Targeted Deep Architectures (TDA)

• TDA is a family of deep learning designs that directs model architectures toward explicit structural objectives, such as topological invariants or influence-function gradients.
• Topology-targeted TDA integrates persistent homology with CNN features, enabling improved representation learning and effective clustering in application domains like semiconductor imagery.
• TMLE-targeted TDA employs selective gradient updates for causal inference, enhancing bias correction and statistical efficiency in semiparametric estimation.

Searching arXiv for the papers on arXiv to ground the synthesis in the current literature. Targeted Deep Architectures (TDA) denotes a family of deep-learning designs in which architectural structure or post-training updates are explicitly organized around a specified target rather than left to generic predictive optimization alone. The expression is not yet standardized. In one usage, a TDA is a topologically informed representation learner in which persistent-homology summaries are fused with learned visual features and carried through self-supervised and transfer-learning pipelines; for disambiguation, this usage may be called “topology-targeted TDA” (Editor’s term) (Giri et al., 5 May 2025). In another usage, “Targeted Deep Architectures” names a TMLE-based framework for causal inference in which a small subset of neural-network parameters is iteratively updated so that an efficient-influence-function estimating equation is approximately solved; this usage may be called “TMLE-targeted TDA” (Editor’s term) (Li et al., 16 Jul 2025). A broader analytical literature based on persistent homology, topological descriptors, bottleneck distance, and persistent homological fractal dimension provides the vocabulary with which architectures are compared as manifold transformers and motivates topology-aware targeting criteria (Magai, 2023).

1. Definition, scope, and terminological structure

The term spans several related but non-identical research programs. In much of the surrounding literature, the acronym TDA still denotes Topological Data Analysis, so “Targeted Deep Architectures” must be read contextually. This suggests that the phrase currently functions less as a single canonical model class than as a label for architectures whose inductive bias, training dynamics, or post-hoc adjustment is directed toward an explicit structural objective.

Usage	Core mechanism	Representative source
Topology-targeted TDA (Editor’s term)	Persistent-homology features are treated as first-class inputs and fused with learned CNN features	(Giri et al., 5 May 2025)
TDA-guided architecture analysis	Persistent homology, bottleneck distance, $E_i(X)$, and PHdim are used to characterize layerwise manifold evolution and compare model families	(Magai, 2023, Świder, 2024)
TMLE-targeted TDA (Editor’s term)	Most weights are frozen and a small targeting subset is updated along projected influence-function gradients	(Li et al., 16 Jul 2025)

Across these usages, the common element is not a single backbone or loss, but an insistence that deep representations should be shaped by an explicit target: topological invariants, desirable manifold profiles, domain-level structural compatibility, or valid semiparametric inference.

2. Topology-targeted TDA in representation learning and clustering

In the semiconductor-image framework, “Deep TDA” denotes a deep learning–driven representation-learning pipeline explicitly augmented with Topological Data Analysis. Images are first converted into topological inputs such as point clouds derived from pixel intensities or feature coordinates. Persistent homology is then computed across scales, focusing on $H_0$ connected components and $H_1$ loops, and the resulting persistence diagrams are vectorized into stable fixed-length feature representations such as persistence landscapes or persistence images. These vectors are not used merely for post-hoc explanation: they are treated as direct inputs to the embedding network, concatenated with CNN features, and optimized in a self-supervised contrastive framework (Giri et al., 5 May 2025).

Formally, if a CNN backbone $f_\theta$ produces $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$ and the topological pipeline produces $z^{\text{TDA}}\in\mathbb{R}^{d_T}$, the joint representation is

$$z = [z^{\text{CNN}}; z^{\text{TDA}}]\in\mathbb{R}^{d_C+d_T},$$

which is then mapped by an MLP projection head $g_\phi$ to a latent contrastive embedding $h = g_\phi(z)\in\mathbb{R}^d$. Training uses a SimCLR-like scheme with two augmented views per image and an NT-Xent objective with cosine similarity. The topological vector is computed on the base sample rather than recomputed for each augmentation. Persistent homology itself is a pre-step and is not differentiable in the reported implementation; gradient-based optimization applies to the downstream fusion network.

