applies to a trained AI model with multiple transformed versions of an input, and aggregates inference outputs to a more accurate result. This approach has the potential to improve inference accuracy and offers a strategy for balancing model size and performance.
Engineering Breakdown
Plain English
This paper addresses a fundamental problem in AI inference: even well-trained models produce errors due to both aleatoric uncertainty (inherent randomness in data) and epistemic uncertainty (model knowledge gaps). The authors discovered that when you apply invariant transformations to an input—variations that preserve the true label—the resulting inference errors show partial independence from epistemic sources. Their solution is a resampling-based inference method that transforms a single input multiple ways, runs inference on each transformed version using the same trained model, and aggregates the outputs to produce a more accurate final prediction. The approach is practical because it requires no model retraining and works with any existing trained model.
Core Technical Contribution
The core insight is that epistemic uncertainty manifests differently across invariant transformations of the same input, meaning errors from these transformed inferences can be partially decorrelated and thus better aggregated. Prior work typically treats uncertainty as monolithic or requires ensemble methods with multiple independently trained models. This paper's novelty is recognizing that a single model's predictions on intelligently transformed inputs can provide uncertainty reduction comparable to ensembles, without the computational and storage overhead of training multiple models. The resampling mechanism is the technical contribution—a lightweight inference-time technique that exploits the geometric structure of epistemic uncertainty in the input transformation space.
How It Works
Given a trained AI model and an input sample, the method first generates multiple transformed versions of that input using invariant transformations—operations that preserve the ground truth label (e.g., rotation, scaling, or augmentation within a safe range). Each transformed input is passed through the same trained model to generate individual predictions. The key mechanism is that epistemic errors (due to model uncertainty) behave differently across these transformations because they depend on regions of the model's learned decision boundary, which are sampled differently by each transformation. The method then aggregates these multiple predictions—typically via voting, averaging, or a learned combination function—to produce a final output that is more robust to epistemic uncertainty. The critical assumption is that invariant transformations are known or can be derived for the task, and that the model's error patterns across these transformations are sufficiently independent to enable meaningful aggregation.
Production Impact
For production systems, this offers immediate accuracy gains without retraining or deploying multiple models, directly addressing the cost and complexity of ensemble methods. On inference latency, you pay a multiplicative cost: if you apply K transformations, inference time increases by factor K (e.g., 3x slower for 3 transformations), which may be prohibitive for real-time systems like recommendation or ad serving. The memory footprint remains unchanged since only one model is loaded; you're trading compute for accuracy. Integration is straightforward—wrap the model's inference in a transformation loop and aggregation function—making it practical to retrofit into existing pipelines. For high-stakes applications (medical imaging, autonomous driving), the accuracy improvement from epistemic uncertainty reduction could justify the latency overhead; for latency-critical systems, you'd need careful benchmarking of accuracy gains versus inference time trade-offs.
Limitations and When Not to Use This
The paper's effectiveness depends entirely on the availability of meaningful invariant transformations, which may not exist or be obvious for all domains (e.g., raw categorical data, time series with no clear augmentations, or structured data like graphs where transformations change semantics). It does not address aleatoric uncertainty, which is irreducible and represents inherent noise in the data—the method only reduces epistemic (model) uncertainty. The approach assumes that errors across transformed inputs are sufficiently independent; this assumption breaks down if the model has systematic biases or if transformations correlate with model failure modes. There's also a practical question about computational feasibility: for expensive models (large language models, computer vision on high-res images), applying 5-10 transformations at inference time may be economically infeasible. Follow-up work needed includes: characterizing which transformation sets are optimal for given domains, theoretical bounds on uncertainty reduction under different transformation schemes, and methods to adaptively select the number of transformations based on model confidence.
Research Context
This work builds on decades of research in uncertainty quantification and ensemble methods, particularly work distinguishing aleatoric from epistemic uncertainty (Kendall & Gal, 2017 style frameworks). It relates to test-time augmentation (TTA) techniques used in computer vision, which also apply transformations at inference, though typically for classification calibration rather than uncertainty reduction. The paper contributes to the emerging area of inference-time adaptation—methods that improve predictions without modifying model weights. It's positioned between Monte Carlo Dropout-style approaches (which sample uncertainty through stochastic forward passes) and ensemble methods (which train multiple models), offering a middle ground. The research opens directions into: optimal transformation design for epistemic uncertainty, theoretical analysis of when partial independence holds, and applications to out-of-distribution detection where epistemic uncertainty is highest.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::