---
title: 'H-SPLID: HSIC-based Latent Info Decomposition'
url: https://www.emergentmind.com/topics/h-splid
type: topic
---

# H-SPLID: HSIC-based Latent Info Decomposition

H-SPLID, short for **HSIC-based Saliency Preserving Latent Information Decomposition**, is a supervised representation-learning method that explicitly partitions a latent representation into a **salient subspace** and a **non-salient subspace**, with classification performed from the salient component only [2510.20627]. Its central objective is to make a model rely on task-relevant information rather than background, context, or other nuisance structure. The method combines latent masking, geometric regularization, and HSIC-based dependence penalties to promote a low-dimensional predictive channel, and it is motivated by the observation that standard neural networks can achieve high clean accuracy while still encoding a mixture of relevant and irrelevant factors in one entangled latent space [2510.20627].

## 1. Problem setting and conceptual basis

H-SPLID targets classification problems in which input signals contain both label-relevant and label-irrelevant variation. The motivating claim is that ordinary training often yields representations that are predictive but not genuinely salient: a classifier may succeed on clean data while still depending on backgrounds, co-occurring context, or redundant visual structure [2510.20627]. In the paper’s diagnostic example, an image contains two MNIST digits side by side while the label is the **left** digit only; a vanilla classifier retains high clean accuracy yet suffers a large accuracy drop when the **right** digit is adversarially perturbed, which is taken as evidence that the learned representation is not restricted to the intended signal [2510.20627].

Within this framework, “saliency preserving” does not denote a post hoc attribution map. It denotes a representational constraint: the latent space is organized so that the information salient for predicting \(Y\) is preserved in one subspace, while remaining variability is pushed into another [2510.20627]. The method is therefore positioned simultaneously as a robustness method, an interpretability method, and a task-relevant representation-learning method.

The paper states that H-SPLID does not require external saliency annotations or a separate background dataset [2510.20627]. This places it in a regime distinct from methods that depend on pixel-level masks, environment labels, or explicitly curated nuisance datasets.

## 2. Latent decomposition and objective function

The model considers \(k\)-class classification on a dataset
\[
\mathcal{D}=\{(x_i,y_i)\}_{i=1}^n \subseteq \mathcal{X}\times\mathcal{Y},
\]
with \(\mathcal{X}\subseteq \mathbb{R}^d\) and \(\mathcal{Y}\subseteq \mathbb{R}^k\), where each \(y_i\) is one-hot encoded [2510.20627]. The network consists of an encoder
\[
f_\psi:\mathcal{X}\to \mathbb{R}^m
\]
and a linear output layer
\[
g_W(z)=Wz,\qquad W\in\mathbb{R}^{k\times m},
\]
so that the full predictor is \(h_\theta=g_W\circ f_\psi\) with \(\theta=\{\psi,W\}\) [2510.20627].

The central mechanism is a diagonal mask
\[
M_s=\operatorname{diag}(b),\qquad b\in\{0,1\}^m,
\]
with complement
\[
M_n=I-M_s.
\]
Given \(z=f_\psi(x)\), H-SPLID defines
\[
z_s=M_s z=b\odot z,\qquad z_n=M_n z=(\mathbf{1}-b)\odot z.
\]
Only \(z_s\) is fed to the classifier [2510.20627]. The non-salient component is not discarded; it is explicitly used in the training objective so that it absorbs variation that should not drive prediction.

Two centroid-based geometric terms organize the latent space. With class index sets \(C_k\), class centroids \(\mu_k\), and global centroid \(\mu\), the paper uses
\[
\mathcal{L}_s(\mathcal{D};\theta,M_s)=\sum_{k=1}^K\sum_{i\in C_k}\|M_s(z_i-\mu_k)\|^2,
\]
and
\[
\mathcal{L}_n(\mathcal{D};\theta,M_n)=\sum_{i=1}^n \|M_n(z_i-\mu)\|^2.
\]
The intended interpretation is that \(\mathcal{L}_s\) promotes within-class clustering in the salient space, while \(\mathcal{L}_n\) promotes global alignment in the non-salient space [2510.20627].

The information-theoretic component is expressed through two HSIC penalties: one between the input and the salient representation, \(\operatorname{HSIC}(X,Z_s)\), and one between the labels and the non-salient representation, \(\operatorname{HSIC}(Y,Z_n)\) [2510.20627]. The first is used to reduce redundant dependence of \(Z_s\) on the full input; the second is used to discourage label information from leaking into \(Z_n\).

The full objective is
\[
\mathcal{L}(\mathcal{D};\theta,M_s,M_n)
=
\lambda_{\mathrm{ce}}\mathcal{L}_{\mathrm{ce}}
+\lambda_s\mathcal{L}_s
+\lambda_n\mathcal{L}_n
+\rho_s\,\operatorname{HSIC}(X,Z_s)
+\rho_n\,\operatorname{HSIC}(Y,Z_n),
\]
where \(\mathcal{L}_{\mathrm{ce}}\) is a masked cross-entropy loss using only the salient coordinates [2510.20627]. The paper characterizes the combined effect as enforcing prediction from \(z_s\) only, reducing redundancy within \(z_s\), and removing class-relevant signal from \(z_n\).

