H2DiLR: Multi-domain Disentanglement & Learning

• H2DiLR is a multifaceted framework unifying techniques across neural decoding, LiDAR pre-training, optimization, and HDR image adaptation.
• It employs structured disentanglement, hierarchical semantic distillation, curvature-informed differential learning rates, and energy-preserving dilation to tackle domain-specific challenges.
• Applications include improved tone decoding in sEEG, state-of-the-art LiDAR perception for autonomous driving, efficient deep learning optimization, and accurate HDR to LDR mapping in virtual production.

H2DiLR refers to several distinct methodologies across scientific disciplines, each unified by the theme of structured disentanglement or hierarchical transformation for improved representation, learning, or signal management. The term appears in intracranial neural decoding (“Homogeneity–Heterogeneity Disentangled Learning for neural Representations”), in LiDAR self-supervised pre-training (“Hierarchical Distillation with Diffusion for LiDAR”), in efficient hyperparameter optimization (“Hessian-informed Differential Learning Rate”), and in high-dynamic-range image adaptation (“HDR Lighting Dilation for Dynamic Range Reduction”). Each usage reflects a context-specific formulation with rigorous mathematical and algorithmic underpinnings.

1. Homogeneity–Heterogeneity Disentangled Learning for Neural Representations

H2DiLR in neural decoding, introduced in the context of invasive brain-machine interface research, addresses the challenge of unified lexical tone decoding from heterogeneous intracranial (sEEG) signals (Wu et al., 2024). Heterogeneity in such signals is intrinsic, arising from both idiosyncratic participant anatomy and variable electrode configurations, which traditionally forced use of subject-specific models incapable of leveraging pooled data. H2DiLR overcomes this limitation by explicit latent variable disentanglement.

The architecture proceeds in two primary stages: a self-supervised vector-quantized autoencoder encodes each trial’s neural time series, partitioning its latent variables token-wise via per-sample ranking into shared (“homogeneous”) and private (“heterogeneous”) discrete codebooks. Shared tokens capture subject-invariant features critical for lexical tone, while private tokens capture subject- or electrode-specific variability. Decoding is performed by a lightweight Transformer over the concatenated codebook embeddings, trained for tonal-classification via cross-entropy.

The H2DiLR loss integrates reconstruction, commitment, and codebook-update terms. In formal notation, for latent $\{z_{n,i,j}\}$ from subject $i$ and instance $n$: $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$ with codebook vectors updated by exponential moving average. The partition ratio $\nu$ governs the tradeoff between homogeneity and heterogeneity, empirically optimal at $0.5$.

Empirically, H2DiLR increases mean tone-decoding accuracy from $\sim$37% (subject-specific) and $\sim$40% (VQ-only UPaNT) to $43.7\%$ on 407-syllable Mandarin sEEG, and up to $46\%$ with enlarged network capacity. Homogeneous codes cluster by tone, while heterogeneous codes index subject identity, with %%%%1$\{z_{n,i,j}\}$1%%%%1 subject-classification accuracy using private tokens. This approach establishes the first unified, cross-subject model for tonal neural decoding, addressing both data scarcity and heterogeneity (Wu et al., 2024).

2. Hierarchical Distillation with Diffusion for LiDAR Pre-training

In LiDAR perception for autonomous driving, H2DiLR (referred to as “Hierarchical Distillation with Diffusion for LiDAR Pre-training”) designates a comprehensive framework integrating discriminative and generative objectives for transferring semantically rich, geometrically coherent representations from camera to LiDAR backbones (Wozniak et al., 18 Jun 2026).

The teacher is a frozen ViT-based Vision Foundation Model (e.g., DINOv2), providing both multi-layer dense per-pixel features and global [CLS] tokens from V synchronized camera images. The student is a sparse convolutional U-Net (e.g., MinkUNet34), extracting per-point features at each down-sampling plane from voxelized LiDAR point clouds.

Three core objectives are optimized:

1. Multi-Layer (Progressive) Semantic Distillation: Calibrated point–pixel correspondences allow progressive, layer-wise projection of LiDAR features onto matched ViT teacher features via minimizing cosine distances. This injects not only “what” semantics but also their hierarchical buildup into the LiDAR backbone.
2. Global Context Distillation: The [CLS] tokens (after max-pooling across views) encode scene-level semantics. The LiDAR backbone produces an analogous max-pooled global descriptor, with alignment enforced by $i$2 loss, ensuring holistic context transfer.
3. Temporal Occupancy Diffusion Objective: A generative auxiliary task, using a DDPM-style conditional denoising objective on future BEV occupancy maps, conditions on multiple past LiDAR sweeps. The student must learn both coarse (global) and fine (temporal/geometric) properties across a noise schedule.

The total loss is a weighted sum: $i$3 with typical $i$4, $i$5, $i$6.

