PyTorchKR Papers

최신 논문

총 523,590편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.05744v1cs.LG

Text-Only Domain Adaptation for Low-Resource Medical ASR: Shallow Fusion and Density Ratio Rescoring for Hausa

Automatic speech recognition (ASR) systems often struggle with specialised medical vocabulary, especially in low-resource languages where domain-specific speech data are scarce. We investigate text-only domain adaptation for Hausa medical ASR using the fixed MMS-1B-FL102 acoustic model and two decoding-time language-model fusion methods: Shallow Fusion (SF) and the Density Ratio Approach (DRA). A 4-gram KenLM trained on 19,421 Hausa health sentences serves as the medical target model, while a 75,372-sentence general-domain model provides the source prior for DRA. Evaluation uses 250 recordings of 25 physician-reviewed prompts from 10 speakers across six northern Nigerian states. Under prompt-disjoint five-fold cross-validation, SF and DRA reduce WER from 29.56% to 23.55% and 23.33%, respectively. Both significantly outperform greedy decoding (p < 0.001), while their difference is not significant (p = 0.423). However, general-LM-only SF achieves 23.96% WER, indicating that the medical LM's additional improvement over a general LM is modest. Medical Term Error Rate also improves under fusion, but does not establish a distinct medical-domain advantage. These findings demonstrate that lightweight, text-only language-model fusion can improve Hausa medical ASR without additional domain-specific speech collection.

arXiv:2611.05742v1cs.CV

Learning Hierarchical Patch Splitting Policies for Faster Vision Transformers

Vision Transformers (ViTs) are bottlenecked by the length of their input token sequence. Standard patchification fixes this length with a single patch size applied across the image, regardless of regional content. Prior work on adaptive patch sizing reduces token count by assigning coarser patches to visually simple regions, but relies on noisy pixel-level heuristics such as local entropy to make patchification decisions. We argue that this compute allocation should be grounded in the model's interpretation of task relevant information, not in raw pixel statistics. We introduce Split Policy Learned Image Tokenization which learns to allocate token resolution based on what the model finds meaningful, producing shorter token sequences that are better matched to task-relevant content. SPLIT matches the accuracy of full-resolution ViT classification baselines while using 30% fewer tokens. The same gains hold on dense prediction: SPLIT achieves competitive accuracy on COCO and ADE20K, demonstrating that semantically aligned patchification generalises well beyond classification.

arXiv:2611.05737v1cs.CV

C-WFDM: A ControlNet-Guided Wavelet-Fusion Diffusion Model for Multimodal Brain MRI Completion

Multimodal brain MRI provides complementary anatomical, tissue, and pathological information to support downstream artificial intelligence (AI) applications. However, modality availability is uneven across public and pooled neuroimaging datasets: T1-weighted (T1w) MRI is widely represented, T2-weighted (T2w) and fluid-attenuated inversion recovery (FLAIR) MRI are less consistently available, and gadolinium-enhanced T1-weighted (T1Gd) MRI is relatively scarce. Pooled datasets therefore contain participants with different combinations of acquired MRI contrasts, reflecting variation in the study designs, acquisition protocols, and clinical or research objectives of the contributing studies. Many missing-modality benchmarks simulate incomplete MRI by withholding contrasts from complete four-modality cases, which may not represent the observed availability patterns in pooled datasets. To address this gap, we introduce C-WFDM, a ControlNet extension of our Wavelet-Fusion Diffusion Model (WFDM), for subject-specific missing-modality completion. C-WFDM is built on an earlier BrainScape snapshot comprising 157 public MRI datasets and 45,880 preprocessed scans. The pretrained WFDM backbone retains target-modality and metadata conditioning, while target-specific ControlNet adapters use MRI contrasts from the same participant to guide synthesis. C-WFDM synthesizes T2w from T1w, FLAIR from T1w and T2w, and T1Gd from T1w, T2w, and FLAIR. Acquired conditioning contrasts are used when available, and generated intermediate contrasts are used only when required. We compare C-WFDM with ALDM, cWDM, and Pix2Pix using paired structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and mean absolute error (MAE). In the BrainScape completion experiment, C-WFDM achieves the best T2w and FLAIR results among the evaluated models, with SSIM values of 0.9232 and 0.8904, respectively.

