---
title: Change-Guided Attention (CGA)
url: https://www.emergentmind.com/topics/change-guided-attention-cga
type: topic
---

# Change-Guided Attention (CGA)

Change-Guided Attention (CGA) denotes attention allocation guided by estimated change signals rather than unguided feature aggregation. In bitemporal remote sensing change detection, CGA is used to steer feature weighting or gating so that changed regions receive higher weight while unchanged regions are suppressed; in semantic change detection it can explicitly connect binary change detection (BCD) and semantic change detection (SCD); and in partially observable agents it can be implemented as an endogenous priority mechanism driven by ignorance, surprise, and staleness rather than by external rewards [2508.08232][2603.09476][2308.14570][2302.10420]. Across these formulations, the common role of CGA is to bias computation toward locations, variables, or transitions that are more likely to have changed.

## 1. Conceptual scope and problem setting

CGA arises in settings where change is sparse, unevenly distributed, or expensive to identify. In semantic change detection from remote sensing imagery, the central questions are “where did change occur?” and “what changed from-to?”; this creates a coupled BCD–SCD problem under severe class imbalance, large spatial extent, and a need for both local detail and long-range context [2508.08232]. In very-high-resolution change detection, existing Siamese encoder–decoder systems were reported to suffer because deep encoder layers focus on irrelevant background regions and decoder confidence in the change regions is inconsistent at different stages [2308.14570]. In partially observable agents, the parallel problem is limited observation capacity: only a subset of variables can be inspected at each tick, so attention must be allocated toward variables most likely to have changed [2603.09476].

A concise way to organize the literature is to distinguish CGA by the source of its guidance signal. Some models use a learned change prior from a BCD pathway, some use bi-temporal similarity or dissimilarity in feature space, some use an explicit changing guide map inside self-attention, and some use internal epistemic-gap signals.

| Work | Guidance source | Mechanism |
|---|---|---|
| Mamba-FCS | stage-wise change maps \(CM_i\) from the BCD decoder | multiplicative guidance \( \hat{X}_i^{T_j} = X_i^{T_j} \odot \sigma(CM_i) \) |
| SAAN | bi-temporal similarity \(Sim_i\) and dissimilarity attention \(DSA_i\) | similarity-guided channel and spatial attention with attention flow |
| HCGMNet | changing guide map \(M\) | self-attention with changing guide map |
| Telogenesis | ignorance, surprise, and staleness | scalar priority function with softmax or top-\(b\) selection |

This spread of definitions also makes clear that CGA is not a single architectural primitive. It is a design principle in which estimated change cues are injected into attention, gating, or selection.

## 2. Stage-wise change priors in Mamba-FCS

In "Mamba-FCS: Joint Spatio- Frequency Feature Fusion, Change-Guided Attention, and SeK Loss for Enhanced Semantic Change Detection in Remote Sensing" [2508.08232], CGA is introduced to inject stage-wise change priors from the BCD decoder into both semantic decoders, ensuring the semantic heads attend to change-dominant regions while preserving the efficiency and global modeling benefits of the Mamba backbone. The framework uses a shared VMamba encoder that extracts four-stage features for each time point, a BCD decoder that produces a full-resolution change map \(Y_{BCD}\) and stage-wise change maps \(CM_i\), and two identical semantic decoders with weights not shared.

The encoder outputs are
$$
X_1^{T1}, X_2^{T1}, X_3^{T1}, X_4^{T1} = F_{encoder}(I_{T1}),
$$
$$
X_1^{T2}, X_2^{T2}, X_3^{T2}, X_4^{T2} = F_{encoder}(I_{T2}),
$$
where \(X_i^{T_j} \in \mathbb{R}^{C_i \times H_i \times W_i}\), with resolutions \(H_i \times W_i = (H/4 \times W/4, H/8 \times W/8, H/16 \times W/16, H/32 \times W/32)\) and channels \(C_i = (128, 256, 512, 1024)\) for VMamba-Base. The BCD decoder outputs
$$
Y_{BCD}, \{CM_i\}_{i=1\ldots4} = F_{BCD}(\{X_i^{T1}, X_i^{T2}\}_{i=1\ldots4}).
$$

The CGA module itself is the exact gating
$$
\hat{X}_i^{T_j} = X_i^{T_j} \odot \sigma(CM_i).
$$
No queries, keys, or values are defined. CGA is not a self-attention or cross-attention mechanism; it is a guided gating where \(\sigma(CM_i)\) acts as a spatial-channel mask. CGA is parameter-free: a sigmoid on \(CM_i\) followed by element-wise multiplication with \(X_i^{T_j}\). The semantic decoder then applies a VSS block and a CBAM-based upsampling unit, proceeding top-down from stage \(4 \rightarrow 3 \rightarrow 2 \rightarrow 1\), with residual additions from coarser scales.

A distinctive aspect of this formulation is that the change prior is not raw differencing alone. Upstream of CGA, the BCD decoder uses joint spatio-frequency fusion with log-amplitude FFT features and spatial difference maps:
$$
F_i^{T1} = \log(1 + |FFT2(X_i^{T1})|), \qquad
F_i^{T2} = \log(1 + |FFT2(X_i^{T2})|),
$$
$$
D_i = |X_i^{T1} - X_i^{T2}|.
$$
These are concatenated and refined by CBAM before being sent into the BCD decoder. As a result, CGA indirectly benefits from frequency cues that suppress illumination artifacts.

The loss design is also coupled to the CGA pathway. The total objective is
$$
L_{total} = L_{CE}^{Y_{BCD}} + 0.5(L_{CE}^{Y_{T1}} + L_{CE}^{Y_{T2}}) + \lambda_1 L_{mIoU} + \lambda_2 L_{SeK},
$$
with \(\lambda_1 = 0

Source: https://www.emergentmind.com/topics/change-guided-attention-cga