Neural Network Latent-Space Outlier Detection for Prediction Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for deep neural networks struggle to efficiently evaluate the reliability of predictions, particularly in real-time systems, as they often require multiple inferences or constrain how the network is constructed, limiting their applicability in critical applications like autonomous cars and power plant control systems.
Innovation Solution
A computer-implemented method using a latent variable model constructed from intermediate output values of a deep neural network to determine if an observation is an outlier, allowing for efficient outlier detection without constraining the network construction and enabling real-time applications by projecting observations into a lower-dimensional subspace using techniques like principal component analysis or autoencoders.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple inferences are performed per prediction to evaluate reliability, then prediction reliability assessment is improved, but computational time and system complexity increase
Solution Approach 1:
The method performs preliminary action by constructing the latent variable model during the training phase using intermediate output values from training data. This pre-computed model captures the essential variability patterns, enabling fast outlier detection during prediction without requiring multiple inferences. The model is built once and reused for efficient reliability assessment.
Solution Approach 2:
The method extracts intermediate output values from specific hidden layers of the deep neural network during training and uses these extracted values to construct the latent variable model. By taking out only the necessary intermediate representations rather than performing multiple full network inferences, the system achieves efficient reliability assessment with reduced computational overhead.
2Reliability
If existing reliability evaluation methods are applied to deep neural networks, then prediction reliability can be assessed, but the network construction is constrained
Solution Approach 1:
The latent variable model serves multiple functions: it enables outlier detection for reliability assessment while being applicable to any deep neural network architecture without imposing construction constraints. The model works with intermediate outputs from any hidden layer, making it universally applicable across different network designs, architectures, and training configurations.
Solution Approach 2:
The latent variable model acts as an intermediary between the deep neural network and the outlier detection mechanism. It takes intermediate output values from the network as input and transforms them into a form suitable for reliability assessment, thereby enabling reliability evaluation without directly constraining the network construction while maintaining flexibility.
3Measurement precision
If intermediate output values from hidden layers are used to construct a latent variable model, then outlier detection accuracy is improved, but computational complexity increases
Solution Approach 1:
The method extracts intermediate output values from specific hidden layers rather than processing entire network outputs or performing multiple inferences. This extraction approach captures sufficient information for accurate outlier detection while significantly reducing the computational complexity compared to alternative methods that require full network re-inference or multiple predictions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method for data analysis is provided. The method comprises: obtaining a deep neural network (100) for processing images and at least a part of a training dataset used for training the deep neural network, the deep neural network comprising a plurality of hidden layers, the training dataset including possible observations that can be input to the deep neural network; obtaining first sets of intermediate output values that are output from at least one of the plurality of hidden layers, each of the first sets of intermediate output values obtained by inputting a different one of the possible input images included in said at least the part of the training dataset; constructing a latent variable model using the first sets of intermediate output values, the latent variable model providing a mapping of the first sets of intermediate output values to first sets of projected values in a sub-space that has a dimension lower than a dimension of the sets of the intermediate outputs; receiving an observation to be input to the deep neural network; obtaining a second set of intermediate output values that are output from said at least one of the plurality of hidden layers by inputting the received observation to the deep neural network; mapping, using the latent variable model, the second set of intermediate output values to a second set of projected values; and determining whether or not the received observation is an outlier with respect to the training dataset based on the latent variable model and the second set of projected values.