--- title: 'CoFilter Method: A Multi-Domain Filter Approach' url: https://www.emergentmind.com/topics/cofilter-method type: topic --- # CoFilter Method: A Multi-Domain Filter Approach Searching arXiv for papers associated with “CoFilter Method” and closely related usages. Search query: "CoFilter arXiv" The expression **“CoFilter Method”** is polysemous in the arXiv literature. In its most specific usage, **CoFilter** denotes **conditional testing after filtering** for multiple testing of partial conjunction hypotheses, introduced as a two-step procedure that removes conservative partial-conjunction $p$-values before multiplicity correction [2110.06692]. In other contexts, the label is also used for the **Co-occurrence Filter (CoF)**, a boundary-preserving image filter [1703.04111]; as an informal shorthand for **CostFilter-AD**, a cost-volume filtering plug-in for unsupervised anomaly detection [2505.01476]; and as an informal shorthand for **CoLoR-Filter**, a data-selection method for targeted language-model pre-training [2406.10670]. A later neuroimaging paper applies the partial-conjunction CoFilter framework to activation localization across subjects or tasks [2508.03675]. This suggests a family resemblance rather than a single method: each variant inserts an explicit filtering stage before a downstream decision rule. ## 1. Terminological scope and disambiguation The term is best understood by distinguishing canonical names from secondary usages. | Label in practice | Canonical name in paper | Domain | |---|---|---| | CoFilter | CoFilter (conditional testing after filtering) | Partial conjunction multiple testing | | CoFilter Method | Co-occurrence Filter (CoF) | Boundary-preserving image filtering | | CoFilter | CostFilter-AD | Unsupervised anomaly detection | | CoFilter | CoLoR-Filter | Targeted LM pre-training | The statistically oriented **CoFilter** is explicitly named in "A procedure for multiple testing of partial conjunction hypotheses based on a hazard rate inequality" [2110.06692]. The image-processing paper "Co-occurrence Filter" states that CoF is also referred to as the **“CoFilter Method”** [1703.04111]. The anomaly-detection paper states that **CostFilter-AD** is abbreviated there as **“CoFilter” for cost filtering**, while also noting that the official terminology used in the paper is **“CostFilter-AD”** [2505.01476]. The language-model paper states that if one encounters **“CoFilter,”** it is an informal shorthand for **CoLoR-Filter**, but that the authors’ canonical name and framing are **CoLoR-Filter** [2406.10670]. A recurring misconception is therefore terminological: the same label does not identify a single established algorithm across fields. In statistical multiple testing, CoFilter is a formal method name; in the other cases, it is either an alias or a shorthand. ## 2. Co-occurrence Filter in image processing In image processing, the **Co-occurrence Filter (CoF)** is a **boundary-preserving, texture-aware image filter** inspired by the bilateral filter (BF), but it replaces BF’s range Gaussian with a learned co-occurrence weighting that reflects which intensity or color pairs tend to appear together in the training data [1703.04111]. Its motivation is that BF preserves any large intensity jump, whereas CoF distinguishes **edges** from **boundaries** by using local co-occurrence statistics: pixel values that frequently co-occur inside a texture are smoothed together even when their intensities differ substantially, while pixel pairs that rarely co-occur across texture boundaries are not averaged. The bilateral filter is written as $$ I'(p) = \frac{1}{Z_p} \sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, w_r\big(I(p)-I(q)\big)\, I(q), $$ with $$ w_s(\|p-q\|) = \exp\!