arXiv:2611.05735v1cs.LG

Self-Supervised Learning of Universal Latent Spaces for Operator Learning

Neural operators fit their representation of functions to the task they are trained on, so every new task rebuilds a reduced space from labeled data. Following the self-supervised pre-training that made representations shareable in natural language processing and computer vision, we decouple representation learning from operator learning. A functional autoencoder is pre-trained on analytic functions synthesized on the fly, with no solver and no labels, to represent functions on arbitrary point clouds, and the latent space it induces is frozen and reused across tasks. The representation is exactly equivariant to translation, dilation, and amplitude, and reconstructs solutions of partial differential equations (PDEs) and natural images on geometries it never saw. Being task-agnostic, the latent space can be interpreted in its own right. The whole space is used and decodes to valid fields, it preserves the ordering of distances between physical fields, its geodesics interpolate physical evolutions, and edit directions transfer across fields. With frozen representation maps, a dataset is encoded once and a neural operator is trained entirely on cached embeddings, which saves memory and training time. On 28 tasks in fluid dynamics, solid mechanics, acoustics, and the Helmholtz and Poisson equations, this paradigm can exceed the accuracy of state-of-the-art end-to-end architectures, with up to an order of magnitude higher throughput and memory savings that fit models nearly two orders of magnitude larger on consumer-grade GPUs. At inference, latent rollouts run up to an order of magnitude faster than regular rollouts, and chunked inference processes functions of over half a billion points on a single consumer-grade GPU.

arXiv:2611.05734v1cs.CV

Mitigating Spurious Invariance in Contrastive Learning: A Probabilistic Self-Supervised Learning Framework for Medical Image Analysis

Owing to the prohibitive cost of manual annotation for enormous medical images, self-supervised learning (SSL) has gained substantial attention and shown promise in various medical imaging tasks. Among SSL approaches, contrastive learning has emerged as a prominent one, encouraging models to encode semantic information that remains invariant between different augmented views. However, this invariance can be spurious for medical images, as subtle but critical medical features can be suppressed during augmentations. In this paper, we mitigate this issue by encoding the learned feature embeddings as probability distributions rather than fixed vectors, thus capturing a richer and more nuanced spectrum of information. Furthermore, we propose a unified probabilistic framework that seamlessly integrates multiple SSL approaches. Our method achieves superior performance over the state-of-the-art SSL methods on various downstream tasks across multiple datasets, including medical image segmentation and classification. We further validate the method through additional analyses of robustness to stronger augmentations, label efficiency under limited fine-tuning labels, and computational trade-offs. The code is released at https://github.com/wwzjer/ProbSSL.

arXiv:2611.05731v1cs.CL

Prospective Memory for Proactive LLM Agents

Most memory systems for LLM agents are retrospective, preserving and retrieving past information to inform current decisions. Proactive agents, however, also require prospective memory: retaining future-directed intentions and realizing them when appropriate opportunities arise, even without explicit requests. Crucially, remembering the right intention does not guarantee acting at the right time: a related observation may call for waiting, while a later cue may make the same intention actionable. Inspired by cognitive theories of human prospective memory, we introduce Pro-Memory, a memory framework that separates maintaining future intentions from deciding when and how to realize them. Pro-Memory preserves pending objectives linked to their supporting evidence, assesses new observations against their original goals and activation conditions, and coordinates appropriate contributions with the agent's ongoing actions. It further updates intention states based on actions actually issued or evidence of resolution, while an experience memory distills feedback from completed episodes into guidance for future decisions. Experiments on two benchmarks spanning delayed-intention tasks and open-ended proactive assistance demonstrate consistent improvements in proactive task performance over representative memory approaches across multiple LLM backbones, while limiting unnecessary interventions.

