Learning From Small Labeled Sets by Using Task and Domain Structure

Abstract

This report details their approach and results for the DARPA LwLL or Learning with Less Labels program. The key problem tackled in this report is the problem of learning from small labeled datasets. This is a problem that is frequent in many application areas. It is especially an issue for defense applications where labeled data might be classified and therefore limited. Unfortunately, modern deep learning systems need large numbers of labeled training examples, making them ineffective in the few label regime. In this report, they describe their efforts to produce new kinds of learning machinery that can learn effectively from few labels. Their key insight for this problem is to leverage structure in the task and in the domain. This structure may take the form of domain knowledge about needed invariances. It might also be more amorphous and difficult to specify, but nevertheless it might dictate which classifiers work well in this domain.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Nov 06, 2023
Accession Number
AD1214374

Entities

People

  • Bharath Hariharan

Organizations

  • Cornell University

Tags

DTIC Thesaurus Topics

  • Air Force
  • Air Force Research Laboratories
  • Computer Vision
  • Computing System Architectures
  • Deep Learning
  • Detection
  • Detectors
  • Information Processing
  • Information Systems
  • Jet Propulsion
  • Machine Learning
  • Military Research
  • Network Architecture
  • Pattern Recognition
  • Recognition
  • Supervised Machine Learning
  • Unmanned Vehicles

Fields of Study

  • Computer science

Readers

  • Artificial Intelligence
  • Speech Processing/Speech Recognition.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Neural Networks