In the world of machine learning, the prevailing wisdom is almost always "more data is better." We spend enormous resources collecting, cleaning, and labeling datasets under the assumption that every new observation brings us closer to the truth. But what if some data is actually making your model worse?
A new paper from researchers at ETH Zürich, University of Cambridge, and Karolinska Institutet challenges this "more is more" dogma. They introduce Forgetting to Improve (F2I), a principled framework for identifying and removing detrimental data points to improve model accuracy and uncertainty quantification.
TL;DR
Standard models often struggle when their assumptions don't perfectly match reality ("model mismatch"), causing some data points to actually degrade performance. This paper proposes a method to greedily "forget" these harmful samples, leading to significantly better predictions and more reliable uncertainties in tasks like Bayesian Optimization and Reinforcement Learning.
What Problem Does This Paper Address?
The authors focus on Gaussian Processes (GPs), which are popular because they provide not just a prediction, but a measure of how certain they are. This uncertainty is usually split into two types: aleatoric (inherent noise in the data) and epistemic (uncertainty due to a lack of data).
The "Aha!" moment of the paper is realizing that in the real world, we often have model mismatch. For example, if you use a smooth kernel to model data that has a sudden, sharp spike, the GP might try to explain that spike by increasing its estimate of global noise. This makes the model overly "cautious" and uncertain even in regions where the data is actually well-behaved.
Standard active learning algorithms are designed to only add data. They have no mechanism to look back and say, "Actually, that observation from ten minutes ago is confusing me; I should ignore it."
Methods
The core of the paper is the F2I algorithm. Instead of just fitting all available data, F2I identifies a task-optimal subset of data.
- Influence Functions: The authors derive mathematical "influence functions" that quantify how much an individual data point contributes to the model's uncertainty in a specific target region.
- Greedy Removal: The algorithm starts with the full dataset and iteratively removes the "worst" sample—the one whose removal most reduces the predictive uncertainty in the target region.
- Aleatoric & Epistemic Tracking: The influence measure accounts for both the information gain (epistemic) and the estimated noise level (aleatoric). This prevents the model from accidentally throwing away "load-bearing" samples that are essential for understanding the underlying structure.
- Sequential Extensions: They extend this to Forgetful BO and Forgetful RL, where the "target region" isn't fixed but moves dynamically as an optimizer searches for a maximum or an agent learns a policy.
Datasets
The researchers evaluated F2I across a diverse set of benchmarks:
- Transductive Regression: Synthetic functions with varying lengthscales and real-world datasets including Dow Jones (stock market data), Boston Housing, CASP (protein structure), and Yacht Hydrodynamics.
- Bayesian Optimization: Eight standard test functions (e.g., Ackley, Branin, Griewank) used to simulate complex optimization landscapes.
- Reinforcement Learning: The Cartpole Balance task, specifically testing if a simple linear model can succeed by "forgetting" non-linear experiences.
Key Insights and Ideas
- Less is More: In the presence of model mismatch, a model trained on a subset of data can be strictly better than one trained on the whole set.
- Uncertainty as a Guide: Since we often don't have ground-truth labels in our target region, we can't use "error" to decide what to remove. Instead, F2I uses the model's own internal uncertainty as a signal for pruning.
- Task-Specific Modeling: A data point might be useful for a global fit but detrimental for a specific task (like finding a local maximum). F2I allows the model to specialize its knowledge.
- Principled "Forgetting": Unlike simple outlier detection which uses heuristics, F2I is grounded in the formal influence of data on the GP's posterior distribution.
Key Results
- Superior Calibration: F2I drastically improves how well the model's "confidence" matches reality. In the Dow Jones example, F2I reduced the Calibration Error (CE) from 0.0664 to a near-perfect 0.0002.
- Faster Optimization: In Bayesian Optimization, "Forgetful BO" consistently achieved lower regret (found better solutions faster) than standard GPs and even other robust methods like Warped GPs.
- Enabling Simple Models for Complex Tasks: In the Cartpole RL task, a standard model-based approach failed because it couldn't handle non-linearities. Forgetful RL succeeded by using a simple linear model and just "forgetting" the non-linear parts of the state space.
Limitations and Future Work
While powerful, the authors acknowledge several hurdles:
- Computational Cost: Each removal step requires updating the model and recalculating influences. While they use efficient approximations, it is still more expensive than a standard GP.
- Theoretical Constraints: Their near-optimality guarantees rely on specific "curvature" conditions of the objective function which might be hard to verify for arbitrary real-world data.
- Beyond GPs: The current math is tailored for Gaussian Processes. A major future direction is investigating how similar "principled forgetting" could be applied to other parametric models, like deep neural networks.
In summary, "Forgetting to Improve" provides a rigorous foundation for the idea that sometimes, to learn better, a model first needs to learn what to ignore.
Summary updated