Associative Graph Search Using KNN Expansion for Accurate ANN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing KNN search algorithms using reduced dimension vectors in Associative Processing Units (APUs) face accuracy issues and increased I/O operations due to the need for a large number of re-rank operations, leading to reduced performance when dealing with large datasets.
Innovation Solution
Implementing a KNN graph to pre-calculate and store W nearest neighbors for each object, allowing the APU to perform a first KNN search with a reduced number of neighbors, and using the KNN graph in the host processor to expand this number while maintaining accuracy, followed by a second KNN search to find the final K nearest neighbors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If reduced dimension vectors are used in APU for KNN search, then search speed and memory utilization are improved, but search accuracy deteriorates
Solution Approach 1:
The system pre-calculates and stores KNN graphs containing neighbor relationships before the actual search query. This preliminary action creates a structured index that allows the APU to perform fast reduced-dimension searches while the pre-computed graph relationships preserve accuracy information, resolving the contradiction between speed and accuracy
Solution Approach 2:
The KNN graph serves as an intermediary data structure between the reduced dimension vectors in APU and the full dimension vectors in host memory. It stores pre-computed neighbor relationships that bridge the accuracy gap created by dimension reduction, allowing fast APU searches to yield accurate results when combined with graph expansion
2Measurement precision
If a large number of reduced dimension neighbors are returned to maintain accuracy, then search accuracy is improved, but the number of I/O operations and re-rank operations increases
Solution Approach 1:
The KNN graph is pre-computed and stored in host memory before search queries are executed. This preliminary action eliminates the need to perform expensive re-rank operations on large numbers of reduced dimension vectors during the actual search, as the graph already contains pre-validated neighbor relationships
Solution Approach 2:
The search process is segmented into two distinct phases: (1) fast reduced-dimension search in APU to retrieve candidate neighbors using the KNN graph, and (2) selective re-ranking only of those candidates in host memory. This segmentation reduces I/O operations compared to re-ranking all possible reduced dimension neighbors
Data Source
AI summary
An associative graph search system includes a KNN graph determiner to determine in advance W neighbors of each item in a dataset and to store each item and its neighbors in a KNN graph, a reduced dimension vector finder implemented on an associative processing unit (APU) to find a first number of first nearest neighbors of a query vector, the APU operating in a constant complexity irrespective of the size of the number, a result expander to find for each first nearest neighbor, W second nearest neighbors using the KNN graph thereby creating a group of neighbors, and a KNN full dimension vector re-ranker to find a final number of full dimension nearest neighbors of the full dimension query vector from the group of neighbors.