Machine Learning Training with Key-Frame Proximity Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face data imbalance issues due to underrepresented classes in training datasets, leading to biased predictions towards more frequent classes, particularly in applications like video game control and image classification, where infrequent actions or attributes are crucial for accurate performance.
Innovation Solution
A filtering technique is applied to attribute sequences to provide context information by modifying values of training frames near key frames, transforming the binary classification problem into a likelihood estimation, enabling the model to learn proximity to key frames and improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard training datasets are used with underrepresented classes, then training efficiency is maintained, but prediction accuracy for infrequent attributes deteriorates
Solution Approach 1:
The filter is applied to the attribute sequence before training the machine learning model, modifying the training dataset in advance to include context information about proximity to key frames. This preliminary processing ensures that the model receives balanced training signals for both frequent and infrequent attributes, preventing bias towards majority classes while maintaining training efficiency.
2Measurement precision
If the model is trained to recognize infrequent attributes, then prediction accuracy for rare events improves, but the model complexity increases
Solution Approach 1:
The filter acts as an intermediary between the raw attribute sequence and the machine learning model. It transforms the original attribute sequence by incorporating context information about proximity to key frames, creating an enhanced training dataset that enables the model to recognize infrequent attributes without increasing model structural complexity. The filter handles the complexity of rare event detection externally.
3Measurement precision
If context information is added to training frames, then the model's ability to predict infrequent attributes improves, but data processing time increases
Solution Approach 1:
The filter is applied offline to the attribute sequence during the data preparation phase, before model training begins. This preliminary processing creates an enhanced training dataset with context information already embedded, allowing the model to learn from enriched data without incurring additional processing time during inference. The time cost is shifted to the one-time preprocessing stage.
Data Source
AI summary
A system for training a machine learning model comprises a receiving unit configured to receive first data comprising an attribute sequence providing a sequence of values each indicating whether a particular attribute is associated with a respective one of a sequence of training frames, training frames so associated being referred to as key frames, a training unit configured to train the machine learning model, using a training dataset, to generate behaviour for an agent to predict occurrence of the particular attribute within a sequence of action frames, and a filtering unit configured to apply a filter to the attribute sequence to generate the training dataset, wherein the filter modifies values of the attribute sequence for at least some training frames proximate to a given key frame, to provide context information in the training dataset indicating proximity to a key frame for frames within the sequence of training frames.