DualCP: Coarse-to-Fine Calibration

• The technique introduces a dual-stage alignment mechanism; first aligning features at a coarse group level and then refining alignment at a fine class level to optimize performance.
• It mitigates catastrophic forgetting in rehearsal-free domain-incremental learning by anchoring features to fixed, neural collapse-inspired prototypes through a dual-objective regression loss.
• Empirical results on benchmarks like DomainNet and CORe50 demonstrate significant accuracy gains and reduced forgetting compared to single-stage alignment methods.

Coarse-to-fine calibration in the context of DualCP refers to a dual-stage alignment strategy for feature representations, designed to address catastrophic forgetting in rehearsal-free domain-incremental learning (RFDIL). The core motivation is to adapt models to new domains while preserving knowledge from previous domains, without replaying prior data. This is accomplished by aligning features to fixed class prototypes using a two-level pipeline, first at a coarse (group) semantic granularity, then at a fine (class) level, with task-specific multilayer perceptrons (MLPs) and a dual-objective regression loss. The resulting mechanism underpins state-of-the-art performance on several RFDIL benchmarks, with extensions inspiring analogous calibration strategies in LLMs.

1. Architectural Framework and Dataflow

The Coarse-to-Fine Calibrator (C2F) constitutes the central architectural unit within DualCP's pipeline (Wang et al., 23 Mar 2025). Given a frozen image backbone and a precomputed set of neural collapse-inspired concept prototypes, C2F comprises:

• Coarse-grained Channel ($g_C$): An $L$-layer MLP mapping the raw image feature vector $x \in \mathbb{R}^d$ into a group-prototype space $x_C = g_C(x;\phi_C) \in \mathbb{R}^d$.
• Fine-grained Channels ($\{g_{F_k}\}$): A collection of group-specific MLPs, one per semantic group, each taking as input the concatenation $[x; x_C] \in \mathbb{R}^{2d}$ and outputting a fine-level embedding $x_F = g_{F_c}([x;x_C];\phi_{F_c}) \in \mathbb{R}^d$.

These MLPs are trained from scratch during each incremental domain step, while both the visual backbone and class/group prototypes (derived from Equiangular Tight Frames, ETFs) are fixed. The design is summarized in the following table:

Stage	Input	MLP Structure
Coarse: $g_C$	$x \in \mathbb{R}^d$	$d \rightarrow d$ (L=2 layers)
Fine: $g_{F_c}$	$[x;x_C] \in \mathbb{R}^{2d}$	$2d \rightarrow d$ (L′=2)

The two-tier mapping explicitly disambiguates group- and class-specific semantic granularity, ensuring hierarchical prototype alignment and improving feature geometry within the latent space.

2. Calibration Objectives and Loss Functions

The targets for C2F's outputs are prototype vectors $e_{C,c}$ (for group $c$) and $e_{F_c,j}$ (class $j$ in group $c$), each unit-normed and sampled from simplex configurations (ETFs). The calibration objective requires each feature mapping to be as closely aligned as possible with its corresponding prototype, using the inner product as the similarity metric. The explicit regression objectives are:

$L_C = (x_C^T e_{C,c} - 1)^2$

$L_F = (x_F^T e_{F_c,j} - 1)^2$

These are combined into a scalar Dual Dot-Regression (DDR) loss:

$$\mathcal{L}_{\mathrm{DDR}}(x_C,x_F; e_{C,c},e_{F_c,j}) = \alpha\,(x_C^T e_{C,c}-1)^2 + (1-\alpha)\,(x_F^T e_{F_c,j}-1)^2$$

where $\alpha \in [0,1]$ (empirically, $\alpha = 0.5$) balances coarse and fine prototype alignment.

3. Training Algorithm and Implementation Details

Training is performed with standard SGD (learning rate 0.1 with cosine decay, weight decay $2 \times 10^{-4}$, batch size 128, typically 20–30 epochs per domain). Prototypes are computed once per class and group, and the C2F MLPs are reinitialized for each new domain. The primary steps are:

1. Extract frozen backbone features.
2. Apply the coarse-grained MLP: $x_C = g_C(x; \phi_C)$.
3. Apply the fine-grained, group-specific MLP: $x_F = g_{F_c}([x; x_C]; \phi_{F_c})$.
4. Compute coarse and fine alignment objectives; aggregate via DDR loss.
5. Update $\phi_C$ and $\{\phi_{F_k}\}$ by SGD.

At inference, group and class assignment are performed by maximizing the inner product of mapped features with their respective prototypes.

Key architectural hyperparameters:

Parameter	Value (ViT-B/16 baseline)
Coarse MLP depth ($L$)	2 (hidden dim $d=768$)
Fine MLP depth ($L'$)	2 (2d $\rightarrow$ d)
$\alpha$ (DDR weight)	0.5
Prototype adjacency threshold	$p=0.85$ for semantic clustering

Ablation studies show that two-layer MLPs with hidden size $d$ yield the best tradeoff: one layer reduces $A_T$ by ≈1.4 points and increases forgetting ($F_T$) (Wang et al., 23 Mar 2025).

4. Empirical Evaluation and Quantitative Impact

DualCP with the Coarse-to-Fine Calibrator demonstrates substantial gains across domain-incremental learning benchmarks:

• DomainNet (345 classes, 6 domains):
  • $A_T = 60.13\%$, $F_T = -1.96\%$
  • VanillaCP (single-level alignment): $A_T = 56.22\%$, $F_T = -2.80\%$
• CORe50 (50 classes, 11 domains):
  • $A_T = 88.10\%$
  • VanillaCP: $A_T = 86.27\%$

These improvements—1.8–3.9% in final accuracy, 0.8–1.2% reduction in forgetting—are directly attributed to the two-stage calibration. Ablation confirms that combining coarse and fine alignment, rather than collapsing both into a single mapping, is essential for high performance and memory retention.

5. Extension to LLM Calibration

The coarse-to-fine calibration paradigm extends naturally to post-training calibration in LLMs, where Dual-Align (Luo et al., 7 Jan 2026) applies a two-stage (coarse and fine) confidence correction. In this setting:

• Coarse stage addresses confidence drift using final-distribution matching between pre-trained and post-trained models.
• Fine stage captures process drift by identifying the peak divergence layer and aligning post-divergence inferential stability via entropy objectives.

A unified loss with sample-specific weighting interpolates between these regimes. Key results on MMLU and MedMCQA show Dual-Align achieves 2.4–5.2% ECE, outperforming single-stage calibrators and approaching supervised temperature scaling performance, all with a single global $\tau$ parameter and zero inference overhead.

6. Theoretical and Practical Significance

The dual-stage calibration enforces an embedding geometry favorable for continual learning, mirroring aspects of neural collapse and prototype theory, and enables efficient, privacy-preserving incremental adaptation with no rehearsal data. The method is extremely lightweight, introducing only small MLP banks and single-stage regression targets, yet effectively mitigates the stability-plasticity dilemma ubiquitous in domain-incremental regimes.

The architectural separation of semantic granularities (group/class or process/confidence) and the explicit, rigorous use of prototype alignment loss make the coarse-to-fine calibrator in DualCP foundational for current and upcoming frameworks in both vision and language domains.

References:

• "DualCP: Rehearsal-Free Domain-Incremental Learning via Dual-Level Concept Prototype" (Wang et al., 23 Mar 2025)
• "Unlocking the Pre-Trained Model as a Dual-Alignment Calibrator for Post-Trained LLMs" (Luo et al., 7 Jan 2026)