MPI-INF Logo
Campus Event Calendar

Event Entry

What and Who

A Data-Driven Approach to the Approximate Nearest Neighbor Problem.

Andreas Kalavas
Max-Planck-Institut für Informatik - D1
AG1 Mittagsseminar (own work)
AG 1  
AG Audience
English

Date, Time and Location

Tuesday, 6 August 2024
13:00
30 Minutes
E1 4
024
Saarbrücken

Abstract

The nearest neighbor search (NNS) problem and its variants have captivated scientists for the past fifty years. This problem is prevalent in applications such as data compression, data mining, and machine learning. Although numerous solutions have been proposed, few offer theoretical guarantees while simultaneously optimizing the structure for the input data. This challenge arises because adapting the structure for a specific dataset can expose vulnerabilities to adversarial queries, leading to suboptimal performance.


We propose a new model to solve the approximate near neighbor problem, aiming to balance theoretical guarantees with dataset adaptability. Our approach involves storing the input point set in a binary tree structure, optimized for
performance on a fixed dataset and query distribution. The core idea of our approach is to extract useful information from the point set to enhance our structure, but to halt this extraction when it becomes potentially harmful. When this happens, we transition to an existing technique that offers theoretical guarantees. This strategy allows us to leverage the efficiency of our model while avoiding elements that could degrade performance. Thus, our structure remains data-driven while maintaining theoretical guarantees. Finally, we conduct experiments to demonstrate our algorithm’s adaptability to a dataset while preserving its theoretical guarantees. Specifically, we assess our model on the MNIST dataset, by performing queries on model instances built on different sized samples. We then compare our results with those of linear search.

Contact

Nidhi Rathi
+49 681 9325 1134
--email hidden

Virtual Meeting Details

Zoom
897 027 2575
passcode not visible
logged in users only

Nidhi Rathi, 08/02/2024 15:32 -- Created document.