Machine Learning on Images with Noisy Human-centric Labels

Data Skeptic · Advanced ·📐 ML Fundamentals ·9y ago

Key Takeaways

Machine learning on images with noisy human-centric labels using a novel architecture to distinguish presence and relevance, enabling web-scale datasets to be useful for training

Original Description

When humans describe images, they have a reporting bias, in that the report only what they consider important. Thus, in addition to considering whether something is present in an image, one should consider whether it is also relevant to the image before labeling it. Ishan Misra joins us this week to discuss his recent paper Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels which explores a novel architecture for learning to distinguish presence and relevance. This work enables web-scale datasets to be useful for training, not just well groomed hand labeled corpora.
Watch on YouTube ↗ (saves to browser)
Sign in to unlock AI tutor explanation · ⚡30

Playlist

Uploads from Data Skeptic · Data Skeptic · 32 of 60

← Previous Next →
1 Data Skeptic book giveaway contest winner selection
Data Skeptic book giveaway contest winner selection
Data Skeptic
 2 OpenHouse - Front end and API overview
OpenHouse - Front end and API overview
Data Skeptic
 3 OpenHouse Crawling with AWS Lambda
OpenHouse Crawling with AWS Lambda
Data Skeptic
 4 [MINI] Logistic Regression on Audio Data
[MINI] Logistic Regression on Audio Data
Data Skeptic
 5 Data Provenance and Reproducibility with Pachyderm
Data Provenance and Reproducibility with Pachyderm
Data Skeptic
 6 [MINI] Primer on Deep Learning
[MINI] Primer on Deep Learning
Data Skeptic
 7 Big Data Tools and Trends
Big Data Tools and Trends
Data Skeptic
 8 [MINI] Automated Feature Engineering
[MINI] Automated Feature Engineering
Data Skeptic
 9 The Data Refuge Project
The Data Refuge Project
Data Skeptic
 10 [MINI] The Perceptron
[MINI] The Perceptron
Data Skeptic
 11 [MINI] Feed Forward Neural Networks
[MINI] Feed Forward Neural Networks
Data Skeptic
 12 Data Science at Patreon
Data Science at Patreon
Data Skeptic
 13 [MINI] Backpropagation
[MINI] Backpropagation
Data Skeptic
 14 [MINI] GPU CPU
[MINI] GPU CPU
Data Skeptic
 15 OpenHouse
OpenHouse
Data Skeptic
 16 [MINI] Generative Adversarial Networks
[MINI] Generative Adversarial Networks
Data Skeptic
 17 [MINI] AdaBoost
[MINI] AdaBoost
Data Skeptic
 18 [MINI] The Bootstrap
[MINI] The Bootstrap
Data Skeptic
 19 [MINI] Dropout
[MINI] Dropout
Data Skeptic
 20 [MINI] Gini Coefficients
[MINI] Gini Coefficients
Data Skeptic
 21 [MINI] Random Forest
[MINI] Random Forest
Data Skeptic
 22 [MINI] Heteroskedasticity
[MINI] Heteroskedasticity
Data Skeptic
 23 [MINI] ANOVA
[MINI] ANOVA
Data Skeptic
 24 Urban Congestion
Urban Congestion
Data Skeptic
 25 [MINI] The CAP Theorem
[MINI] The CAP Theorem
Data Skeptic
 26 Unstructured Data for Finance
Unstructured Data for Finance
Data Skeptic
 27 Detecting Terrorists with Facial Recognition?
Detecting Terrorists with Facial Recognition?
Data Skeptic
 28 Predictive Models on Random Data
Predictive Models on Random Data
Data Skeptic
 29 [MINI] Entropy
[MINI] Entropy
Data Skeptic
 30 [MINI] F1 Score
[MINI] F1 Score
Data Skeptic
 31 Causal Impact
Causal Impact
Data Skeptic
 ▶ Machine Learning on Images with Noisy Human-centric Labels
Machine Learning on Images with Noisy Human-centric Labels
Data Skeptic
 33 The Library Problem
The Library Problem
Data Skeptic
 34 Stealing Models from the Cloud
Stealing Models from the Cloud
Data Skeptic
 35 Data Science at eHarmony
Data Science at eHarmony
Data Skeptic
 36 Multiple Comparisons and Conversion Optimization
Multiple Comparisons and Conversion Optimization
Data Skeptic
 37 Election Predictions
Election Predictions
Data Skeptic
 38 [MINI] Calculating Feature Importance
[MINI] Calculating Feature Importance
Data Skeptic
 39 MS Connect Conference
MS Connect Conference
Data Skeptic
 40 Music21
Music21
Data Skeptic
 41 The Police Data and the Data Driven Justice Initiatives
The Police Data and the Data Driven Justice Initiatives
Data Skeptic
 42 Studying Competition and Gender Through Chess
Studying Competition and Gender Through Chess
Data Skeptic
 43 [MINI] Goodhart's Law
[MINI] Goodhart's Law
Data Skeptic
 44 Trusting Machine Learning Models with LIME
Trusting Machine Learning Models with LIME
Data Skeptic
 45 [MINI] Leakage
[MINI] Leakage
Data Skeptic
 46 Predictive Policing
Predictive Policing
Data Skeptic
 47 Mutli-Agent Diverse Generative Adversarial Networks
Mutli-Agent Diverse Generative Adversarial Networks
Data Skeptic
 48 [MINI] Convolutional Neural Networks
[MINI] Convolutional Neural Networks
Data Skeptic
 49 Unsupervised Depth Perception
Unsupervised Depth Perception
Data Skeptic
 50 [MINI] Max-pooling
[MINI] Max-pooling
Data Skeptic
 51 MS Build 2017
MS Build 2017
Data Skeptic
 52 Activation Functions
Activation Functions
Data Skeptic
 53 Doctor AI
Doctor AI
Data Skeptic
 54 [MINI] The Vanishing Gradient
[MINI] The Vanishing Gradient
Data Skeptic
 55 CosmosDB
CosmosDB
Data Skeptic
 56 Estimating Sheep Pain with Facial Recognition
Estimating Sheep Pain with Facial Recognition
Data Skeptic
 57 [MINI] Conditional Independence
[MINI] Conditional Independence
Data Skeptic
 58 MINI: Bayesian Belief Networks
MINI: Bayesian Belief Networks
Data Skeptic
 59 Project Common Voice
Project Common Voice
Data Skeptic
 60 [MINI] Recurrent Neural Networks
[MINI] Recurrent Neural Networks
Data Skeptic

