Daily Summary for 2026-11-10
daily
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: It's the tenth of November, twenty twenty-six, and this is the day's research.
Dev: 30 new papers came out today.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We'll take the day in one pass, then pull out the papers we're staying with.
Dev: Start us off with Spectral Regimes of Truncated sinc-Power and Pulse-Train Kernels.
The day's research: Rosa: We are focusing on improving how embodied AI systems select communication beams in millimeter-wave networks.
Dev: Fast and reliable beam selection is a major bottleneck when offloading massive sensory data like video and LiDAR point clouds.
Rosa: The main effort involves proposing a novel framework called In-Context Learning driven Cross-Validation Conformal Prediction to tackle this problem.
Dev: This framework uses in-context learning to adapt quickly from just a few examples. It combines this with conformal prediction to create compact candidate beam sets that are provably reliable, which is much more efficient than traditional methods.
Rosa: This approach builds on classical conformal prediction ideas where a dataset is used to set confidence bounds for new predictions.
Dev: The most significant development is the BVH-accelerated dynamic ray-tracing channel model, which addresses a major bottleneck in evaluating sixth-generation networks by shifting from time-driven to spatial-driven updates.
Rosa: This method is crucial because accurate channel modeling is necessary for 6G evaluation. It achieves a substantial computational time reduction of seventy-five point six percent compared to previous approaches.
Dev: This framework proposes a spatial-driven update mechanism to accelerate dynamic ray tracing in vehicular communication systems using bounding volume hierarchies.
Rosa: That paper is A Novel BVH-accelerated Dynamic Ray-Tracing Channel Model for Vehicular Communication Systems. What about the spectral regimes of truncated sinc-Power and pulse-train kernels?
Dev: That paper classifies spectral regimes based on the regularity of symbol calculus operations. It helps us understand signal characteristics better.
Rosa: How does that classification relate to beam selection efficiency in the embodied AI systems work?
Dev: The classification informs which beam selection strategies might perform best under certain spectral conditions for those systems.
Rosa: We also have prNet: Data-Driven Phase Retrieval via Stochastic Refinement, which uses Langevin dynamics and model updates to improve phase retrieval by balancing fidelity and perceptual quality.
Dev: That focuses on phase retrieval quality; does it impact the beam selection framework's accuracy?
Rosa: Not directly, but better phase information could refine the input data used for beam prediction. It is a separate concern.
Dev: Agreed. We also have A Particle-Based Multiuser Communication with a Budget Limit, which determines the minimum particle budget required for exact counting in a multiuser communication network.
Rosa: That deals with resource limits in communication, distinct from the beam selection mechanism itself.
Dev: Precisely. Decoder-Guided Lossy Contour Coding Via Anchor Refinement proposes a coarse-to-fine contour coding framework that uses decoder guidance to achieve lossy compression with rate-distortion control.
Rosa: That is about compression efficiency; how does it compare to the computational savings from BVH acceleration?
Dev: They address different bottlenecks. One is data throughput, the other is channel estimation speed for 6G evaluation.
Rosa: Then there is Finite Element Elastographic Analysis of Kidney MRI for Computational Assessment of Chronic Kidney Disease, which develops a finite element method to estimate kidney viscoelasticity from MRI and classifies chronic kidney disease using machine learning.
Dev: That's biomedical imaging; no direct link to millimeter-wave beamforming research here.
Rosa: Indeed. A Unified Multi-Encoder Framework for Music Lyric Intelligibility Assessment: Bridging Invasive and Non-Invasive Paradigms uses a unified multi-encoder approach to predict music lyric intelligibility by combining acoustic, semantic, and textual cues.
Dev: That's audio processing; does the multimodal input complexity affect the AI system's decision making under uncertainty?
Rosa: The complexity is high, but the beam selection model must handle that data stream effectively for embodied agents.
Dev: FUSEMOS: Perceptual Evaluation of Text-to-Music Generation with Dual-Encoder Fusion and Ranking-Aware Composite Loss proposes a dual-encoder fusion framework with a ranking-aware loss for better perceptual evaluation of text-to-music generation.
Rosa: That is about generative audio quality; it doesn't directly inform our conformal prediction bounds.