## 3. Optimization procedure and mask mechanics

H-SPLID is trained by alternating optimization over network parameters and the salient mask [2510.20627]. With the mask fixed, the network parameters \(\theta=\{\psi,W\}\) are updated for one epoch using minibatch SGD on the full objective. For each minibatch, the algorithm computes latent codes, minibatch class means, the global mean, the H-SPLID loss, and then applies a gradient step [2510.20627].

With \(\theta\) fixed, the mask is updated by minimizing the geometric part of the objective,
\[
\lambda_s \mathcal{L}_s(Z^{(t)},M_s)+\lambda_n \mathcal{L}_n(Z^{(t)},M_n),
\]
subject to \(M_s=\operatorname{diag}(b)\) and \(M_n=I-M_s\) [2510.20627]. The paper states that this yields a closed-form coordinate update for \(b_i\), interpreted as follows: a latent dimension that varies strongly at the global level but weakly within classes is assigned higher saliency weight, whereas a dimension that mainly reflects global or shared variation is pushed toward the non-salient side [2510.20627]. The optimization produces a continuous mask \(b\in[0,1]^m\), which is thresholded at \(0.5\) at the end, although the reported behavior is that many coordinates already lie near \(0\) or \(1\) [2510.20627].

To stabilize training, the paper uses a moving-average update for the mask and initializes the procedure with \(M_s^{(0)}=I\) [2510.20627]. For ImageNet-scale experiments, the implementation introduces \(\beta_{\text{init\_fraction}}\) and \(\beta_{\text{update\_fraction}}\), which specify the fraction of the data used for mask initialization and the fraction processed before each mask update [2510.20627].

This alternating scheme induces moderate overhead. On COCO, the reported average epoch times are \(37.1\) s for Vanilla, \(45.7\) s for HBaR, and \(42.9\) s for H-SPLID [2510.20627]. On ImageNet-1K, total runtimes for 20 epochs are reported as 11h27m for Vanilla, 12h04m for HBaR, and 14h06m for H-SPLID [2510.20627]. The paper attributes the overhead to HSIC computation and mask updates.

## 4. Theoretical properties

The main theoretical claim is a robustness bound linking perturbation sensitivity to both the **dimension of the salient subspace** and the **HSIC between inputs and salient representations** [2510.20627]. Let
\[
z_s=M_s f_\psi(x),\qquad s=\|M_s\|_0,
\]
where \(s\) is the number of active salient dimensions. Under the assumptions that \(x\) is sampled from a truncated multivariate normal, the encoder \(f_\psi\) is \(L\)-Lipschitz, the output layer is linear and bounded, and the kernels used for HSIC are continuous, positive definite, universal on compact domains, and uniformly bounded in sup norm, the paper proves that the expected prediction deviation under bounded perturbations is upper-bounded by a term that scales with \(\sqrt{s}\) and \(\operatorname{HSIC}(x,z_s)\) [2510.20627].

The conceptual interpretation is explicit: reducing the dimensionality of the predictive channel and reducing the amount of input information preserved in that channel both improve robustness [2510.20627]. The paper presents this as a robustness–compression link rather than as a purely empirical regularization effect.

A corollary bounds the probability of encountering an input at which the Frobenius norm of the input gradient exceeds a threshold. In the paper’s formulation, the measure of the “salient-active” region is controlled by the same structural quantities: salient-space dimension and \(\operatorname{HSIC}(x,z_s)\) [2510.20627]. The proof sketch proceeds by bounding network outputs using salient sparsity, relating bounded neural functions to RKHS function classes, using HSIC to control covariances, applying a Stein-type identity for truncated Gaussian inputs, and then converting the gradient bound into a perturbation bound through a first-order Taylor argument [2510.20627].

The theory is not presented as a general theorem for arbitrary data distributions. The assumptions are specific, and the paper treats them as the analytic setting under which the robustness–compression relationship can be formalized [2510.20627].

## 5. Empirical results

The empirical evaluation covers five datasets: C-MNIST, a four-class COCO subset, ISIC-2017, ImageNet-9, and CounterAnimal [2510.20627]. Baselines include Vanilla cross-entropy, weight decay, \(L_1\) sparse weights, \(L_1\) sparse activations, group-lasso weights, group-lasso activations, and HBaR [2510.20627]. All methods are trained on clean data only, with no adversarial training and no saliency masks used during training [2510.20627].

On the diagnostic C-MNIST task, the paper reports that Vanilla achieves clean accuracy \(96.86\%\) but only \(31.76\%\) under strong PGD applied to the irrelevant right digit, whereas H-SPLID achieves clean accuracy \(97.14\%\) and \(87.46\%\) under the same attack [2510.20627]. The non-salient space alone attains only \(8.57\%\) classification accuracy, which the paper interprets as evidence that class information has largely been removed from that channel [2510.20627].