\left(-\frac{\|p-q\|^2}{2\sigma_s^2}\right),\quad w_r(\Delta I) = \exp\!\left(-\frac{\|\Delta I\|^2}{2\sigma_r^2}\right), $$ and $$ Z_p = \sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, w_r\big(I(p)-I(q)\big). $$ CoF replaces $w_r$ with a co-occurrence weight $w_c$: $$ I'(p) = \frac{1}{Z_p} \sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, w_c\big(I(p), I(q)\big)\, I(q), $$ where $$ Z_p = \sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, w_c\big(I(p), I(q)\big). $$ For grayscale images, intensities are quantized to $K$ bins, and the spatially weighted co-occurrence counts are $$ C(i,j) = \sum_{p} \sum_{q \in \mathcal{N}(p)} K_s\big(\|p-q\|\big)\, \mathbf{1}[Q(I(p))=i] \, \mathbf{1}[Q(I(q))=j]. $$ With marginal counts $$ h(i) = \sum_{p} \mathbf{1}[Q(I(p))=i], $$ the paper uses the symmetric normalization $$ M(i,j) = \frac{C(i,j)}{\big(h(i)+\epsilon\big)\big(h(j)+\epsilon\big)},\quad w_c(i,j) = M(i,j). $$ The paper also lists conditional and symmetrized-conditional alternatives. The limiting cases are instructive: as the spatial kernel bandwidth $\sigma \to 0$, CoF becomes identity; as $\sigma \to \infty$, $M(i,j) \to 1$ and CoF reduces to Gaussian smoothing in the image plane. For color images, direct RGB co-occurrence is infeasible, so the method clusters colors, typically in CIELab, using $k$-means. Let $$ T(p) \in \{1,\dots,K\},\quad \{\tau_\ell\}_{\ell=1}^K \in \mathbb{R}^3. $$ Hard cluster-space co-occurrence is $$ C_{\text{hard}(\tau_a,\tau_b) = \sum_{p} \sum_{q \in \mathcal{N}(p)} K_s(\|p-q\|)\, \mathbf{1}[T(p)=a]\, \mathbf{1}[T(q)=b], $$ and the soft approximation is $$ C_{\text{soft}(\tau_a,\tau_b) \approx \sum_{k_1=1}^K \sum_{k_2=1}^K K(\tau_a,\tau_{k_1})\, K(\tau_b,\tau_{k_2})\, C_{\text{hard}(\tau_{k_1},\tau_{k_2})}. $$ The resulting guided filter is $$ I'(p) = \frac{\sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, M_T\big(T(p),T(q)\big)\, I(q)}{\sum_{q \in \mathcal{N}(p)} w_s(\|p-q\|)\, M_T\big(T(p),T(q)\big)}. $$ The reported practical profile is local rather than non-local: hard co-occurrence collection is $O(nr^2)$, filtering is $O(nr^2)$, and the soft approximation adds $O(K^4)$, with typical settings **Lab color space**, **$K=32$ clusters**, and **window $15\times 15$**. The paper also distinguishes **Iterative CoF (I-CoF)**, which fixes $M_T$ across passes and converges quickly while preserving boundaries, from **Rolling CoF (R-CoF)**, which recomputes $M_T$ after each pass and yields stronger homogenization at the cost of possible boundary degradation. Relative to BF, Guided Filter, Domain Transform, L0 smoothing, Rolling Guided Filter, semantic filtering, WLS with diffusion distances, and Non-Local Means, CoF is positioned as a method that smooths repeated textures without sacrificing texture boundaries. ## 3. CoFilter for partial conjunction multiple testing In statistics, **CoFilter** is a two-step procedure for multiple testing of **partial conjunction (PC) hypotheses**, designed to reduce the conservativeness of standard PC $p$-value testing [2110.06692]. For $m$ studies and target replicability level $r$, the PC null hypothesis is $$ H_0^{r/m}:\ \text{at most } r-1 \text{ of the } m \text{ studies have a true signal}. $$ If the ordered study-level $p$-values are $$ p_{(1)} \le \cdots \le p_{(m)}, $$ the Fisher-based PC statistic combines the $m-r+1$ largest $p$-values: $$ T_{PC}(r,m) = -2 \sum_{i=r}^{m} \log p_{(i)}, $$ with $$ p_{PC} = \Pr\big( \chi^2_{2(m-r+1)} \ge T_{PC}(r,m) \big). $$ The method’s motivation is that these PC $p$-values are often conservative under the composite null, so many very large $p_{PC}$ contribute to multiplicity but