Application of Out-of-distribution Detection for Reliability Determination of Machine Learning Predictions

Chen, Edward; Bao, Han; Dinh, Nam

Permalink

https://hdl.handle.net/2142/121835 Copy

Description

Title

Application of Out-of-distribution Detection for Reliability Determination of Machine Learning Predictions

Author(s)

• Chen, Edward
• Bao, Han
• Dinh, Nam

Issue Date

2023

Keyword(s)

• Out-of-distribution detection
• Reliability
• Neural networks

Abstract

Artificial intelligence (AI) and machine learning (ML), specifically neural networks, have garnered significant interest. However, the reliability of AI/ML models in safety relevant systems is still under debate. Reliability, is defined commonly as the ability of an item to perform as required, without failure, for a given time interval, under given conditions [1, 2]. As neural network-based AI/ML models are generic by construction, functionality cannot be defined explicitly in code (unlike conventional software). Rather, through tuning of non-descript weights and biases, a hidden underlying correlation between observable states and desired target can be ascribed to a model. The function of an AI/ML model is thus designated by the training data which is typically a limited subset of the intended operational space. For a well-trained neural network model (𝑓(⋅)), test samples (𝑥′) that are near, in absolute distance, to training dataset samples (𝑥) typically match performance achieved during training (i.e., 𝑓(𝑥) − 𝑓(𝑥′) ≈ 0 when |𝑥′ − 𝑥| < 𝜖). These tasks are known as interpolation tasks. For these tasks, the AI/ML model can accomplish the desired functionality. However, for extrapolation tasks (i.e., when |𝑥′ − 𝑥| > 𝜖), the performance of the model typically degrades significantly. The reliability of an AI/ML model is thus dependent on the test samples remaining within the range and domain of the training dataset. As it is currently intractable to develop a training set that is a perfect representation of the operational space, the reliability of an AI/ML model is also undetermined. Furthermore, operational spaces can change over time, rendering initially relevant AI/ML models to become irrelevant, or potentially erroneous, over time. The shift of the operational space from interpolation to extrapolation tasks is known as out-of-distribution detection. Thus, in this work, we demonstrate a real-time model-agnostic method to quantify the relative reliability of AI/ML predictions by incorporating out-of-distribution detection on the training dataset. First, the range of interpolative capability of an AI/ML model is determined by assigning a multivariate exponential decay distribution to every training data sample. Test samples that are far from training data samples thereby decay in reliability the further away they are. Second, predictions made by the model are accepted or rejected in real-time to identify model relevancy over sudden or gradual changes in the operational space. The method, named Laplacian distributed decay for reliability (LADDR), determines the differences between an operational sample against the entire training dataset. LADDR is demonstrated on a feedforward neural network-based digital twin used for the prediction of safety significant factors during different loss-of-flow transient conditions. We demonstrate that for out-of-distribution test samples, LADDR can successfully reject poor predictions and reduce overall predictive error. Ultimately, LADDR illustrates how training data can be used as evidence to support the reliability of AI/ML predictions when utilized for conventional interpolation tasks

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121835

Owning Collections