YoVDO

Exploring the Origins and Prevalence of Texture Bias in Convolutional Neural Networks

Offered By: Launchpad via YouTube

Tags

Convolutional Neural Networks (CNN) Courses Deep Learning Courses Hyperparameters Courses Model Interpretability Courses

Course Description

Overview

Explore the concept of texture bias in convolutional neural networks through this 26-minute Launchpad video. Delve into the origins and prevalence of this phenomenon, examining its potential problems and impact on model performance. Investigate how datasets influence texture or shape bias, and analyze the effects of training objectives and hyperparameters. Compare shape bias in various models, including high-performing ImageNet models and more "brain-like" architectures. Gain insights into the layers of AlexNet and the implications of different pre-processing techniques. Understand the interplay between shape and texture information in neural networks, and consider how this knowledge can be applied to improve model design and performance.

Syllabus

Intro
Shape vs. Texture Bias
Why Texture Bias could cause problems?
How does the dataset influence texture or shape bias?
Geirhos Style Transfer cat
Does the model know about shape even though it's talking about texture?
Alexnet Layers
Effects of Training Objective
Training Objective Results
High Performing ImageNet Models
Shape Bias in (more) "Brain-like" Models
Effects of pre-processing (random vs. center crop)
How do hyper parameters influence shape bias?


Taught by

Launchpad

Related Courses

Machine Learning Interpretable: interpretML y LIME
Coursera Project Network via Coursera
Machine Learning Interpretable: SHAP, PDP y permutacion
Coursera Project Network via Coursera
Evaluating Model Effectiveness in Microsoft Azure
Pluralsight
MIT Deep Learning in Life Sciences Spring 2020
Massachusetts Institute of Technology via YouTube
Applied Data Science Ethics
statistics.com via edX