not to discoveries. CoFilter therefore applies a filter threshold $\tau \in (0,1]$: $$ S_\tau = \{ j : p^{r/m}_j \le \tau \}. $$ For selected features, it forms **conditional PC $p$-values** $$ \tilde{p}^{r/m}_j = \frac{1}{\tau} p^{r/m}_j, \qquad j \in S_\tau, $$ and then applies a multiple testing procedure such as Bonferroni, Holm, BH, Storey, or BY on the filtered family. The theoretical contribution is the validity of these conditional PC $p$-values when **the null $p$-values are uniform and Fisher’s combining method is used**. The proof is based on a **novel inequality in hazard rate order of partial sums of order statistics**. The paper states the theorem as follows: if $X_1,\dots,X_\ell$ and $Z_1,\dots,Z_\ell$ are i.i.d. $\mathrm{Exp}(1)$, and for some $n>\ell$, $X_{\ell+1},\dots,X_n$ are any positive random variables, independent of $X_1,\dots,X_\ell$, then $$ X_{(1):n} + \cdots + X_{(\ell):n} \le_{hr} Z_1 + \cdots + Z_\ell. $$ From this, the paper derives the corollary that $$ \left[ \frac{1}{\tau} p_{PC} \,\middle|\, p_{PC} \le \tau \right] $$ is a valid conditional $p$-value for $H_0^{r/m}$. The method includes a data-adaptive threshold. For BH at level $\alpha$, the threshold at a given $\tau$ is $$ \hat{t}(\tau) = \max\{ t \in [0,1] : |S_\tau| \cdot t \le \alpha \cdot |S_{\tau t}| \}, $$ and one may choose $$ \hat{\tau} \in \arg\max_{\tau \in \{\tau_1,\dots,\tau_K\}} |S_{\tau \hat{t}(\tau)}|. $$ Under PRDS-type conditions, finite-sample BH control is available; under weak dependence, asymptotic FDP control is given for fixed or random $\tau$. The reported application is to **multiple GWAS of Crohn’s disease** with **8 independent GWAS studies** and about **953,241 autosomal SNPs per study**. Using Fisher PC $p$-values for $r \in \{2,3,4\}$, CoFilter with fixed $\tau=0.1$ or greedy $\tau$ yielded substantially more discoveries than unfiltered BH or adaptive BH on PC $p$-values. The paper reports, for $\tau=0.1$, **$|S_\tau|=30{,}385$** for $r=2$, **$7{,}328$** for $r=3$, and **$1{,}973$** for $r=4$. ## 4. Neuroimaging adaptation of CoFilter "A New Approach to Partial Conjunction Analysis in Neuroimaging" applies the recently proposed CoFilter method to voxelwise activation localization across subjects or tasks [2508.03675]. For subject $i \in [s]$ and voxel $j \in [m]$, let $p_{ji}$ be the voxelwise $p$-value for testing inactivity. The number of truly active subjects at voxel $j$ is $$ \delta_j = \big|\{\,i \in [s] : H_{ji}\ \text{is false}\,\}\big|, $$ and the $\gamma$-level PC hypothesis is $$ H_{0,j}^{\mathrm{PC}(\gamma)}: \ \delta_j < \gamma \quad \text{vs} \quad H_{1,j}^{\mathrm{PC}(\gamma)}: \ \delta_j \ge \gamma. $$ If $$ p_{j(1)} \le p_{j(2)} \le \cdots \le p_{j(s)} $$ are the ordered subject-level $p$-values at voxel $j$, the paper uses Fisher’s combination on the largest $s-\gamma+1$ ordered $p$-values: $$ p^{\mathrm{PC}_j(\gamma)}=1-F_{\chi^2_{2(s-\gamma+1)}}\!\left(-2\sum_{i=\gamma}^s \log p_{j(i)}\right). $$ For a threshold $\tau$, the filtered set is $$ S_\tau(\gamma)=\big\{j\in[m] \,:\, p^{\mathrm{PC}_j(\gamma)}\le\tau\big\}, $$ and the conditional PC $p$-values are $$ \widetilde{p}^{\mathrm{PC}_{j,\tau}(\gamma)}=\frac{p^{\mathrm{PC}_j(\gamma)}}{\tau}, \qquad j\in S_\tau(\gamma). $$ BH is then applied at level $\alpha$ on the filtered family. The procedure is repeated for all $\gamma=1,\dots,s$, and the voxelwise replicability lower bound is $$ d_j=\max\{\gamma\in [s]\;:\; H_{0,j}^{\mathrm{PC}(\gamma)} \ \text{is rejected}\}. $$ The overall error criterion is $$ \mathrm{FDP}=\frac{|\{j\,:\,d_j>\delta_j\}|}{|\{j\,:\,d_j>0\}|}, \qquad \mathrm{FDR}=\mathbb{E}(\mathrm{FDP}). $$ The paper also studies an adaptive or **“greedy”** threshold