This video discusses a novel approach to machine learning on images with noisy human-centric labels, enabling web-scale datasets to be useful for training. The approach involves distinguishing between presence and relevance in images, and can be used to improve model performance.

Key Takeaways
  1. Identify the problem of reporting bias in human-centric labels
  2. Develop a novel architecture to distinguish presence and relevance
  3. Train models on web-scale datasets
  4. Evaluate model performance with noisy labels
💡 Noisy human-centric labels can be handled by distinguishing between presence and relevance in images, enabling web-scale datasets to be useful for training

Related Reads

📰
What Is MLIR and Why Does It Exist?
Learn about MLIR, a intermediate representation for machine learning models, and its purpose in optimizing ML workflows
Dev.to · Fedor Nikolaev
 📰
Why Choosing the Right Machine Learning Development Company Matters More Than the AI Model
Choosing the right machine learning development company is crucial for turning AI investments into measurable results, as it can make or break the success of AI projects
Medium · Machine Learning
 📰
Data privacy in AI training: federated learning, differential privacy, and synthetic data
Learn how federated learning, differential privacy, and synthetic data preserve data privacy in AI training, and why they matter for secure machine learning
Dev.to AI
 📰
Data Preprocessing: Encoding and Feature Scaling in Machine Learning
Learn to preprocess data by encoding and scaling features for better machine learning model performance
Medium · Machine Learning
Up next
Is Python Dead in 2026?| Truth About Python in AI Era | 90 Days Roadmap @FameWorldEducationalHub
FAME WORLD EDUCATIONAL HUB
Watch →