Significance of Slow Concept Drift in AI Models
The constantly changing nature of the world around us poses a challenge for the development of AI models. The figure below shows how visual features of objects evolve significantly over a 10-year period, posing a challenge for object categorization models. In our recent work, “Instance-Conditional Timescales of Decay for Non-Stationary Learning,” we address this challenge and achieve significant gains over other robust learning methods on a range of benchmark datasets for nonstationary learning.
The Challenge of Concept Drift for Supervised Learning
We compared offline training, which iterated over all the training data multiple times in random order, and continual training, which iterated multiple times over each month of data in sequential order. Both approaches have their advantages and disadvantages, but our new method combining the benefits of offline learning and continual learning addresses the challenge of slow concept drift.
Time-Sensitive Reweighting of Training Data
We propose to train a helper model that assigns a weight to each point based on its contents and age. This weight scales the contribution from that data point in the training objective for the model. Our method improves substantially over the no-reweighting baseline as well as many other robust learning techniques, showing broad applicability on a wide range of nonstationary learning challenge datasets sourced from the academic literature that spans different data sources and modalities.
Extensions to Continual Learning
We also consider an interesting extension of our work, adapting our approach to continual learning in a straightforward manner by applying temporal reweighting within the context of each bucket of data being used to sequentially update the model. Our approach consistently beats regular continual learning as well as a wide range of other baselines.
In conclusion, our work on slow concept drift in AI models addresses a significant challenge and achieves significant gains over other robust learning methods on a variety of tasks studying natural concept drift.