The architecture is “targeted” because it reserves a dedicated branch for topology and forces every learned embedding to carry a topological summary. This is particularly aligned with wafer-defect imagery, where defect classes have explicit global or meso-scale signatures such as edge rings, donut patterns, scratches, local clusters, and near-full wafers. The reported pipeline uses ResNet18 or EfficientNet backbones, grid-searches TDA parameters such as the proximity parameter $\beta$ and the embedding-space metric, and then clusters the learned embeddings with a TDA-enhanced density-based procedure represented as a mapper-like graph.

The reported applications are domain-specific and concrete. On WM811K, comprising 25,527 images across 8 single defect types, the chosen TDA configuration was $H_0$0 with Euclidean metric; the TDA map yielded 4 clusters + noise, with the largest cluster containing ~69% of images, another cluster almost purely Edge-Ring, and the rare Near-Full class appearing as a tail. On Mixed WM38, comprising 38,015 images across 38 classes with 79% multi-label defects, the selected configuration was $H_0$1 with Euclidean metric, and 31 clusters were produced. In zero-shot transfer, a pre-trained foundational model separated good and faulty images on SPVD and recovered largely defect-type-aligned clusters on SWED, with 20 clusters on SPVD and 8 clusters on SWED. A plausible implication is that topology-targeted fusion is particularly useful when the label structure is naturally expressed in terms of connectivity, holes, and large-scale spatial organization rather than purely local texture (Giri et al., 5 May 2025).

3. Manifold-based analysis and topology-guided design criteria

A second strand of the literature does not define a single TDA architecture, but provides a design language for targeted deep models by treating a deep network as a manifold transformer. In this view,

$H_0$2

with each block mapping a representation manifold $H_0$3 to $H_0$4. Persistent homology is computed on layerwise point clouds, usually via Vietoris–Rips filtrations, and summarized by topological descriptors

$H_0$5

where $H_0$6 is the lifetime of the $H_0$7-th homological feature, together with persistent homological fractal dimension (PHdim), which is estimated from the scaling of $H_0$8 across sample sizes. These quantities are used to measure connectivity complexity, cyclomatic capacity, and intrinsic dimensionality of representations (Magai, 2023).

The reported empirical regularities are architecture-dependent. For CNNs such as ResNet, SE-ResNet, MobileNetV2, and VGG, $H_0$9 and $H_1$0 generally decrease with depth during training, while PHdim exhibits a hump: it rises in early and intermediate layers and then declines sharply near the output. ViT behaves differently: topology across depth “almost does not change” relative to CNNs, although absolute descriptor values vary with training epoch. ConvMixer is closer to CNNs than to ViT. BERT and RoBERTa show decreasing topological descriptors across depth when well trained, but PHdim tends to decrease more steadily and earlier than in CNNs. The same work reports strong correlation magnitudes between final-layer topology and performance, including approximately 0.85–0.95 for CNNs and 0.872, 0.862, 0.932, and 0.954 for ViT, ConvMixer, BERT, and RoBERTa, respectively, in analyses where PHdim decreases as accuracy increases (Magai, 2023).

A related comparative study of ResNet, VGG19, and ViT sharpens the methodological conditions under which such claims are reliable. It finds that removing outliers has little impact on the results, that representations should be compared with the same number of elements, that models with similar architecture tend to have similar topology of representations, that models with a larger number of layers change their topology more smoothly, and that pre-trained and fine-tuned models remain similar in initial layers but diverge in middle and final layers (Świder, 2024). In that study, VGG19 and ResNets show CNN-like growth of $H_1$1 and $H_1$2 structure with depth, whereas ViT exhibits a rise–plateau–fall pattern in the number of $H_1$3 and $H_1$4 features, together with stronger late-block reorganization.

