DiPS: Diverse Methods Across Domains

• DiPS is a reused term that denotes diverse methods across seismic imaging, computer vision, reinforcement learning, and astrophysics.
• In seismic imaging, DiPS provides a finite-offset slope relation for validating velocity models, explaining reflector curvature, and estimating migration apertures.
• In machine learning and scientific computing, related frameworks convert unlabeled data into pseudo-supervision signals and enable differentiable PDE solving and protein interface prediction.

“DiPS” and “DIPS” are reused labels in several arXiv literatures rather than a single canonical concept. In the sources considered here, the term denotes a prestack-seismic dip relation, pseudo-supervision and pseudo-labeling frameworks in computer vision, policy-selection and sketching methods in machine learning, a protein-interface dataset and a differentiable PDE solver, and two distinct astrophysical dip phenomena (Singh, 2022, Al-Hindawi et al., 2023, Zhang et al., 2 Jul 2026, Ghosh et al., 2021, Morehead et al., 2021, Mistani et al., 2022, Easeman et al., 2022, Panizo-Espinar et al., 2023).

1. Terminological scope

Usage	Domain	Reference
DiPS as reflector-dip relations	Seismic imaging	(Singh, 2022)
Domain-knowledge Inspired Pseudo Supervision	UI2I and cross-domain classification	(Al-Hindawi et al., 2023)
Dialogue Policy Selection	High-stakes persuasion agents	(Zhang et al., 2 Jul 2026)
Differentiable Policy for Sketching	Sequential recommender systems	(Ghosh et al., 2021)
Discriminative Pseudo-Label Sampling	Weakly supervised object localization	(Murtaza et al., 2023)
DIPS / DIPS-Plus; JAX-DIPS	Protein interfaces; differentiable PDE solvers	(Morehead et al., 2021, Mistani et al., 2022)
Dip phenomena	Galaxy metallicity; black-hole transients	(Easeman et al., 2022, Panizo-Espinar et al., 2023)

The shared orthography masks strong domain divergence. In some papers, DiPS is a method name; in others, it is shorthand for a physical dip phenomenon or part of a software or dataset title. A related but distinct singular form, “DiP,” denotes “Discriminative implicit Parts” for person re-identification (Li et al., 2022).

2. DiPS in seismic imaging

In seismic imaging, DiPS concerns the relation between reflector dips on unmigrated constant-offset data and the true dip measured on prestack time-migrated constant-offset sections under 2D, constant-velocity, isotropic, homogeneous assumptions (Singh, 2022). The migrated dip $\theta$ and unmigrated dip $\alpha_u$ are expressed on time sections through

$$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$

The central prestack result is the finite-offset slope relation

$$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$

which extends the zero-offset post-stack relation of Chun and Jacewitz. Here $h$ is half-offset and $t_0$ is zero-offset two-way time.

This formulation has three stated uses. First, it provides a velocity-QC test: given $\delta\tau/\delta x$, $t_0$, $v$, and $h$, one predicts $\alpha_u$0 and compares it with the measured unmigrated slope. In the synthetic example with $\alpha_u$1 m, $\alpha_u$2 km/s, $\alpha_u$3 ms/m, $\alpha_u$4 ms/m, and $\alpha_u$5 s, the predicted unmigrated slope is $\alpha_u$6 ms/m, differing by about $\alpha_u$7. Second, the theory explains why a planar constant-dip reflector appears curved on constant-offset gathers: for fixed $\alpha_u$8 and $\alpha_u$9, the apparent unmigrated dip varies with the geometric parameter $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$0, so curvature increases with offset and decreases asymptotically with depth. Third, the paper derives a prestack migration-aperture estimate,

$$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$1

and studies its dependence on offset, depth, and dip angle. A plausible implication is that DiPS, in this usage, is less a named algorithm than a compact way of referring to dip behavior and dip transforms in prestack time migration.

3. Pseudo-supervision and pseudo-labeling frameworks

In computer vision and cross-domain learning, DIPS most explicitly denotes “Domain-knowledge Inspired Pseudo Supervision,” a model-selection framework for unsupervised image-to-image translation checkpoints in cross-domain classification (Al-Hindawi et al., 2023). The setting assumes a labeled source domain and an unlabeled target domain with the same semantic label space. A source classifier $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$2 is trained on labeled source images, a UI2I model $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$3 translates target images to source style, and the unresolved problem is checkpoint selection without target labels. DIPS addresses this by extracting InceptionV3 features from the untranslated target validation set, fitting a Gaussian Mixture Model with number of components equal to the known number of classes, and using the resulting cluster assignments as pseudo labels. Standard supervised metrics such as balanced accuracy and AUC are then computed between pseudo labels and classifier predictions on translated images. In the binary boiling-crisis case study, DIPS either selected the true best checkpoint or a very close one: for DS1 $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$4 DS2, the true best balanced accuracy was $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$5 and DIPS selected a checkpoint with $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$6, whereas FID selected one with $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$7; for AUC, the true best was $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$8 and DIPS selected $$\tan\theta=\frac{v\,\delta\tau}{2\,\delta x}, \qquad \tan\alpha_u=\frac{v\,\delta t}{2\,\delta x}.$$9.

