Towards Robust Evaluations of Continual Learning

Motivation

• Recent research work relate the problem of continual learning to increasing the dataset difficulty.
• Current experimental set-ups and evaluations are misleading as they are biased towards prior focused approaches.
• Need to encourage continual strategies that work when all the fundamental desideratum are enforced and not just a subset of them.

Contributions

• Thorough analysis of flaws in existing evaluations used by our community.
• Empirically show that current evaluations are biased towards prior focused approaches.
• Propose fundamental desiderata for future evaluations of continual learning strategies.
• Can be applied irrespective of the dataset.
• Propose new experimental set ups to overcome the issues of exiting ones.

Proposed Desiderata

- Core Desiderata

• A: Cross-task resemblance : Data for new task must resemble from old task enough to produce confident predictions of old classes sometimes early in training. Permuted MNIST violates this property.
• B: Shared output head
• C: No test time assumed task labels
• D: No unconstrained re-training on old tasks
• E: More than two tasks : The more the number of tasks that a CL strategy can deal with, the better.

- Other desiderata

• Unclear task demarcation
• Continuous tasks
• Overlapping tasks
• Long task sequences
• Time/Compute/Memory constraints
• Strict privacy guarantees

Critical analysis of existing evaluations

• Permuted MNIST doesn’t represent real world scenarios for continual learning as it violates desiderata A.
• Multi-headed version of split MNIST requires knowledge of the task and classes in eah task apriori.
• Prior based approaches tend to perform much better in multi-headed versions than in single headed versions.
• Two task transfer is not a realistic evaluation in continual learning as algorithms might fail with more number of tasks.

Empirical analysis of existing evaluations

• Experiments considering all the five core desiderata show that prior focussed methods suffer the most.
• Missing any desiderata can lead to blindspots in the evaluation pipeline.
• Model uncertainity can be used to detect task changes.
• Training time and accuracy must be traded off. Memory can be treated like time here.

Insights and Conclusion

• Current evaluations are misleading as they are biased towards prior focussed approaches.
• Continual learning strategies should be evaluated by taking into consideration all the proposed core five desiderata. Single-headed split MNIST is the most simplistic experiment set-up.