Negative Backward Transfer (NBT)

• Negative Backward Transfer (NBT) is the negative change in a model’s performance on previously learned tasks after acquiring new tasks, quantitatively indicating catastrophic forgetting.
• It arises from factors like task dissimilarity, gradient misalignment, and limited model capacity, and is measured using backward transfer metrics in continual and multi-task learning frameworks.
• Mitigation strategies such as gradient-based projection methods and uncertainty-weighted regularization effectively reduce NBT, enhancing the robustness of continual learning systems.

Negative Backward Transfer (NBT) refers to the degradation of a machine learning model’s performance on previously learned tasks as a result of acquiring new tasks in continual, multi-task, or lifelong learning scenarios. Formally, NBT is quantified by negative values of the backward transfer metric (BWT), which measures the average change in accuracy (or error) on earlier tasks after completing a training sequence. NBT serves as a direct signal of catastrophic forgetting and is a central obstacle in the development of robust continual learning systems across various modalities and architectures, including deep networks, quantum machine learning, and LLMs.

1. Formal Definition and Measurement of Negative Backward Transfer

Let $T$ denote the total number of tasks, and $A_{i,j}$ (or $R_{i,j}$) denote the test performance (such as accuracy) on task $j$ after the model has finished training through task $i$. The backward transfer (BWT) is defined as

$$\mathrm{BWT} = \frac{1}{T-1} \sum_{k=1}^{T-1} (A_{T,k} - A_{k,k})$$

or equivalently with $R_{T,k}$ and $R_{k,k}$. If $\mathrm{BWT} < 0$, the average performance on old tasks has declined after learning all tasks—this is termed negative backward transfer (NBT) (Situ et al., 2022, Lin et al., 2022, Wu et al., 2024).

In multi-task learning (MTL), a related notion of negative transfer occurs for a task $t$ if its loss under MTL, $L_t^{\mathrm{MTL}}$, exceeds that of a single-task baseline $L_t^{\mathrm{STL}}$: $$\Delta_t = L_t^{\mathrm{MTL}} - L_t^{\mathrm{STL}} > 0$$ (Silva et al., 2020).

Summary Table: NBT Measurement

Scenario	Metric	NBT Signature
Continual Learning	$\mathrm{BWT} < 0$	Forgetting
Multi-task Learning	$\Delta_t > 0$ for some $t$	Per-task degradation
Prompt Tuning (LLM)	$\mathrm{BWT} < 0$ or FWT $< 0$	Loss of past skills

2. Theoretical Causes of NBT

Negative backward transfer arises when updates made to optimize new tasks interfere with, or overconstrain, parameters useful for previous tasks. Theoretical frameworks (Benavides-Prado et al., 2022, Lin et al., 2022) identify key contributing factors:

• Task Dissimilarity: When new tasks are not sufficiently related to previous ones, enforcing invariances suitable for the new task may degrade prior solutions (Benavides-Prado et al., 2022).
• Gradient Misalignment: If the gradient of a new task update projects strongly into the subspace of an old task but the alignment is negative, parameter updates can increase the old task loss—a direct mechanism for NBT (Lin et al., 2022).
• Model Capacity and Overlap: Capacity constraints or overly shared feature representations can exacerbate interference. Large VC-dimension and insufficient data amplify the potential for NBT (Benavides-Prado et al., 2022).

Within gradient-based learning, two criteria delineate regimes:

• Sufficient Projection: Strong component of new-task gradient within old-task subspace.
• Positive Correlation: New and old task gradients are aligned ($\langle g_j, g_t \rangle > 0$), enabling beneficial update; otherwise, NBT is possible (Lin et al., 2022).

3. Algorithms for Avoiding or Mitigating NBT

Multiple algorithmic strategies have been proposed to prevent or reduce NBT by constraining parameter updates or structuring knowledge transfer:

Gradient-based Projection Methods

Gradient Episodic Memory (GEM): At each update, project the new-task gradient $g$ so that for all prior tasks $k<t$, the loss does not increase (i.e., $\langle \tilde g, g_k \rangle \geq 0$). This is achieved via quadratic programming, ensuring no NBT (no loss increase) for any previous task (Situ et al., 2022).

• Empirical results in quantum continual learning show GEM delivers positive BWT (forgetting is eliminated), outperforming both naive sequential training and EWC regularization (Situ et al., 2022).

CUBER: Builds on a geometric analysis, categorizing each layer and each old-new task pair into regimes: “no overlap,” “forward transfer only,” and “backward transfer possible.” Only when gradients are positively correlated is the model permitted to make updates that affect old-task parameters, allowing for selective positive backward transfer. Otherwise, projections or penalties prevent drift (Lin et al., 2022).

