Neural Network OOD Detection Using Layer-Wise Feature Distance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current deep neural networks (DNNs) face challenges in accurately detecting out-of-distribution (OOD) data, particularly in large-scale applications like automatic target recognition (ATR), due to data drift, leading to unreliable predictions and a lack of effective systems for uncertainty estimation.
Innovation Solution
A system and method that utilize a trained neural network to calculate average feature information, classify incoming observations, and determine OOD samples by calculating distance scores between feature information and average feature information, with methods to downsample activation graphs for scalable and efficient OOD detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If topological descriptors are used to monitor neural network performance for OOD detection, then measurement precision is improved, but productivity deteriorates due to poor scaling to large neural networks
Solution Approach 1:
The patent segments the neural network into multiple layers and computes topological descriptors separately for each layer rather than for the entire network at once. This layer-wise segmentation enables parallel computation and reduces the computational burden, allowing OOD detection to scale to large neural networks while maintaining measurement precision.
Solution Approach 2:
The patent computes topological descriptors for a selected subset of layers rather than all layers, or uses a reduced set of topological features. This partial computation approach maintains sufficient OOD detection accuracy while significantly reducing processing time and computational resources required.
2Measurement precision
If topological descriptors are computed for every layer of large CNNs, then measurement precision is improved, but device complexity increases
Solution Approach 1:
The patent extracts topological descriptors from intermediate layer activations rather than computing them for the entire network architecture. By taking out and analyzing only the relevant feature representations from specific layers, the method reduces computational complexity while preserving the essential information needed for accurate OOD detection.
Solution Approach 2:
The patent performs preliminary computation of topological descriptors during the forward pass of neural network inference, integrating the OOD detection computation into the existing inference workflow. This preliminary action avoids separate complex computations and reduces overall device complexity.
Data Source
AI summary
A system may include at least one processor configured to: obtain a trained neural network (NN); obtain or calculate at least one average feature information associated with the trained NN, each of the at least one average feature information including a given average feature information summarizing in-class statistics that each layer of the trained NN uses for a given class; receive an incoming observation influencing a given layer; calculate a corresponding feature information of the incoming observation, the corresponding feature information summarizing statistics of the incoming observation for the given layer; classify the incoming observation as being in the given class; calculate a distance score associated with a distance between the corresponding feature information of the incoming observation and the given average feature information; determine the incoming observation is an out-of-distribution (OOD) sample; and output an alert indicating the incoming observation is OOD and/or discard the incoming observation's classification.