Core Concepts
Objects are learned through prediction, mimicking human infant abilities.
Abstract
Objects are fundamental for mental representation.
Humans perceive objects in 3D environments without supervision.
Models lack the ability to learn like infants.
A novel network architecture learns object segmentation, 3D locations, and depth.
Objects are treated as latent causes for efficient predictions.
Prediction error guides the brain to improve segmentation accuracy.
The model, OPPLE, integrates prediction approaches for object perception.
Results show OPPLE outperforms other models in object segmentation.
OPPLE learns depth perception and 3D object localization.
The model relaxes assumptions to improve learning performance.
Dataset generation and evaluation methods are detailed.
OPPLE's performance is compared to other models in object segmentation, localization, and depth perception.
Stats
"Our model outperforms all compared models on both metrics (Table 3)."
"Object viewing angles are better estimated (correlation r = 0.86) than distances (r = 0.51)."
"OPPLE also learns to infer depth (distance of pixels from the camera)."
"We trained and tested a version of our network in which the rules of rigid body motion and self-motion induced apparent motion are replaced by neural networks with two FC layers."
Quotes
"Objects are treated as latent causes for efficient predictions."
"Prediction error guides the brain to improve segmentation accuracy."
"OPPLE outperforms all compared models on both metrics."