over a grid $\mathcal{T}$, choosing the $\tau$ that maximizes the number of rejections. It states that Dickhaus et al. (2024) show FDR control for fixed $\gamma$ for any choice of $\tau$, including adaptive choices, when using Fisher’s combination and the conditional scaling $p^{PC}/\tau$. For varying $\gamma$, the paper reports strong empirical evidence of overall FDR control, while noting that a full formal proof is not yet provided. The empirical evaluation includes two simulation regimes and a real dataset. In **Simulation 1**, with **$s=10$ subjects**, **$m=1000$ voxels**, **$n=50$ time points**, and equi-correlation $\rho \in \{0,0.3,0.6,0.9\}$, the reported FDR values over **500 repetitions** are **0.0262, 0.0256, 0.0234, 0.0191** for CoFilter, compared with **0.0403, 0.0408, 0.0377, 0.0298** for Benjamini–Heller (2009) and **0.0207, 0.0197, 0.0174, 0.0147** for AdaFilter. In **Simulation 2**, with **$s=8$ subjects**, **$m=1000$ voxels** in a **$10\times10\times10$** array and **SNR $\in \{1,1.5,2,2.5\}$**, the reported FDR values are **0.0478, 0.0475, 0.0473, 0.0470** for CoFilter, **0.0477, 0.0474, 0.0472, 0.0469** for Benjamini–Heller (2009), and **0.1066, 0.1152, 0.1142, 0.1141** for AdaFilter. The real-data study uses the **Smeets et al. (2013)** food-cue fMRI dataset from **OpenNeuro ds000157**, comprising **30 healthy women**. The paper reports that CoFilter discovers more voxels at very high $\gamma$ than Benjamini–Heller, makes on average **0.56** more rejections on the full voxel set, differs from Benjamini–Heller in about **18.7\%** of voxels, and in about **70\%** of those differing voxels CoFilter finds more activation, with about **3** more rejections on average in that subset. Reported Talairach coordinates include **$(-29,-78,-11)$** and **$(25,-80,-9)$** in visual cortex. ## 5. CostFilter-AD as a cost-filtering “CoFilter” method In unsupervised anomaly detection, **CostFilter-AD** introduces explicit **matching cost volume filtering** into UAD and is abbreviated in the paper as **“CoFilter” for cost filtering** [2505.01476]. The method addresses a failure mode shared by reconstruction-based and embedding-based UAD systems: anomaly maps are typically derived from image-level or feature-level matching, but that matching can be noisy and inaccurate due to imperfect reconstructions, misalignment across views or scales, and the absence of ideal normal templates. The method constructs a dense matching cost volume from frozen multi-layer features of an input image $I$ and normal templates $\{I_{\mathcal T}\}$. Using a frozen encoder such as **DINO ViT-B/8** or **EfficientNet-B4**, the paper extracts $$ f_{\mathcal S} \in \mathbb{R}^{L\times C\times H'\times W'} \quad\text{and}\quad f_{\mathcal T} \in \mathbb{R}^{N\times L\times C\times H'\times W'}. $$ Global cosine matching is then defined by $$ \mathcal{V}(j,n,l,i) = \frac{f_{\mathcal S}^{i,l} \cdot f_{\mathcal T}^{n,j,l}} {\|f_{\mathcal S}^{i,l}\| \cdot \|f_{\mathcal T}^{n,j,l}\|}, $$ and converted to anomaly cost by $$ \mathcal{C}(j,n,l,i)=1-\mathcal{V}(j,n,l,i). $$ After merging template and spatial candidate indices, the cost volume is reshaped to $$ \mathcal{C} \in \mathbb{R}^{(DN)\times L\times H'\times W'}. $$ A coarse anomaly map is obtained by global min pooling along the matching dimension: $$ \bar{\mathcal{M}}^l(x)=\min_d \mathcal{C}^l(x,d). $$ Filtering is performed by a **3D U-Net** with dual-stream guidance. **Spatial guidance (SG)** uses multi-layer input features $f_{\mathcal S}$ to preserve edges and contours, while **matching guidance (MG)** uses the coarse anomaly map $\bar{\mathcal M}$ to bias the network toward channels likely to contain anomalies. The