A second pseudo-labeling usage appears in weakly supervised object localization as “Discriminative Pseudo-Label Sampling” (Murtaza et al., 2023). There, a frozen self-supervised transformer encoder produces multiple class-agnostic attention maps. DiPS thresholds each map with Otsu thresholding, extracts connected components, converts them to bounding boxes, and scores each box with a separate pretrained classifier via $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$0, where $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$1 retains the content inside the box and blurs the outside. The top-$$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$2 proposals form a diverse pseudo-label pool, and one proposal is randomly sampled at each training step to supervise a dedicated localization head. The localization objective combines partial cross-entropy on sampled foreground/background pixels with a CRF regularizer. Reported results include CUB performance of $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$3 MaxBoxAccV2, $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$4 Top-1 Loc, and $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$5 Top-5 Loc, ILSVRC mean MaxBoxAccV2 of $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$6 with $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$7 at $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$8, and TelDrone MaxBoxAccV2 of $$\frac{\delta t}{\delta x} = \frac{\delta\tau}{\delta x} \left[ 1+\frac{v^2}{4}\left(\frac{\delta\tau}{\delta x}\right)^2+\left(\frac{2h}{v t_0}\right)^2 \right]^{1/2},$$9.

These two uses are methodologically adjacent. Both convert unlabeled or class-agnostic structure into pseudo-supervised signals, but they solve different problems: DIPS for checkpoint ranking in UI2I pipelines and DiPS for training a localization head from self-supervised transformer proposals.

4. Policy selection and memory management

In high-stakes dialogue, DiPS denotes “Dialogue Policy Selection,” a framework that applies offline reinforcement learning, specifically Implicit Q-Learning, to choose among discrete persuasion personas in a wildfire evacuation scenario (Zhang et al., 2 Jul 2026). The dialogue is modeled as a POMDP $h$0 in which the observed state representation is a MiniLM embedding of recent resident utterances, actions are persona indices, and reward is sparse and terminal: success is whether the resident agrees to evacuate. The critic estimates $h$1, and inference uses

$h$2

The system is therefore a policy-over-policies rather than an end-to-end text generator. In human-resident experiments, overall success was $h$3 for DiPS, compared with $h$4 for a zero-shot operator and $h$5 for a generic RAG baseline; in improved simulation, full DiPS reached $h$6 success in $h$7 average turns, versus $h$8 and $h$9 for zero-shot, and $t_0$0 and $t_0$1 for global RAG.

In recommender systems, DiPS instead denotes “Differentiable Policy for Sketching,” a bilevel framework for learning which items to retain in a fixed-size sketch of a user’s long interaction history (Ghosh et al., 2021). The outer problem minimizes future prediction loss over recommender parameters $t_0$2 and policy parameters $t_0$3, while the inner problem adapts user-specific parameters from the current sketch. In the online setting, the sketch update is

$t_0$4

and in the batch setting DiPS uses a differentiable Top-$t_0$5 projection. The central technical issue is gradient estimation through discrete keep/evict decisions; the paper uses straight-through estimators and a queue of past intermediate sketches to approximate the total derivative efficiently. Empirically, the method achieves up to $t_0$6 fewer sketch items for the same predictive quality than heuristic policies. On MovieLens 10M in the online setting with $t_0$7, DiPS obtained RMSE $t_0$8, compared with $t_0$9 for Random, $\delta\tau/\delta x$0 for Hardest, and $\delta\tau/\delta x$1 for Influence.

The common thread is explicit control at a higher decision layer. Dialogue DiPS selects persuasion strategies turn by turn; recommender DiPS selects memory contents over time. In both cases, the learned policy acts on a compressed state rather than raw long-horizon history.

