Abstract
In this work, we present CRIB (Continual Recognition Inspired by Babies), a synthetic incremental object learning environment that can produce data that models visual imagery produced by object exploration in early infancy. CRIB is coupled with a new 3D object dataset, Toys-200, that contains 200 unique toy-like object instances, and is also compatible with existing 3D datasets. Through extensive empirical evaluation of state-of-the-art incremental learning algorithms, we find the novel empirical result that repetition can significantly ameliorate the effects of catastrophic forgetting. Furthermore, we find that in certain cases repetition allows for performance approaching that of batch learning algorithms. Finally, we propose an unsupervised incremental learning task with intriguing baseline results.
Repeated Exposure Ameliorates Catastrophic Forgetting
In general, incremental learning algorithms only see one concept at a time and never again, usually leading to catastrophic forgetting. Using CRIB to repeatedly sample the concept class of object instances, we find that repeated exposure to object instances leads to a gradual improvement in accuracy that eventually reaches batch performance.
CRIB Learning Environment Samples
Instance Learning Exposures
Instance Testing Image Data
CRIB is a data generator accessible using a simple Python API that given a query for instance id and sequence length, renders a video clip of a smoothly rotating object overlaid on a dynamic background. Data is returned as 3xHxW array and can be saved to disk or directly used as input data for training. Each generated clip has randomized lighting, object scale and object rotation trajectory. This allows for virtually unlimited data.