Dev: It shows how perceptual metrics are integrated into complex models, which could influence the required reliability thresholds in our selection process.
Rosa: We also have Phase-Matched Design of Pinching-Antenna Elements on a 60-GHz Dielectric Rod Waveguide, which derives a phase-matching rule to guide the fabrication of pinching antennas based on element shape for optimal performance.
Dev: That is hardware design; it focuses on antenna physical realization rather than network protocol selection.
Rosa: True. StreamAZ: Line-by-Line SAR Focusing with State Space Models presents an online SAR image formation framework that uses state-space models to achieve low latency and high focusing quality in radar.
Dev: Low latency is important for real-time sensing, but it's a different domain than communication beam steering.
Rosa: Copula-Based Dependence Modeling for Resilient 6G Multi-Hop Wireless Networks develops a copula-based framework to model dependence among link failures in multi-hop wireless networks to improve outage analysis.
Dev: That models network topology resilience, which is related but not identical to the beam selection mechanism itself.
Rosa: A Cytology Dataset for Early Detection of Oral Squamous Cell Carcinoma introduces the first large, multicenter oral cytology dataset to train AI models for early detection of oral squamous cell carcinoma.
Dev: Medical datasets are useful for training general AI models, but not specific to the beam selection optimization here.
Rosa: Re-Planning Budgets under Admission Control in Grant-Free LEO Beam-Hopping Access derives an explicit budget expression for worst-case delivery rate loss in grant-free LEO beam hopping access.
Dev: That addresses resource allocation limits in access networks, which is a constraint on our system design.
Rosa: SubRoformer: Subband Time-Domain Representation for Music Source Separation combines a subband module with a residual RoFormer to extract vocals and instruments from mixed music recordings.
Dev: Signal separation is complex; it highlights the need for robust feature extraction, similar to what we do with sensory data.
Rosa: Spectral Convergence of Sparse Bipartite Graphs via the Line-Graphon recovers eigenvalue-level spectral convergence for sparse bipartite graph sequences by exploiting the properties of their line-graph sequences.
Dev: That is graph theory applied to spectral analysis; it's abstract compared to our physical channel modeling.
Rosa: Vertical Segmented Waveguide-Enabled Pinching Antenna Assisted EDMA-NOMA Integrated Sensing and Communication Systems proposes a framework using vertical segmented antennas to simultaneously support communication and sensing in NOMA networks.
Dev: Integrating sensing and communication is key; it forces us to consider multi-task optimization in our selection process.
Rosa: ST-CGMDiff: Sparse-Conditioned Cross-Frequency Spatio-Temporal Latent Diffusion for Dynamic 3D Channel Gain Map Reconstruction proposes a latent diffusion framework to reconstruct dynamic three-dimensional channel gain maps across multiple frequency bands.
Dev: Reconstructing the channel dynamically is what we are trying to do, but this uses diffusion models for map reconstruction.
Rosa: One ISAC Array, Two Worst Cases: Hardware Impairments in Beamforming and AoA Estimation analyzes how element-level errors affect beamforming and angle of arrival estimation in integrated sensing and communication systems.
Dev: Element error analysis is critical; it sets the practical limits on how well our theoretical models can perform in practice.
Rosa: Adapting Sleep Foundation Models for Dense Sleep Spindle Localization adapts a sleep foundation model to localize sleep spindles in electroencephalography data using local time-frequency pathways.
Dev: Again, specialized AI adaptation; it shows model flexibility but not direct beam selection improvement.
Rosa: From Vision to Context: A Review of Prompt-Driven Target Volume Segmentation in Radiotherapy systematically compares different prompt-driven segmentation techniques for radiotherapy target volume delineation.
Dev: Segmentation accuracy is vital for data interpretation, but the beam selection relies on channel state information, not image segmentation quality directly.
Rosa: Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM proposes a hardware architecture that uses candidate pruning to achieve sort-free MIMO box decoding at high QAM orders efficiently.
Dev: That is about efficient decoding hardware; it's focused on implementation speed rather than the selection algorithm itself.
Rosa: Delay-Doppler Synchronization for Distributed Uplink Interference Nulling in LEO Networks derives a closed-form approximation for residual interference in two-branch interference nulling under LEO network conditions.
