Learning Similarity Metrics with Invariance Properties

Posted in Science on September 16, 2008


Learning Similarity Metrics with Invariance Properties

Lecture slides:

  • Learning Similarity Metrics with Invariances
  • Learning an Invariant Dissimilarity Metric
  • Dissimilarity Metric for Face Recognition
  • Siamese Architecture
  • Dissimilarity Metric vs Traditional Classification
  • Trainable Metric vs Other Dimensionality Reduction Methods
  • Trainable Metrics vs handcrafted invariances
  • Siamese Architecture for Comparing TimeSeries Data
  • 1D Convolutional Net (TDNN)
  • Examples
  • Siamese Architecture
  • Probabilistic Training: Maximum Likelihood
  • Solution?
  • Another Loss Function
  • Loss Function
  • Examples of Loss Functions
  • Loss Function: SquareExponential
  • Face Verification datasets: AT&T, FERET, and AR/Purdue
  • Face Verification datasets: AT&T, FERET, and AR/Purdue
  • Face Verification datasets: AT&T, FERET, and AR/Purdue
  • Face Verification dataset: AR/Purdue
  • Preprocessing
  • Centering with a Gaussianblurred face template
  • Alternated Convolutions and Subsampling
  • Architecture for the Mapping Function Gw(X)
  • Internal state for genuine and impostor pairs
  • Gaussian Face Model in the output space
  • Dataset for Verification - Verification Results
  • Classification Examples
  • lInternal State
  • DrLim: Dimensionality; Reduction by Learning an Invariant Mapping
  • “Traditional” Manifold Learning
  • Learning a FUNCTION from input to output
  • Learning an INVARIANT FUNCTION from input to output
  • Learning Invariant Manifolds with EBMs
  • Step 1: Building a Neighborhood Graph
  • Step 2: Pick a Parameterized Family of Function
  • Step 3: Pick a Loss function and Minimize it w.r.t. W
  • Architecture
  • Architecture and loss function
  • Loss function
  • Mechanical Analogy
  • MNIST Dataset
  • MNIST Handwritten Digits. Sanity Check
  • Architecture of the Gw(X) Function:
  • Alternated Convolutions and Subsampling
  • Learning a mapping that is invariant to shifts
  • Simple Experiment with Shifted MNIST
  • Shifted MNIST: LLE Result
  • Shifted MNIST: Injecting Prior Knowledge
  • Discovering the Viewpoint Manifold
  • Generic Object Detection and Recognition with Invariance to Pose and Illumination
  • Data Collection, Sample Generation
  • NORB Dataset: LLE
  • Automatic Discovery of the Viewpoint Manifold with Invariant to Illumination
  • NORB Dataset: Learned Hidden Units

Author: Yann Le Cun, New York University

Watch Video

Tags: Science, Lectures, Computer Science, Machine Learning, VideoLectures.Net, Structured Output