These results do not by themselves instantiate a targeted architecture, but they provide operational targeting criteria. This suggests that low final PHdim, low $H_1$5, and architecture-specific layerwise topology profiles can be used to rank model families, activation functions, and hyperparameters without appealing solely to test-set performance. It also suggests that “targeting” may be understood not only as inserting a new module, but as selecting architectures whose manifold evolution matches the structural demands of the task.

4. Topology regularization for transfer learning and domain adaptation

A more direct targeted use of topology appears in unsupervised domain adaptation. Here the objective is to align labeled source and unlabeled target domains not only by matching latent distributions, but also by shaping their global manifold structure. The reported architecture combines a LeNet-style feature extractor with latent bottleneck $H_1$6, a source-label classifier, a decoder enforcing reconstruction, a domain discriminator trained adversarially, a Bregman-divergence alignment term, and a differentiable topological layer acting on batches of latent codes (Weeks et al., 2021).

The added topology loss compares source and target persistence diagrams in the latent space: $H_1$7 with the Wasserstein distance approximated through Sinkhorn iterations. The training schedule is a cascade: autoencoder loss, classification loss, domain-adversarial loss, Bregman divergence, and finally topology loss. In the reported experiments, topological regularization is applied only during the last 3 epochs after 80 total epochs of standard training.

The central empirical result is negative but informative. On MNIST $H_1$8 USPS, a BD-only configuration achieved $H_1$9 target accuracy without topology, $f_\theta$0 with $f_\theta$1 topology, and $f_\theta$2 with $f_\theta$3. A DAd + Autoencoder configuration achieved $f_\theta$4 without topology, $f_\theta$5 with $f_\theta$6, and $f_\theta$7 with $f_\theta$8. A DAd + Autoencoder + BD configuration achieved $f_\theta$9 without topology, $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$0 with $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$1, and $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$2 with $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$3 (Weeks et al., 2021).

The paper’s interpretation is precise: aligning persistence alone is insufficient for transfer and must be considered together with the lifetimes of topological singularities. Longer lifetimes are associated with robust discriminative features and more favorable structure in data, whereas naive diagram matching can be satisfied by making both domains noisier or less discriminative. The same work therefore recommends homology-specific and lifetime-aware losses—for example, aligning $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$4 cluster structure while penalizing undesirable persistent $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$5 holes—rather than treating topological similarity as a scalar objective detached from task semantics. This is one of the clearest objective controversies in the TDA literature: topology matters, but the way it is encoded in the loss matters just as much.

5. TMLE-targeted TDA for causal and semiparametric inference

In causal inference, Targeted Deep Architectures denotes a specific framework that embeds Targeted Maximum Likelihood Estimation directly into neural-network parameter space. The setup begins with a standard predictive network $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$6, then partitions its parameters as

$z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$7

freezes $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$8, and iteratively updates only a small targeting subset $z^{\text{CNN}} = f_\theta(x)\in\mathbb{R}^{d_C}$9. The update direction is obtained by projecting an influence function onto the span of the gradients of the loss with respect to those targeting weights. This produces a plug-in estimator that removes first-order bias and yields asymptotically valid confidence intervals while imposing no restrictions on the backbone architecture (Li et al., 16 Jul 2025).

At a given targeting iterate, the projected score is obtained from the regularized least-squares problem

$z^{\text{TDA}}\in\mathbb{R}^{d_T}$0

where $z^{\text{TDA}}\in\mathbb{R}^{d_T}$1 is a raw influence-function estimate such as an AIPW score. The projected EIF is then

$z^{\text{TDA}}\in\mathbb{R}^{d_T}$2

and the targeting update has the form

$z^{\text{TDA}}\in\mathbb{R}^{d_T}$3

Convergence is monitored via

$z^{\text{TDA}}\in\mathbb{R}^{d_T}$4

The framework extends beyond scalar targets. For a vector-valued estimand $z^{\text{TDA}}\in\mathbb{R}^{d_T}$5, each component influence function is separately projected onto the same score space, producing coefficients $z^{\text{TDA}}\in\mathbb{R}^{d_T}$6. These are then merged into a single universal targeting direction

$z^{\text{TDA}}\in\mathbb{R}^{d_T}$7

with weights determined by the empirical means of the projected component EIFs. This enables joint targeting of high-dimensional parameters such as entire survival curves without requiring a separate external fluctuation model for each time point.

