Model Variance Compensation Using Multi-Orientation Scene Inputs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ensemble machine-learning methods require significant computational resources, which are often limited in applications such as autonomous vehicles, leading to reduced accuracy and stability in predictions due to 'flickering' outputs from machine-learned models that are not rotationally invariant.
Innovation Solution
Generating multiple orientations of scene data (e.g., rotated, shifted, reflected) and inputting them as a batch into a single machine-learned model to simulate a pseudo ensemble, thereby aggregating outputs to form a final prediction while conserving computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ensemble machine-learning methods are used to improve prediction accuracy, then prediction accuracy is improved, but computational resources increase significantly
Solution Approach 1:
The input scene data is segmented into multiple rotated orientations (e.g., 0°, 60°, 120°, 180°, 240°, 300°) which are then processed individually through the same machine-learned model. This segmentation of input data allows the system to achieve ensemble-like accuracy by considering multiple perspectives without requiring multiple complete model instances, thus improving prediction accuracy while controlling computational resource usage.
2Stability of the object's composition
If ensemble machine-learning methods are used to improve prediction stability, then prediction stability is improved, but computational resources increase significantly
Solution Approach 1:
The input scene data is segmented into multiple rotated orientations (e.g., 0°, 60°, 120°, 180°, 240°, 300°) which are then processed individually through the same machine-learned model. This segmentation of input data allows the system to achieve ensemble-like accuracy by considering multiple perspectives without requiring multiple complete model instances, thus improving prediction accuracy while controlling computational resource usage.
Solution Approach 2:
The outputs from processing multiple rotated orientations are merged by computing their mean value to produce a final aggregated output. This merging of multiple orientation-based predictions stabilizes the overall prediction by averaging out variations, achieving prediction stability similar to ensemble methods while using a single model instance and thus conserving computational resources.
3Use of energy by moving object
If a single machine-learned model is used to conserve computational resources, then computational resources are conserved, but prediction accuracy and stability decrease due to flickering outputs
Solution Approach 1:
The scene data is pre-processed by generating multiple rotated orientations (e.g., 0°, 60°, 120°, 180°, 240°, 300°) before being input into the machine-learned model. This preliminary action of rotating the input data allows the single model to receive diverse perspectives that would otherwise require multiple model instances, thereby improving prediction accuracy and stability without increasing computational resource requirements.
Solution Approach 2:
The outputs from processing multiple rotated orientations are merged by computing their mean value to produce a final aggregated output. This merging of multiple orientation-based predictions stabilizes the overall prediction by averaging out variations, achieving prediction stability similar to ensemble methods while using a single model instance and thus conserving computational resources.
Data Source
AI summary
Techniques for increasing performance of machine-learned models while conserving computational resources generally required by ensemble machine-learning methods are described herein. The techniques may include determining multiple views of a scene that is to be input into a machine-learned model. In some examples, a scene data input may be rotated by 90, 180, and 270 degrees to generate four scene inputs (e.g., 0-, 90-, 180-, and 270-degree rotated inputs) that can be passed through the machine-learned model and the results per scene can be aggregated to determine a final prediction/decision. Similarly, scene inputs may be shifted, reflected, translated, and/or the like before being input into the machine-learned model. The predictions may be associated with one or more objects in the environment that are represented in the scenes.