CSE Seminar - Minglun Gong - Generalization in Machine Learning for Counting Problems
- Calendar
- Mathematics & Statistics
- Date
- 03.29.2023 12:30 pm - 2:00 pm
Description
Date: Wednesday March 29, 2023
Time: 12:30-2:00pm
Speaker: Minglun Gong, University of Guelph
Title: Generalization in Machine Learning for Counting Problems
Abstract: Machine learning relies on training data to learn patterns by minimizing the difference between predicted and annotated output. Generalization in machine learning refers to a model's ability to perform well on unseen data. Achieving generalization is particularly challenging for counting problems, which require a single, non-constrained value as output. This talk presents several approaches to achieve generalization in counting problems. Conventional approaches overfit to similar-density inputs, require location-level annotations, and rely on explicit exemplars for counting different objects. The presented approaches address these issues by handling variations in density, removing the burden of location annotation, and recognizing adaptive exemplars within input images.
To handle large scale variation and density shift in test data, the first approach extracts scale-invariant features through interlayer multi-scale integration and a novel intralayer Scale-invariant Transformation. The second approach uses a tree structure to parse scale information hierarchically and employs a Multi-level Auxiliator to recognize cluttered backgrounds. To remove the model’s dependency on location annotation, a novel and efficient counter is presented, which is trained using total counts only. It models global dependencies of embeddings and regresses total counts using a multigranularity MLP regressor. A self-supervised proxy task called Split-Counting is proposed to evade the limited samples and shortage of spatial hints. To count objects of different categories without explicitly specified exemplars, a zero-shot generalized counter is developed. It uses a pseudo-Siamese structure to learn pseudo exemplar clues from inherent repetition patterns. The general counter is capable of adaptively capturing spatial location hints through a carefully-designed self-similarity learning strategy. Extensive experiments on widely used benchmarks demonstrate the effectiveness of the proposed strategies in generalizing machine learning in counting problems. Ablation studies and visualization analysis are also performed to better understand the impacts and behaviors of individual components.
Location: DSB AB103 & Zoom
Zoom Meeting Info:
Meeting ID: 993 2631 5722
Passcode: 913408
https://mcmaster.zoom.us/j/99326315722?pwd=YVpYVW8yUHpyaU5RTmdjNTI4WGlPUT09
Time: 12:30-2:00pm
Speaker: Minglun Gong, University of Guelph
Title: Generalization in Machine Learning for Counting Problems
Abstract: Machine learning relies on training data to learn patterns by minimizing the difference between predicted and annotated output. Generalization in machine learning refers to a model's ability to perform well on unseen data. Achieving generalization is particularly challenging for counting problems, which require a single, non-constrained value as output. This talk presents several approaches to achieve generalization in counting problems. Conventional approaches overfit to similar-density inputs, require location-level annotations, and rely on explicit exemplars for counting different objects. The presented approaches address these issues by handling variations in density, removing the burden of location annotation, and recognizing adaptive exemplars within input images.
To handle large scale variation and density shift in test data, the first approach extracts scale-invariant features through interlayer multi-scale integration and a novel intralayer Scale-invariant Transformation. The second approach uses a tree structure to parse scale information hierarchically and employs a Multi-level Auxiliator to recognize cluttered backgrounds. To remove the model’s dependency on location annotation, a novel and efficient counter is presented, which is trained using total counts only. It models global dependencies of embeddings and regresses total counts using a multigranularity MLP regressor. A self-supervised proxy task called Split-Counting is proposed to evade the limited samples and shortage of spatial hints. To count objects of different categories without explicitly specified exemplars, a zero-shot generalized counter is developed. It uses a pseudo-Siamese structure to learn pseudo exemplar clues from inherent repetition patterns. The general counter is capable of adaptively capturing spatial location hints through a carefully-designed self-similarity learning strategy. Extensive experiments on widely used benchmarks demonstrate the effectiveness of the proposed strategies in generalizing machine learning in counting problems. Ablation studies and visualization analysis are also performed to better understand the impacts and behaviors of individual components.
Location: DSB AB103 & Zoom
Zoom Meeting Info:
Meeting ID: 993 2631 5722
Passcode: 913408
https://mcmaster.zoom.us/j/99326315722?pwd=YVpYVW8yUHpyaU5RTmdjNTI4WGlPUT09
Go Back