Theoretical claims are framed as a specialization of Adaptive Debiased Machine Learning. Under the stated regularity, projection-closeness, and empirical-process conditions, the estimator is asymptotically linear, semiparametrically efficient for an oracle parameter, and doubly robust when the underlying influence function has that structure. The two working examples are the average treatment effect and a marginal survival curve under right censoring with informative censoring.

The empirical evidence is correspondingly concrete. On the IHDP benchmark, reported coverage values were 73.5% for the naive plug-in, 84.0% for targeted regularization loss, 91.5% for A-IPTW and post-hoc TMLE, 92.0% for TDA with last-layer targeting, and 92.4% for TDA with full-head targeting. On simulated survival data with informative censoring, time-averaged MSE was 0.0029 for the initial neural hazard model, 0.0028 for Kaplan–Meier, and 0.0015 for TDA; absolute bias was 0.0413, 0.0407, and 0.0306, respectively; coverage was 75.5%, 55.8%, and 91.0% (Li et al., 16 Jul 2025). In this sense, TMLE-targeted TDA is not a topology method at all, but a neural implementation of targeting for valid statistical inference.

6. Related architectures, misconceptions, and open problems

Several common misconceptions follow from the overloaded acronym. The first is that TDA always denotes a topology method. In fact, one major use of the term refers to TMLE-style targeting in neural weight space rather than Topological Data Analysis (Li et al., 16 Jul 2025). The second is that all topology-targeted TDAs are end-to-end differentiable. In the semiconductor clustering framework, persistent homology is computed as a pre-step, vectorized, and injected as a fixed feature pathway; differentiable TDA layers are mentioned only as future work (Giri et al., 5 May 2025). The third is that any topology loss should improve transfer. The domain-adaptation results argue against this: naive persistence alignment can degrade target accuracy, and lifetime structure matters (Weeks et al., 2021).

A neighboring but distinct line is the Top-Down Attention Framework (TDAF), which implements an explicit controller–worker decomposition through a Recursive Dual-Directional Nested Structure and Attention Network Across Recurrence. TDAF is not presented as “Targeted Deep Architectures,” but it exemplifies architectural targeting in another sense: coarse, global attention features recursively modulate finer bottom-up features across multiple flows. Reported gains include 2.0% improvements on ImageNet for ResNet with TDAF, 2.7% AP over FCOS for object detection, 1.6% for pose estimation, and 1.7% accuracy for 3D-ResNet action recognition (Pang et al., 2020). This suggests that explicit targeting can also be realized through top-down control pathways rather than topological branches or statistical targeting updates.

Open problems are correspondingly heterogeneous. In topology-targeted models, reported limitations include proprietary or fixed PH vectorization, absence of full numeric ablations, sensitivity of raw persistent homology to preprocessing, and lack of end-to-end learned topological embeddings (Giri et al., 5 May 2025). In topology-regularized domain adaptation, the main unresolved issue is how to formulate lifetime-aware, homology-specific losses that reward favorable structure rather than mere diagram similarity (Weeks et al., 2021). In manifold-analysis work, computational cost still restricts routine use of higher-dimensional homology and makes PHdim difficult to integrate directly into gradient-based training, even though PHdim and $z^{\text{TDA}}\in\mathbb{R}^{d_T}$8 appear useful for test-free model selection (Magai, 2023). In TMLE-targeted TDA, the main practical issues are non-convex convergence, memory costs associated with storing per-sample gradients, choice of targeting submodel, and extension beyond i.i.d. settings (Li et al., 16 Jul 2025).

Taken together, these strands indicate that “Targeted Deep Architectures” presently names a research direction rather than a single settled architecture family. What unifies them is the deliberate replacement of generic representation learning by structurally specified guidance: topological invariants in one case, topology-aware manifold diagnostics in another, lifetime-sensitive alignment in transfer, and efficient-influence-function targeting in causal inference.