Compressed Histograms for Neural Network Activation Characterization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to efficiently characterize activation spaces and detect anomalies in data, particularly in large-scale deep neural networks, due to high memory demands and susceptibility to adversarial attacks.
Innovation Solution
The method involves using pre-trained deep neural networks to evaluate new test data, extracting activations, and generating compressed node-specific histograms to compute p-values, which are then used to identify and retain anomalous data portions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional methods are used to characterize activation spaces in deep neural networks, then measurement precision is improved, but memory usage increases significantly
Solution Approach 1:
The patent extracts only the essential statistical characteristics (mean, standard deviation, min, max) from the full activation distributions, storing these compressed representations instead of complete activation data. This extraction principle maintains measurement precision for anomaly detection while dramatically reducing memory requirements from storing full activation matrices to storing compact statistical summaries.
Solution Approach 2:
Instead of storing complete activation data and computing statistics on-demand (which would require high memory), the patent inverts the approach by pre-computing and storing only the essential statistical moments. This inversion allows precise anomaly detection through comparison of test activations against stored statistics without retaining large volumes of raw activation data.
2Reliability
If comprehensive activation data is retained for anomaly detection, then reliability is improved, but loss of time increases due to larger data processing requirements
Solution Approach 1:
The patent extracts only the critical statistical features (mean, std, min, max) needed for reliable anomaly detection, discarding redundant activation data. This extraction maintains detection reliability by preserving essential distribution characteristics while reducing processing time through smaller data volumes for comparison operations.
Solution Approach 2:
The patent applies partial action by computing and storing only the subset of statistical moments necessary for effective anomaly detection (first two moments for distribution shape, min/max for bounds), rather than retaining or computing all possible activation statistics. This partial computation achieves sufficient reliability with reduced processing overhead.
3Measurement precision
If standard histogram methods are used for activation characterization, then measurement precision is maintained, but device complexity increases
Solution Approach 1:
The patent changes the parameters represented in activation histograms from fine-grained binning to coarse-grained statistical moments (mean, std, min, max). This parameter transformation simplifies the computational device requirements by replacing complex histogram binning operations with straightforward statistical computations, while maintaining sufficient precision for anomaly detection through these aggregated parameters.
Data Source
AI summary
A computer-implemented method, according to one approach, includes: receiving a new set of test data and evaluating the test data using a pre-trained deep neural network. In response to evaluating the test data, activations are extracted from layers of the deep neural network. Compressed histograms are further used to determine p-values for the extracted activations. The p-values are evaluated and portions of the test data that are determined as being anomalous, based at least in part on the evaluation of the p-values, are retained.