Pre-training with H2DiLR yields state-of-the-art cross-modal distillation accuracy for 3D object detection, scene flow, and semantic occupancy, surpassing prior transfer schemes. Individually, each component improves accuracy and representation quality; together, they produce sharper occupancy boundaries, improved temporal consistency, and reduced misclassification on global scene types (Wozniak et al., 18 Jun 2026).

3. Hessian-Informed Differential Learning Rate (Hi-DLR)

H2DiLR (“Hi-DLR”) also names an optimization paradigm for efficient, curvature-aware differential learning rate adaptation in deep neural training (Xu et al., 12 Jan 2025). Differential learning rate (DLR) strategies allow per-group step-sizes, commonly in PEFT settings (e.g., LoRA). Hi-DLR frames the selection of group-wise learning rates as a (typically small, $i$7-dimensional) hyperparameter optimization problem at each (or every $i$8-th) step, using a second-order Taylor approximation: $i$9 The minimizer for each group is then

$n$0

Estimation of the linear and curvature coefficients ($n$1) is performed efficiently using “forward-sampling”: 4$n$2 forward-pass loss evaluations per hyperparameter adaptation (with Gaussian-sampled per-group step-size perturbations), fitting per-group univariate quadratics.

Parameter grouping is flexible (per-layer, per-module, per-task head). Increased $n$3 improves adaptation granularity but raises computational overhead, mitigated by reducing adaptation frequency $n$4. Hi-DLR maintains computational efficiency ($n$51/[1+4$n$6/3$n$7] of base optimizer speed with $n$8 groups, adaptation every $n$9 steps), as no full Hessian needs formation or storage.

Empirical results demonstrate that Hi-DLR consistently accelerates convergence and achieves higher test accuracy than uniform or static learning-rate schedules in LoRA PEFT, ViT classification, NAM regression, and multi-task head tuning settings. The technique provides a principled, lightweight alternative to full Newton methods, and enables parameter group selection or pruning based on a computed influence statistic (PPI), permitting “meta-PEFT” pruning with negligible accuracy loss (<1% parameter retention) (Xu et al., 12 Jan 2025).

4. HDR Lighting Dilation for Dynamic Range Reduction

H2DiLR in image-based lighting denotes “HDR Lighting Dilation for Dynamic Range Reduction on Virtual Production Stages” (Debevec et al., 2022). This technique addresses the practical problem of mapping HDR environment maps (commonly exceeding the dynamic range of LED panels) to displayable LDR equivalents, while preserving total radiometric energy and avoiding color artifacts.

The method identifies over-bright regions in the HDR map ($$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$0), partitions them into connected components, then grows each region iteratively via symmetric dilation until the average (solid-angle-weighted) radiance of the region falls below the panel display threshold $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$1. Each dilated component's pixels are then replaced by the region's average. The procedure is mathematically guaranteed to preserve the sum of pixel energies: $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$2 Pseudocode details this process, with complexity $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$3 for $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$4 average dilation iterations over $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$5 pixels. Quantitative evaluations show the method eliminates over-bright pixels (no pixel exceeds $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$6 post-processing), preserves total energy ($$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$7 error 0), and produces LDR renderings matching original contrast and color balance to within $$\hat z_{n,i,j} = \begin{cases} e^S(M^S_{n,i,j}), & R_{n,i}(j) \leq \nu L\ e^P_i(M^P_{n,i,j}), & R_{n,i}(j) > \nu L \end{cases}$$810% mean luminance. Color fringing is avoided by dilating all RGB channels jointly and using symmetric structuring elements.

On virtual production LED stages, H2DiLR enables accurate reproduction of challenging IBL scenarios despite hardware limitations (Debevec et al., 2022).

5. Comparative Table of H2DiLR Meanings

H2DiLR Context	Core Mechanism	Application Domain
Neural decoding	Latent disentanglement; VQ codebooks	sEEG/ECoG tonal language decoding
LiDAR pre-training	Hierarchical distillation, diffusion objective	3D AD perception (LiDAR backbones)
Learning rate tuning	Hessian-informed groupwise step-size HPO	Deep learning optimization
HDR mapping	Morphological dilation, energy-preserving avg.	Virtual production/lighting

The diversity of domains underscores H2DiLR’s generalizable emphasis on the principled separation, transfer, or allocation of structure—be it in latent codes, learned semantic hierarchies, hyperparameter assignment, or illumination energy.

6. Future Directions and Limitations

Across contexts, open questions pertain to scaling H2DiLR variants to larger cohorts (in sEEG/BCI), further interpreting or adapting discrete neural codes, generalizing distillation frameworks in cross-modality transfer, and ensuring computational scalability as model or application complexity increases. Extension of meta-learning strategies and anatomically/semantically informed disentanglement, as well as broader adoption of second-order-informed hyperparameter adaptation, represent important research avenues (Wu et al., 2024, Wozniak et al., 18 Jun 2026, Xu et al., 12 Jan 2025, Debevec et al., 2022).

A plausible implication is that the H2DiLR paradigm—explicitly structuring or adapting based on interpretable axes—could serve as a common foundation for advances in data-scarce or cross-domain learning settings.