core **Residual Channel-Spatial Attention (RCSA)** module concatenates the cost feature with projected guidance, $$ x_l'=\mathrm{cat}(x_l, h(\bar{\mathcal M}), h(f_{\mathcal S}^l)), $$ then applies residual channel attention $$ x_l^{ca} = \sigma( \mathrm{conv}( \mathrm{MP}(x_l') ) + \mathrm{conv}( \mathrm{AP}(x_l') ) ) \odot x_l' + x_l', $$ and residual spatial attention $$ x_l^{sa} = \sigma( \mathrm{conv}( \mathrm{cat}( \mu(x_l^{ca}), \max(x_l^{ca}) ) ) ) \odot x_l^{ca} + x_l^{ca}. $$ After filtering, the final anomaly map is $$ \mathcal{M}=\mathrm{softmax}(\mathrm{conv}(\min_d x)), $$ and the image-level anomaly score is the average of the **top-$K$ pixels with $K=250$**. The paper emphasizes that CostFilter-AD is a **generic post-processing plug-in** that can be attached to both reconstruction-based and embedding-based pipelines. In reconstruction-based settings, templates can include intermediate denoising steps $$ I_{t\to0} = \frac{1}{\sqrt{\bar{\alpha}_t}} \left[ I_t - \sqrt{1-\bar{\alpha}_t}\cdot \epsilon_\theta(I_t,t) \right], $$ where $\epsilon_\theta$ is the frozen diffusion noise predictor. In embedding-based settings, a small number $N$ of normal images per category is sampled instead of using large memory banks. The filtering network is trained on synthetic anomalies under the **GLAD protocol** with $$ \mathcal{L} = \mathcal{L}_{Focal}(\mathcal{M}, \mathcal{M}_s, \sigma(\hat{Y}_c)) + \mathcal{L}_{CE}(\hat{Y}_c, Y) + \alpha \big[ \mathcal{L}_{Soft-Iou}(\mathcal{M}, \mathcal{M}_s) + \mathcal{L}_{SSIM}(\mathcal{M}, \mathcal{M}_s) \big], $$ using **$\alpha=0.1$** and **$\gamma_0=3$** in the class-aware focal adaptor. The implementation details reported are **$L=4$** feature layers, **$N=3$** for GLAD and AnomalDF, **$N=1$** for HVQ-Trans, **input sizes $256\times256$** for GLAD and AnomalDF and **$224\times224$** for HVQ-Trans, and **Adam**, **lr $=1\mathrm{e}{-3}$**, **batch size 8**, **40 epochs**, with **ReduceLROnPlateau**. On **A100 40GB**, the added runtime and memory are reported as **+0.37 s and +2.07 GB** for GLAD, **+0.07 s and +0.94 GB** for HVQ-Trans, and **+0.32 s and +0.82 GB** for AnomalDF. The reported gains are benchmark-specific but consistently positive. On **MVTec-AD** multi-class UAD, mean image/pixel AUROC improves from **97.5/97.3 to 98.7/98.2** for GLAD, from **98.0/97.3 to 99.0/98.0** for HVQ-Trans, and from **96.8/98.1 to 98.5/98.8** for AnomalDF. On **VisA**, the corresponding changes are **90.1/97.4 to 93.2/98.1**, **91.3/98.5 to 93.4/98.6**, and **90.5/97.5 to 94.3/99.2**. The paper also reports improvements for a unified single-class model and ablations showing that mapping $DN$ to channels is superior to mapping $DN$ to depth, that multiple denoised templates outperform the final-step template alone, and that dual-stream attention plus the full loss gives the best results. ## 6. CoLoR-Filter as an informal “CoFilter” in language-model pre-training In language-model pre-training, **CoLoR-Filter** stands for **Conditional Loss Reduction Filtering**. The paper states that if one encounters **“CoFilter,”** it is an informal shorthand for the same method, but that the canonical name is **CoLoR-Filter** because the method is defined by the **conditional reduction in loss** of a target-domain model relative to a background model [2406.10670]. The setting is targeted data selection. Let $D_{\text{train}}$ be a large corpus, $D_{\text{down}}$ a small sample from the downstream target distribution, and $D_{\text{prior}}$ a prior dataset, typically sampled from the same distribution as $D_{\text{train}}$. For an autoregressive language model, $$ \ell_\theta(x) = -\log p_\theta(x) = -\sum_{t=1}^T \log p_\theta(x_t \mid x_{