EXPLORING EXAMPLE INFLUENCE IN CL

Problem Statement:

• To understand example influence, one classic technique is the Influence Function (IF), which leverages the chain rule from a test objective to training examples. However, directly applying it leads to high computational complexity.

Research Goals:

• Explore a reasonable influence from each training example to Stability (S) and Plasticity (P), and apply example influence to continual learning training.
• Propose a meta‑learning algorithm, MetaSP, to simulate IF and avoid expensive chain‑rule calculations.

Proposed Approach:

• Hold a pseudo‑update with per‑example perturbations, then use two small validation sets (from old and new data) to compute gradients on those perturbations—these gradients estimate the example’s influence on S and P.

• Fuse S‑ and P‑influences via a Pareto‑optimal mixing weight γ using MGDA, yielding a influence vector I*.

• Use I* to regularize model updates (upweight helpful examples, downweight harmful ones) and to rank examples for rehearsal buffer selection.

Summarized Steps:

Pseudo‑Update Simulation

• From current parameters θ, perform a one‑step “virtual” update adding small per‑example perturbation E to the batch loss.

Compute S‑ and P‑Aware Influences

• Measure how the pseudo‑update changes losses on two validation sets (V_old and V_new), and take gradients w.r.t. E.

SP Fusion via Pareto Optimality

• Solve a two‑objective optimization for γ via MGDA to combine S‑ and P‑gradients into I*.

Leveraging Influences

• Model Update: Regularize SGD by subtracting I* >L.
• Rehearsal Selection: Rank examples by I* and store the top ones, dropping the least helpful.

Experimental Setup:

• Benchmarks: Split‑CIFAR‑10, Split‑CIFAR‑100, Split‑Mini‑ImageNet (class‑ and task‑incremental).
• Model: ResNet‑18 from scratch; replay batch size 32; use 10% of buffer & seen data as V_old/V_new.
• Training: 50 epochs per task; varied replay memory sizes.

Results:

Limitations:

• Relies on a replay buffer, which may not suit privacy‑sensitive or zero‑memory settings.
• Pseudo‑update and dual validation add extra compute overhead.

Strengths:

1. Efficiency: Simulates IFs with first‑order gradients—no Hessian inversions—scalable to deep nets.
2. Principled Fusion: Uses multi‑objective optimization to balance remembering old tasks and adapting to new ones.

Future Work:

• Explore lighter‑weight validation schemes or proxy metrics to reduce compute overhead.