arXiv:2611.05728v1stat.ML

Generative Modeling for Cascading Extremes via Autoregressive Transformers

Extreme events often occur sequentially, from earthquake-tsunami cascades to successive extreme losses in financial markets. We introduce an extremal token representation that encodes multivariate observations as a categorical sequence. A quiet token records the absence of threshold exceedances; otherwise, the token identifies the exceeding component with the largest marginal probability-scale score. The representation is invariant to componentwise strictly increasing transformations. It preserves the identity and temporal ordering of dominant exceedances but discards their magnitudes and the identities of nondominant simultaneous exceedances. A decoder-only transformer estimates the history-dependent next-token kernel on this representation, and recursive sampling generates future token sequences from the induced joint predictive law. We amortize our generative model by pretraining across simulated process families with varying parameters and dependence persistence. Next-token predictions are evaluated against exact conditional probabilities for held-out categorical processes and particle filter references for heavy-tailed vector autoregressive processes. Applied to ten-second returns for five semiconductor stocks, the pretrained model supports zero-shot transfer, while a few epochs of adaptation improve agreement between observed and generated token frequencies and transition patterns. These results support task-specific tokenization and simulation-based pretraining for modeling which dimensions dominate multivariate extremes.

arXiv:2611.05724v1cs.LG

Give the Drafter a Second Chance to Guess: Parallel Reconsideration for Speculative Decoding

Speculative decoding accelerates large language model inference with a small drafter, whose per-guess acceptance rate determines the achievable speedup. Feature-level drafters share a design choice: every token, easy or hard, is drafted under a single forward pass. Many of the tokens the target rejects are ones the drafter is itself unsure about yet the target answers decisively. To address this, we introduce Parallel Reconsideration: guided by its own uncertainty, the drafter reconsiders its hesitant guesses with extra passes that run in parallel with the one-pass draft, adding negligible latency. A co-designed training recipe and serving-time drafting algorithm append these reconsidered guesses to the original draft tree as refined candidates. Across models and benchmarks, on average, Parallel Reconsideration improves over the baselines by 5.3% to 10.2% on acceptance length and by 3.5% to 9.2% on speedup.

arXiv:2611.05723v1stat.ME

Balance-Guided Forest for Causal Effect Estimation

Covariate balancing is widely used for estimating causal effects in observational studies. A central challenge is determining which covariate features should be balanced. Forest-kernel balancing provides a data-driven way to learn informative balancing features. Existing forest-kernel balancing methods first learn a tree-based representation and then construct balancing weights. However, this procedure creates a mismatch, as limited tree capacity may be spent on features that are already well represented. We propose Balance-Guided Forest (BGF), which uses the current balancing weights to guide subsequent tree construction. BGF grows trees using a novel splitting score that combines residual imbalance with an outcome contrast, favoring splits that separate outcome-relevant regions where imbalance remains. This couples representation learning with covariate balancing rather than treating balancing as a downstream step. We establish an oracle-type bound that links the iterative construction to weighting error and causal bias. Simulations exhibit improvements in causal estimation relative to existing forest-kernel approaches. Real-data applications illustrate the practical behavior of BGF across observational studies.

arXiv:2611.05720v1cs.CL

Cognitive Bias Signatures in Search-Equipped LLM Agents: A Stage-Resolved Analysis

Search can help LLM agents correct mistakes, but their starting conditions may shape which evidence they find and use. Drawing on cognitive-bias paradigms, we develop a controlled search harness to examine result order, records of prior work, and initial hypotheses while holding each question and document collection fixed. Matched controls distinguish order effects from differences in document composition and the effects of failed work from context length. Across 10 models and 1,000 synthetic search worlds, we track queries, retrieved documents, document openings, and final answers. Our comparisons suggest that final answers are more sensitive to initial evidence order during active search than when all evidence is supplied in full. Dead-end histories can reduce strategy switching relative to matched neutral histories, indicating an effect beyond context length. Under wrong hypotheses, stage measurements distinguish reduced retrieval or opening from errors that persist after counter-evidence is opened. Exploratory interventions show that opening counter-evidence more often can coincide with lower accuracy, while accuracy can improve with little change in opening.

