Treatment Effect Estimation to Guide Model Optimization in Continual Learning

Jonas Seng, Florian Peter Busch, Matej Zecevic, Moritz Willig

Abstract

Continual Learning (CL) systems are faced with potentially large numbers of tasks to be learned while models have limited capacity. This often makes it impossible to learn all required tasks within a single model without performance degradation. To detect when a model might reach its limits, we propose using Treatment Effect Estimation (TEE) techniques to estimate the effect of training a model on a new task with respect to a global performance measure. This framework allows for proactive capacity planning and informed knowledge management.

Introduction

Continual Learning (CL) is the pursuit of artificial intelligence that can learn a sequence of tasks over time without suffering from "catastrophic forgetting"—the tendency for neural networks to overwrite previously learned information when exposed to new data. Despite significant progress, CL researchers still struggle with two fundamental questions:

  1. Capacity Planning: How can we determine if a model is "full" and whether adding a new task will degrade existing knowledge beyond an acceptable threshold?
  2. Knowledge Management: In memory-constrained environments, how do we identify which specific parameters or memories are least critical to maintain when making room for new information?

Existing approaches often rely on heuristics or computationally expensive replay buffers. This paper presents a vision for framing these challenges through the lens of Causal Inference, specifically using Treatment Effect Estimation (TEE) to ask "what if" questions about model performance before committing to training.

A Causal Framework for Continual Learning

The core innovation of this work is treating the training process (e.g., Stochastic Gradient Descent) not merely as an optimization problem, but as a causal intervention—a $do$-operation—within a system. By framing training as a "treatment," we can apply the tools of causal inference to predict its outcomes.

The Causal Graph

To formalize this, we represent the training process using a causal graph. This graph illustrates the dependencies between the new task, the decision to train, the resulting model parameters, and the final performance metric.

Causal graph representing the training process as an intervention, where $t_i$ is the task, $\tau_i$ is the treatment decision, $\theta_i$ are the parameters, and $l_i$ is the performance.

In this model:

  • $t_i$ represents the new task to be learned.
  • $\tau_i$ is the binary treatment variable: the decision of whether or not to update the model on task $t_i$.
  • $\theta_{i-1}$ and $\theta_i$ represent the model parameters before and after the potential update.
  • $T_{i-1}$ represents the set of previously learned tasks.
  • $l_i$ is the overall performance measure across all tasks.

Estimating Effects: Factual and Counterfactual Reasoning

By adopting this causal perspective, we can utilize two primary types of estimation to guide the CL system:

1. Factual Reasoning (CATE): We can estimate the Conditional Average Treatment Effect (CATE). This answers the question: "If I update my current parameters $\theta_{i-1}$ on this new task $t_i$, how much will the average performance $l_i$ across all tasks change?" This serves as a "safety valve," allowing a system to autonomously decide to skip a task if the predicted negative impact on old knowledge is too high.

2. Counterfactual Reasoning: This allows the system to ask: "What would my performance be right now if I had never trained on Task X?" By identifying which previous tasks or parameters have the lowest counterfactual impact on current performance, the system can perform "causal pruning"—surgically removing low-impact knowledge to free up capacity for new tasks.

Implementation via Influence Functions

Actually retraining a model multiple times to observe these effects is computationally prohibitive. To address this, we propose the use of Influence Functions. These mathematical tools from robust statistics allow us to approximate how a model's parameters would shift if a specific task were added or removed without actually performing the gradient descent. This provides a computationally feasible way to estimate the causal effects of training interventions.

Discussion and Future Work

While this theoretical framework provides a rigorous way to manage model capacity, several hurdles remain for empirical implementation. Influence functions, while powerful, can be unstable and difficult to scale to high-dimensional neural network parameter spaces. Furthermore, estimating the distribution of parameters $p(\theta_i | t_i, \theta_{i-1})$ is challenging due to the non-convex nature of deep learning landscapes.

Future research will focus on moving from this conceptual vision to empirical validation. This will likely involve using representation learning to compress parameter spaces into more manageable dimensions for causal estimation, ultimately leading to CL systems that can autonomously manage their own knowledge and capacity limits.

Summary updated

Abstract

Continual Learning systems are faced with a potentially large numbers of tasks to be learned while the models employed have only limited capacity available, which makes it potentially impossible to learn all required tasks within a single model. In order to detect on when a model might break we propose to use treatment effect estimation techniques to estimate the effect of training a model on a new task w.r.t. some suitable performance measure.
BibTeX
@inproceedings{seng2023treatment,
  author = {Jonas Seng and Florian Peter Busch and Matej Zecevic and Moritz Willig},
  title = {{Treatment Effect Estimation to Guide Model Optimization in Continual Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=cbo9QaQv95},
}
Figure