On COCO, the most direct evaluation concerns attacks on backgrounds. Under background-only PGD at \(\epsilon=3/255\), the reported accuracies are \(34.4\%\) for Vanilla, \(43.0\%\) for weight decay, \(39.9\%\) for HBaR, and \(57.5\%\) for H-SPLID [2510.20627]. Under full-image PGD at \(\epsilon=2/255\), Vanilla attains \(34.2\%\), HBaR \(41.9\%\), and H-SPLID \(58.9\%\) [2510.20627]. Under background-only AutoAttack at \(\epsilon=3/255\), Vanilla is \(26.4\%\), HBaR \(31.4\%\), and H-SPLID \(52.6\%\); under full-image AutoAttack at \(\epsilon=2/255\), Vanilla is \(20.9\%\), HBaR \(25.2\%\), and H-SPLID \(48.8\%\) [2510.20627]. These results are presented as evidence that H-SPLID reduces dependence on non-salient image regions.

On ISIC-2017, where perturbations affect non-lesion pixels, H-SPLID again yields the best reported performance: \(76.78\pm0.78\) without perturbation, \(70.00\pm1.62\) under brightness perturbation, \(68.38\pm1.38\) under defocus blur, and \(69.50\pm1.72\) under occlusion-like corruption; the corresponding HBaR values are \(75.90\pm0.84\), \(68.70\pm1.94\), \(65.62\pm2.06\), and \(66.18\pm3.01\) [2510.20627]. The paper treats this as evidence that the decomposition is not limited to synthetic backgrounds.

On ImageNet-9 and CounterAnimal, the reported values are \(76.7\) on IN-9 original, \(64.5\) on Only-FG, \(59.5\) on MixedRand, \(80.3\) on CA-Common, and \(60.6\) on CA-Counter for H-SPLID [2510.20627]. The gains on MixedRand and CA-Counter are emphasized because these settings expose reliance on spurious background context [2510.20627].

The ablation study on COCO separates the contributions of geometry and HSIC. The paper reports \(33.75\%\) background-attack and \(35.29\%\) full-attack accuracy for cross-entropy only, \(43.69\%\) and \(44.12\%\) for cross-entropy plus \(\mathcal{L}_s+\mathcal{L}_n\), \(42.71\%\) and \(45.87\%\) for cross-entropy plus HSIC only, and \(57.12\%\) and \(58.44\%\) for the full H-SPLID objective [2510.20627]. It therefore argues that neither ingredient alone accounts for the observed behavior.

The paper also reports evidence for low-dimensional saliency. On COCO, one of the best configurations yields a salient subspace of **14 out of 512** dimensions [2510.20627]. Increasing \(\lambda_n\) or \(\rho_n\) tends to shrink the salient space and improve robustness up to a point, while excessive regularization can collapse the model to trivial solutions [2510.20627].

## 6. Interpretation, limitations, and disambiguation

H-SPLID is best understood as a task-directed latent partitioning method rather than as a generic sparsification scheme. Compared with disentanglement methods, it is not trying to recover independent generative factors in an unsupervised sense; “salient” is defined operationally as useful for the label [2510.20627]. Compared with invariant learning, it seeks insensitivity to nuisance factors through explicit latent partitioning and HSIC regularization rather than through environment annotations or causal invariance assumptions [2510.20627]. Compared with information bottleneck-style approaches such as HBaR, it imposes distinct pressures on the two latent subspaces: it compresses input dependence in \(Z_s\), removes label information from \(Z_n\), and assigns geometric roles to both spaces [2510.20627]. Compared with \(L_1\) or group-lasso regularization, it is not merely making weights or activations small; it is attempting to determine which latent dimensions should carry discriminative information and which should not [2510.20627].

The method’s stated applicability is strongest when irrelevant information is present in the input and the dataset is diverse enough that salient and non-salient factors do not always co-occur [2510.20627]. The paper explicitly notes that if every class always appears in the same context, the model cannot disentangle object from background from labels alone; in that case additional supervision or prior knowledge would be needed [2510.20627]. Other caveats identified in the paper are the computational overhead, the image-centric experimental scope, the distributional and kernel assumptions required by the theory, and the possibility of mask collapse under poor hyperparameter settings [2510.20627].

The acronym also benefits from disambiguation. In current arXiv usage, **H-SPLID** most directly refers to **HSIC-based Saliency Preserving Latent Information Decomposition** [2510.20627]. Other supplied papers discuss unrelated methods that have been suggested as possible intended matches when the string is used ambiguously, including the **Hierarchical Hamiltonian Splitting** family for collisionless \(N\)-body simulation [2009.06133] and the HerMES prior-based SPIRE deblending and cross-identification framework associated with the **HerMES XID algorithm** [1009.1658]. Those methods are distinct in domain, formalism, and purpose.

Source: https://www.emergentmind.com/topics/h-splid