arXiv:2611.05715v1cs.LG

CoReFlow: Consistency-Relaxed Flow Maps for Probabilistic Traffic Forecasting

Probabilistic traffic forecasting is essential for reliable Intelligent Transportation Systems (ITS), as it captures uncertainties in complex road networks. Recent diffusion- and flow-based generative models have facilitated uncertainty modeling through a multi-step sampling process. However, beyond slow inference, this process is designed to capture complex multimodal distributions, which misaligns with the limited complexity of the conditional traffic data distribution, causing models to underperform deterministic baselines. To address this, we introduce the \underline{Co}nsistency-\underline{Re}laxed \underline{Flow} Map (CoReFlow) model, presenting the first effort to leverage the one-step generation capability of flow maps, known as solution operators of probability-flow ordinary differential equations. Besides, we improve accuracy using a concise training objective that supervises the model via interpolation endpoints over finite time intervals. Unlike existing methods (e.g., self-distillation and MeanFlow Identity) that maintain trajectory consistency to model multimodal distributions, our objective leverages conditional mean to restrict the consistency, thereby achieving higher accuracy in the probabilistic traffic forecasting task. Extensive experiments on four real-world datasets show CoReFlow achieves state-of-the-art forecasting accuracy. It surpasses recent baselines, provides competitive uncertainty modeling, and reduces inference costs compared to existing generative baselines.

arXiv:2611.05706v1cs.NE

Neural networks and games

Hopfield neural networks and noncooperative games provide seemingly different approaches to understanding how collective outcomes emerge from local interactions. In a neural network, individual units adjust their states according to local signals; in a game, players choose actions to maximise individual payoffs. I establish an exact equivalence between symmetric binary Hopfield networks and binary-action games that admit a quadratic potential. The equivalence identifies stable network configurations with pure-strategy Nash equilibria, and asynchronous neural updating with best-response dynamics. More broadly, the storage and retrieval of memories in neural networks correspond, respectively, to the design and selection of equilibria in games.

arXiv:2611.05697v1cs.CV

GeoDIVE: Geospatial Descent Informed by Value Estimation

Finding specific objects, structures, and spatial patterns supports critical work in agriculture, infrastructure planning, environmental monitoring, and disaster response. These features, however, are often dispersed across very large regions and span dramatically different spatial scales, which defines a search space that is costly to embed, maintain, and search naively. We introduce GeoDIVE, a model that treats a multiresolution satellite image gallery as a forest and uses best-first traversal to determine which tiles to encode. At each visited node, a lightweight model then predicts an optimistic estimate of the best query similarity hidden among its descendants using the query and subtree representations. In evaluations with image and open-vocabulary text queries, GeoDIVE matches exhaustive search at 10% of the encoding cost for image queries and leads all tested baselines at every budget in both modalities. To support future evaluation of multiscale geospatial search, we also release object-level image and text queries for Git-10M, constructed from Microsoft building footprints and OpenStreetMap polygons with VLM-generated captions. Source code and models will be made publicly available.

arXiv:2611.05693v1cs.CV

Anchored Slice Propagation with Boundary Transport for 3D CT Segmentation from Three Orthogonal Annotated Slices

Sli2Vol propagates a binary mask through affinities between adjacent slices, but an error can persist because each prediction becomes the input to the next step. Three orthogonal annotations reduce dependence on any one chain, yet they offer more than three starting masks: where a chain crosses another annotated plane, the correct label is already known. Those labels could correct drift during propagation and indicate how much to trust each chain at fusion. Between annotated planes, the information carried by the chain may matter as well. A binary mask records foreground membership but discards distance to the boundary, a cue that could help follow gradual shape changes between slices.We address these two sources of error with a fixed pretrained affinity backbone. We anchor each chain by correcting its prediction where it crosses another annotated plane, then use its agreement with those labels to weight fusion. We compare this anchored mask with two ways of transporting boundary geometry: a truncated signed-distance field (SDF) propagated by dense affinity averaging, and a contour whose vertices follow motion cues from the same affinity. The contour tracker starts from a hand-crafted rule and learns a motion correction from supervoxel pseudo-objects in 506 CT volumes, without manual organ labels. On 2,091 held-out structure instances from approximately 403 CT scans spanning 17 abdominal structures, anchoring raises mean Dice from 0.737 with orthogonal mask propagation to 0.769 and improves every structure. Beyond the anchored mask, learned contour tracking improves Dice by 5.7\% for the esophagus and 5.9\% for the postcava, while SDF transport improves right-adrenal Dice by 2.8\%. The anchored methods have similar mean Dice (0.769--0.772); contour tracking performs worse on the kidneys, stomach, and hepatic vessel tree.