5. DIPS as data infrastructure and differentiable scientific computing

In structural biology, DIPS originally referred to the “Database of Interacting Protein Structures,” and DIPS-Plus is its feature-rich extension for protein interface prediction (Morehead et al., 2021). The original DIPS contains $\delta\tau/\delta x$2 binary complexes. DIPS-Plus contains $\delta\tau/\delta x$3 complexes because $\delta\tau/\delta x$4 were dropped after PDB unavailability or prohibitively expensive MSA generation; it includes $\delta\tau/\delta x$5 residues, $\delta\tau/\delta x$6 residue-residue interactions, and eight residue feature types: Secondary Structure, Relative Solvent Accessibility, Residue Depth, Protrusion Index, Half-Sphere Amino Acid Composition, Coordinate Number, Profile HMM Features, and Amide Normal Vector. Townshend et al.’s atom-level DIPS is thus extended into a residue-level geometric-deep-learning resource. Retraining NeiA+HOPI on DIPS-Plus yields MedAUROC $\delta\tau/\delta x$7 on DB5-Plus test complexes, exceeding previous values such as $\delta\tau/\delta x$8 for NeiWA+HOPI and $\delta\tau/\delta x$9 for SASNet.

In scientific computing, JAX-DIPS is a “differentiable interfacial PDE solver” that uses the Neural Bootstrapping Method to train neural surrogates from finite discretization residuals rather than from direct symbolic differentiation of the PDE operator (Mistani et al., 2022). The target PDE class is elliptic with jump conditions across irregular interfaces:

$t_0$0

NBM constructs implicit Cartesian cells around collocation points, evaluates finite-volume interface residuals there, preconditions them, and differentiates only with respect to network parameters using first-order AD. The implementation uses jax.jit, jax.vmap, and jax.pmap. The paper reports convergence with increasing collocation points and residual preconditioning, and argues that the approach inherits conservation laws and interface structure from the finite discretization. In a bulk test without interfaces, RMSE fell from $t_0$1 to $t_0$2 as training grids increased from $t_0$3 to $t_0$4.

These usages place DIPS closer to infrastructure than to a single model family: one is a large curated benchmark and feature bank for protein interfaces, the other a JAX-based neuro-symbolic solver architecture for PDEs with discontinuities.

6. Dip phenomena in astrophysics

In extragalactic spectroscopy, DiPS refers to central metallicity dips in radial gas-phase metallicity profiles (Easeman et al., 2022). Using $t_0$5 low-inclination local star-forming galaxies from SDSS-IV MaNGA, the study asks whether observed depressions in $t_0$6 near galaxy centers are physical or diagnostic artefacts. The work finds no clear evidence that the dips are caused by changing ionization parameter within galaxies. Galaxies exhibiting central dips in the O3N2 metallicity profile have, on average, lower H$t_0$7EW values out to $t_0$8 and higher central $t_0$9; dips are also more prevalent at high stellar mass and low global specific star-formation rate. The paper therefore suggests a link to central quenching, while emphasizing that the result is diagnostic-dependent and should be interpreted cautiously.

In compact-object astrophysics, DiPS denotes the recurrent optical dips observed in the black-hole transient Swift J1357.2-0933 (Panizo-Espinar et al., 2023). The source has $v$0, orbital period $v$1 h, and inclination $v$2. Optical dips were observed in every outburst: about $v$3 mag in 2011, $v$4 mag in 2017, $v$5 mag in 2019, and $v$6 mag in 2021, with individual durations of about $v$7 minutes. The dip recurrence period varied from $v$8–$v$9 min in 2011, $h$0–$h$1 min in 2017, $h$2–$h$3 min in 2019, $h$4–$h$5 min in 2021, and about $h$6 min in quiescence. The four outbursts do not share a common DRP–time evolution, but DRP correlates strongly with flux: for outburst data, the Pearson coefficient is $h$7 versus X-ray flux and about $h$8 versus optical flux. The proposed interpretation is a connection between optical dips and vertically extended disc or outflow structures, especially because previous work found blue-shifted optical absorption during dips.

Taken together, these astrophysical usages preserve the literal sense of “dips” rather than the acronymic sense dominant in machine learning. One concerns central depressions in metallicity profiles; the other concerns recurrent flux decrements in a high-inclination black-hole transient.

7. Cross-domain significance

Across these literatures, DiPS/DIPS functions as a compact label for three different kinds of objects. First, it can denote a derived relation or observable morphology, as in seismic reflector-dip transforms and astrophysical dip phenomena (Singh, 2022, Easeman et al., 2022, Panizo-Espinar et al., 2023). Second, it can denote a learning framework that converts weak, indirect, or delayed information into actionable supervision, as in Domain-knowledge Inspired Pseudo Supervision, Discriminative Pseudo-Label Sampling, Dialogue Policy Selection, and Differentiable Policy for Sketching (Al-Hindawi et al., 2023, Murtaza et al., 2023, Zhang et al., 2 Jul 2026, Ghosh et al., 2021). Third, it can designate enabling infrastructure, as in DIPS-Plus and JAX-DIPS (Morehead et al., 2021, Mistani et al., 2022).

This suggests that “DiPS” is best understood bibliographically as a family of acronymic reuses rather than a unified research program. The commonality lies in nomenclature, not in method, data modality, or mathematical formalism.