ocl

Learners and base classes for Online Continual Learning.

Online Continual Learning (OCL) is a setting where learners train on a sequence of tasks. A task is a specific concept or data distribution. After training the learner on each task, we evaluate the learner on all tasks.

Continual learning is an important problem to deep learning because these models suffer from catastrophic forgetting, which occurs when a model forgets how to perform well after training on a new task. This is a consequence of a neural network’s distributed representation. The term Continual Learning is often synonymous with overcoming catastrophic forgetting. Non-deep learning methods do not suffer from catastrophic forgetting. Care should be taken to distinguish between online continual learning with and without deep learning.

Online continual learning (OCL) differs from data stream learning because the objective is performance on historic tasks rather than adaptation. Unlike traditional continual learning, OCL restricts training to a single data pass.

>>> from capymoa.classifier import HoeffdingTree
>>> from capymoa.datasets.ocl import TinySplitMNIST
>>> from capymoa.evaluation.ocl import ocl_train_eval_loop
>>> import numpy as np
>>> scenario = TinySplitMNIST()
>>> learner = HoeffdingTree(scenario.schema)
>>> metrics = ocl_train_eval_loop(learner, scenario.train_streams, scenario.test_streams)

The final accuracy is the accuracy on all tasks after finishing training on all tasks:

>>> print(f"Final Accuracy: {metrics.accuracy_final:0.2f}")
Final Accuracy: 0.69

The accuracy on each task after training on each task:

>>> with np.printoptions(precision=2):
...     print(metrics.accuracy_matrix)
[[0.9  0.   0.   0.3  0.  ]
 [0.88 0.9  0.   0.12 0.  ]
 [0.77 0.82 0.62 0.12 0.  ]
 [0.77 0.82 0.6  0.52 0.  ]
 [0.77 0.82 0.57 0.52 0.75]]

Notice that the accuracies in the upper triangle are close to zero because the learner has not trained on those tasks yet. The diagonal contains the accuracy on each task after training on that task. The lower triangle contains the accuracy on each task after training on all tasks.

>>> print(f"Forward Transfer: {metrics.forward_transfer:0.2f}")
Forward Transfer: 0.05

>>> print(f"Backward Transfer: {metrics.backward_transfer:0.2f}")
Backward Transfer: -0.07

Classes

TaskAware	Interface for learners that are aware of the task during evaluation.
TaskBoundaryAware	Interface for learners that are aware of the transition between tasks.