arXiv:2611.05689v1cs.CV

TRUST-Stereo: Teacher-Assisted Robust Uncertainty-Guided Search and Transfer for Fast Stereo Matching

Coarse-to-fine stereo matching enables efficient disparity estimation but remains sensitive to errors in propagated disparity, which may place ground-truth correspondences outside subsequent local search ranges. We propose TRUST-Stereo (Teacher-Assisted Robust Uncertainty-Guided Search and Transfer for Fast Stereo Matching), which addresses this limitation through an Uncertainty-Guided Residual Search Module (UGRSM). UGRSM estimates the reliability of propagated disparity and predicts residual offsets to adaptively relocate local cost-volume centers. Additionally, we introduce a two-stage Senior-Junior Teacher Assistant knowledge distillation strategy to transfer rich monocular representations from a large foundation model to a lightweight backbone to achieve fast stereo matching performance. Experiments on KITTI 2012, KITTI 2015, ETH3D, and Middlebury demonstrate strong zero-shot generalization. Using only 1.4M synthetic stereo pairs, TRUST-Stereo achieves state-of-the-art results on several benchmark metrics, outperforming competing efficient methods trained with substantially larger datasets, including real-world pseudo-labeled samples. Specifically, it achieves a KITTI 2015 EPE of 0.93, ETH3D Bad 1.0 of 2.13% and Middlebury Bad 2.0 of 6.79% while maintaining inference at approximately 50 ms per stereo pair.

arXiv:2611.05680v1cs.CV

Toward an Optimal Configuration for Tree Species Classification from Bark Images: A Leakage-Safe Ablation Study

Background: The automatic identification of tree species from bark photographs is of practical interest for forestry inventory and the wood-processing industry, yet the design choices that govern the accuracy of deep-learning approaches have not been studied systematically, and some prior results may be inflated by leakage of images of the same trunk across training and test partitions. Methods: A systematic ablation study is conducted on the public BarkNet 1.0 dataset (20 species) under the established tree-level, leakage-safe partitioning scheme. Using ConvNeXt-V2 backbones, the following factors are varied in turn and evaluated by five-fold cross-validation: patch size, the strategy by which per-patch predictions are aggregated into an image-level decision, the training regime (single- versus two-stage), backbone capacity, the whole-image baseline, and the number of patches available at inference. Results: Per-patch accuracy rises monotonically with patch size (from 75.6% at 160 px to 89.6% at 512 px), and image-level accuracy rises in step, plateauing near 94.9% around 512 px. Learned aggregation exceeds majority voting by at most 0.7 pp, an advantage that shrinks as patches grow and attention becomes near-uniform. Backbone capacity (pico to tiny) alters accuracy by less than a point, and two-stage training offers no advantage over one-stage. Whole-image classification (91.9%) trails patch-based aggregation, since it admits no bag to average over; accuracy increases with the number of patches at inference, saturating around 64. Conclusions: The study yields a simple recommended configuration: patches as large as possible while leaving enough to average over (around 512 px here), with a nano backbone. It also quantifies the second-order role of learned aggregation and model capacity, providing a reproducible reference point for practitioners and future research

