Decision Tree Ensemble Classification Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current architectures for decision tree ensemble classification, whether serial or parallel, fall short in optimizing speed, energy consumption, and area efficiency simultaneously.
Innovation Solution
The proposed architecture involves a rank computing block with computation units per feature, using pre-sorted threshold values for efficient binary search and lookup tables to quickly compute node comparisons, reducing the need for complex memory structures and parallel comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If parallel architecture with one comparator per node is used to compute all node comparisons at once, then speed is improved, but device complexity and hardware area increase
Solution Approach 1:
The patent segments the decision tree evaluation process into two distinct phases: a preprocessing phase that computes and stores ranks for all node comparisons, and an inference phase that uses these precomputed ranks for rapid classification. This segmentation allows the complex parallel comparison logic to be moved offline, reducing online hardware complexity while maintaining speed advantages.
Solution Approach 2:
The patent performs preliminary computation of node comparison results during an offline preprocessing stage, storing the outcomes as rank values in memory. During actual classification, these precomputed ranks are directly utilized without requiring real-time comparison operations, thereby eliminating the need for extensive parallel comparator hardware while preserving fast classification performance.
2Device complexity
If serial architecture is used to traverse decision trees from root to leaf, then device complexity is reduced, but speed and productivity deteriorate
Solution Approach 1:
The patent precomputes the results of all node comparisons during an offline phase and stores them as rank values. During classification, the system directly retrieves and uses these precomputed ranks without performing sequential comparisons, thereby achieving fast classification throughput with simple serial hardware that does not require complex tree traversal logic.
3Adaptability or versatility
If one tree processing unit per decision tree is used in the ensemble, then adaptability is improved, but device complexity and hardware area increase
Solution Approach 1:
The patent designs a universal tree processing unit that can handle multiple decision trees through configuration rather than requiring dedicated hardware for each tree. The single processing unit is made adaptable to different tree structures and ensembles by loading appropriate parameters and precomputed ranks, achieving multi-functionality without proportional increases in hardware complexity.
Solution Approach 2:
The patent merges the processing of multiple decision trees into a single configurable tree processing unit. Instead of having separate hardware accelerators for each tree in the ensemble, the system combines them into one unit that processes trees sequentially or in batches using shared resources and precomputed data, thereby reducing overall hardware complexity while maintaining ensemble capability.
4Measurement precision
If all threshold values are stored in memory for each node comparison, then measurement precision is maintained, but loss of substance increases due to memory requirements
Solution Approach 1:
The patent extracts only the essential information needed for comparison—the rank values that represent the relative ordering of threshold values—rather than storing all raw threshold values. This extraction reduces memory requirements significantly while preserving the precision needed for accurate node comparisons, as the rank information suffices for determining comparison outcomes.
Solution Approach 2:
The patent transforms the stored data from raw threshold values to compressed rank representations. By changing the parameter format from continuous threshold values to discrete rank indices, the system achieves the same comparison precision with dramatically reduced memory storage requirements, as ranks encode the necessary ordering information in a compact form.
Data Source
AI summary
In an embodiment a method for classification of a decision tree ensemble include saving of threshold values representative of decision trees in the decision tree ensemble into one group per feature to be classified, sorting of the threshold values of a group according to a threshold index, conducting a node comparison of the threshold values of a corresponding group of threshold values, outputting a rank as a result of the node comparison, wherein the rank represents a encoded address to representative of the threshold value and determining a class of the features to be classified as a function of ranks.