Source-Specific Features Can Consume Model Capacity
When you add images from a different source, the network may spend part of its capacity learning patterns tied to that source, such as a more uniform background or a different camera angle. If those source-specific patterns differ a lot from the target data, fewer representational resources remain for the target task, which can reduce performance.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Shared Visual Structure Makes Extra Source Data Helpful
Source-Specific Features Can Consume Model Capacity
When is it reasonable to mix a large set of public web photos into a product-image training set?
A very large network trained long enough can, in principle, learn from two different image collections without an unavoidable tradeoff.
Complete the capacity requirement before adding more training data
Match each training-scenario element with its role in the learning argument.
Order the reasoning steps for deciding whether to add public web photos to training.
Why a larger model can learn from two image sources
Whether to Use 180,000 Public-Web Photos Alongside 4,000 App Photos
When is it reasonable to mix outside web images with app-captured images?
What does the 40:1 dataset ratio suggest in this example?
Combining a second image source always improves model performance.
Learn After
Auxiliary Data Can Strain Limited Model Capacity
Why can adding studio photos hurt recognition of handheld-device photos?
Features tied to one data source can consume capacity needed for another data source.
Domain-specific patterns can consume a model's _____ capacity.
Match each concept to its role in a capacity tradeoff.
Order the reasoning from added studio photos to possible performance harm.
Why source differences can reduce model capacity for the target domain
Assess the risk of mixing website photos into a phone-photo model
How can source-domain details consume a model’s capacity?
Which difference between two data sources most directly creates the risk of wasted model capacity?
Heavy reliance on source-only patterns can reduce deployment accuracy.