arXiv:2611.05666v1cs.LG

Revisiting Critic Estimation in Soft Actor-Critic Algorithms

Soft Actor-Critic (SAC) is a widely used deep reinforcement learning algorithm that combines entropy regularization with an actor-critic architecture. Standard implementations rely on one-step temporal-difference learning of Q-functions, yet alternative critic design choices remain underexplored. In this work, we revisit three such choices: learning soft advantages instead of soft Q-values, the use of multi-step return targets, and the training of multi-step critics from replay buffer data. Keeping the policy objective fixed, we compare the six critic variants these choices generate: (i) advantage and (ii) dueling-advantage critics learned from one-step transitions, and $\lambda$-return and soft Generalized Advantage Estimation (GAE) critics learned either (iii, iv) on-policy or (v, vi) from replayed trajectory segments. Each variant is evaluated across an entire random hyperparameter search on fourteen tasks, treating the hyperparameters as part of the algorithm. Learning soft advantages instead of soft Q-values does not improve performance, and the soft GAE critic adds nothing over the simpler $\lambda$-return critic. What does improve performance, most clearly on continuous control, is training a $\lambda$-return critic on replayed trajectory segments, even without importance weighting or trace cutting. This extends earlier evidence that uncorrected multi-step targets work well under replay to the full soft $\lambda$-return, entropy terms included. By design, the underlying ideas studied here extend beyond SAC, and our empirical findings aim to encourage their application to a broader class of reinforcement learning algorithms.

arXiv:2611.05659v1cs.LG

Adaptive Domain Decomposition for Finite Basis Physics-Informed Neural Networks

In this work, we present adaptive domain decomposition finite basis physics-informed neural networks (ADD-FBPINNs). The proposed method integrates Partition of Unity Networks (POUNets), a data-driven approach for learning domain partitions, with finite Basis physics-informed neural networks (FBPINNs) to perform more efficient domain decomposition and achieve more accurate approximations. FBPINNs improve upon the accuracy of standard physics-informed neural networks (PINNs) through user-defined subdivisions of the computational domain. By adapting the domain decomposition based on the solution's features using a separable POUNet, the framework can automatically learn partitions that are stable and can be geometrically interpretable. We test this approach on the two-dimensional Helmholtz equation, two-dimensional Poisson equation with multiscale features, and viscous Burgers equation and observe improved accuracy compared with FBPINNs with fixed domain decomposition.

arXiv:2611.05658v1cs.LG

Beyond Accuracy: Information-Theoretic Decoder Evaluation Under Latent Intent

A central challenge in assistive brain-to-robot control is decoding a user's latent intent from complex and noisy brain signals. Prior approaches evaluate decoders as single-trial classifiers with accuracy-based metrics, but this neglects the inherently sequential nature of assistive control. Since a planner must integrate intermittent, noisy decoder outputs over time, single-step accuracy can only be a proxy when predicting downstream task performance. We formalize brain-to-robot control as a partially observable Markov decision process (POMDP), and prove that there is a sharp phase transition: regret is uniformly bounded across the horizon as long as pairs of goals are all decoder-distinguishable, but grows linearly otherwise. We show that a single channel statistic, the worst-pair Chernoff information $\Gamma_{\min}$ sets this boundary. Then, we derive a closed-form regret predictor $U_{\mathrm{id}}(C)$, that can be computed from any confusion matrix in $O(K^2)$ time. Across two domains spanning orthogonal information channels, we show that $U_{\mathrm{id}}$ predicts cumulative planning regret beyond what accuracy can explain. We then highlight a pathological accuracy trap, in which two decoders have identical validation accuracy 35.1% on real EEG data (BNCI2014001 subject 5) but result in a difference of $9.6\times$ in held-out regret, which $U_{\mathrm{id}}$ predicts. Same-signal control confirms $U_{\mathrm{id}}$'s advantage with decoders that span independent feature channels. Ultimately, our work provides a drop-in diagnostic to evaluate and select sensors and decoders in sequential assistive control.

arXiv:2611.05655v1cs.LG

When Are Two SAE Latents the Same Feature?

Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data. We develop a framework that connects these tasks through an energy-weighted optimal-transport plan under a specified similarity geometry. The same plan defines a dictionary-level discrepancy, decomposes it into latent contributions, and quantifies how each latent's mass is distributed across counterparts. We characterize relationships between decoder, activation, and first-order causal similarities; idealized models clarify the interpretation of splitting and absorption. For fixed trained SAEs, we derive joint asymptotic inference for transport costs and fractions, accounting for activation similarities and both energy marginals estimated from the same documents. A joint document bootstrap gives asymptotically valid simultaneous confidence bands. Experiments across four language models locate cross-seed discrepancy and separate context-dependent changes.