• CUBER achieves the first demonstration of positive BWT (elimination of NBT) in fixed-capacity, no-replay settings across standard computer vision benchmarks.

Regularization and Task Uncertainty

Aleatoric Uncertainty-weighted MTL (U-AMTFL): Introduces learned, task-specific uncertainty parameters $\sigma_t$ into the loss: $$L_{\mathrm{total}} = \sum_t \frac{1}{2\sigma_t^2} L_t + \log \sigma_t + R(\Theta,B,A)$$ Tasks with high uncertainty contribute less to the shared feature space, reducing the risk of NBT due to unreliable auxiliary tasks (Silva et al., 2020).

4. Application to Prompt Tuning and Lifelong Learning

Prompt tuning for LLMs exhibits unique NBT patterns. Techniques such as per-task prompt freezing avoid parameter overwrite, but do not guarantee positive forward transfer.

Similarity Heuristic Lifelong Prompt Tuning (SHLPT): Introduces an instance-level similarity estimator, partitioning prior tasks into "similar" and "dissimilar" sets. Mixtures of similar prompts initialize new task learning, while contrastive loss components drive apart features of dissimilar tasks. Prompt-freezing removes catastrophic forgetting (BWT $=0$), while SHLPT's similarity-guided mechanism uniquely achieves positive FWT and mitigates NBT even on hard, negatively correlated task sequences (Wu et al., 2024).

5. Error Bounds and Theoretical Guarantees

Theoretical work (Benavides-Prado et al., 2022) establishes uniform-convergence-type error bounds for negative backward transfer. For two tasks $T^{(s)}$ (source) and $T^{(t)}$ (target), the increase in risk after backward transfer is bounded by

$$Er^{P^{(s)}}(h_{\mathrm{withBT}}) - Er^{P^{(s)}}(h_{\mathrm{noBT}}) \leq 2(\epsilon_1 + \epsilon_2)$$

where the $\epsilon_i$ terms depend on VC-dimensions, sample sizes, and hypothesis class modularity. NBT is provably controlled by the size of these slack terms. The bounds suggest that ensuring adequate relatedness, sufficient task-specific data, and modular model designs mitigates the upper bound on NBT.

6. Experimental Benchmarks and Empirical Evidence

Across domains, the prevalence and mitigation of NBT have been examined:

• Quantum Variational Classifier (QVC): GEM eliminates NBT (BWT $=+0.0372$ vs. $-0.0268$ for EWC) and enhances robustness on sequential quantum classification tasks (Situ et al., 2022).
• Computer Vision CL: On Split CIFAR-100 and Split Mini-ImageNet, CUBER achieves positive or near-zero BWT (e.g., $+0.13\%$), outperforming prior constrained baselines (Lin et al., 2022).
• MTL with Imbalanced Data: U-AMTFL reduces the number and severity of tasks suffering NBT compared to conventional MTL and STL baselines (Silva et al., 2020).
• Prompt Tuning for LLMs: SHLPT yields BWT $=0$ (no catastrophic forgetting) and is the only method with consistent positive FWT on standard, large, and adversarial task sequences (Wu et al., 2024).

7. Implications, Limitations, and Open Questions

Avoiding NBT is fundamental for scalable, robust continual learning. Key implications include:

• Strictly guarding against any parameter drift (orthogonal projection) can prevent forgetting but eliminates potential for beneficial backward transfer; relaxing these constraints, when gradients are aligned, is essential for positive transfer.
• Theoretical and empirical analyses indicate that NBT is minimized when tasks are sufficiently related, sample sizes are adequate, and model architectures exhibit suitable modularity. Poor task similarity and limited data raise NBT risk.
• Current techniques for NBT mitigation are often sensitive to hyperparameters (e.g., similarity thresholds in SHLPT) and lack universal guarantees across arbitrary task sequences or domains.
• In MTL, existing uncertainty-driven and asymmetric feature learning approaches reduce but do not always eliminate NBT, especially in extreme data regimes or under epistemic uncertainty.
• Open questions include formalizing the relationship between learned similarity metrics and the reduction of NBT, integrating epistemic uncertainty, and further narrowing the theory-practice gap for deep continual architectures with large-scale, heterogeneous tasks.

Negative backward transfer remains an active focus across continual, multi-task, and parameter-efficient lifelong learning. Its rigorous quantification, theoretical analysis, and mitigation are critical for the next generation of adaptive, robust AI systems (Situ et al., 2022, Silva et al., 2020, Benavides-Prado et al., 2022, Lin et al., 2022, Wu et al., 2024).