Dev: Interference nulling is related to beamforming, but this is about mitigating external noise rather than optimizing the initial beam choice.
Rosa: Intra-Cell Spatial Reuse in LEO Downlinks: Beam-Leakage-Aware Satellite Scheduling proposes a beam-leakage-aware scheduling algorithm to maximize capacity by serving multiple terminals concurrently in dense LEO downlinks.
Dev: Beam leakage awareness is relevant because our system operates in dense networks where interference between beams matters.
Rosa: Joint Association and Beamforming Design for Cooperative ISAC Systems develops a framework to jointly optimize association and beamforming for cooperative integrated sensing and communication systems.
Dev: That directly addresses the trade-off between selection and association, which is central to our problem.
Rosa: Deep Plug-and-Play Proximal Distance Algorithm for Robust Principal Component Analysis proposes a plug-and-play method that integrates deep priors into robust principal component analysis to improve performance.
Dev: Improving the robustness of the underlying estimation algorithms is certainly valuable for reliable beam selection.
Rosa: Unified Motion Correction and Reconstruction for Accelerated Zero Echo Time MRI develops a MAP estimation method to jointly estimate 3D images and rigid motion parameters in accelerated zero echo time MRI.
Dev: That's medical image reconstruction; it deals with spatial alignment, which is a different kind of geometric problem than beam steering.
Rosa: We have covered the main papers on channel modeling and selection frameworks. It shows the breadth of related research outside our immediate bottleneck area.
Dev: Agreed. The BVH acceleration and the conformal prediction framework are most directly applicable to improving our selection efficiency today.
Rosa: Indeed, those two developments define the core of what we achieved in this review cycle for embodied AI systems.
Lucky paper: 2611.08401: Rosa: Okay, so we're looking at the Spectral Regimes of Truncated sinc-Power and Pulse-Train Kernels today.
Dev: The paper investigates how a signal band-limited to one rectangular band behaves when that rectangular band is deformed through various symbol calculus operations.
Rosa: It sets up a taxonomy for spectral regimes based on the regularity of the symbol, identifying four types: indicator, smooth, atomic, and self-similar.
Dev: The indicator regime features a universal coefficient of one over pi squared per band edge because there's a specific jump in the function at the band edge.
Rosa: That jump is what sets that constant value for that regime. What about the smooth regime?
Dev: In the smooth regime, we replace that initial jump with an algebraic edge of exponent p minus one, and the spectral edge exponent is governed by this power p minus one.
Rosa: So in contrast to the indicator case where it's a jump, in the smooth case it becomes a rate at which the smallest eigenvalues approach zero.
Dev: Exactly; that rate is what defines the spectral-edge exponent for that regime. How does that relate to what we see in beam selection?
Rosa: It gives us insight into how well we can estimate channel characteristics when dealing with signals that aren't perfectly smooth.
Taro: From a representation learning standpoint, this classification helps us understand the underlying structure of the signal representation itself, whether it’s sparse or continuous.
Dev: That's right; Taro is looking at how the mathematical structure dictates the observable spectral behavior.
Rosa: What about atomic symbols in that paper? What do we get when you take the limit where each band shrinks to zero width?
Dev: In that atomic regime, the operator becomes finite rank, and there's a universal descriptor which is an integer two M.
Rosa: That integer value tells us about the rank structure, which is directly related to subspace estimation in communication theory.
Dev: It’s a clear structural property derived from shrinking those bands down to zero width. How does that compare to what we do with our candidate set size reduction?
Rosa: Both rely on understanding the underlying structure of the signal or channel characteristics, but one is about prediction and the other is about signal analysis.
Taro: I wonder if knowing when we are in an atomic versus a self-similar regime could guide our choice between different candidate beam set generation methods.
Dev: That’s a good thought; if we know the signal structure, we can choose the most efficient way to generate those candidates.
Rosa: And then there's the self-similar limit, which yields a cluster hierarchy whose count versus mass scaling is governed by the Cantor dimension D equals log two over log three.
Dev: That Cantor dimension D equals log two over log three governs how that hierarchy scales as you look at different levels of detail.
Rosa: That fractal dimension is independent of the vertical labeling of plateau heights, which suggests a very robust scaling property.
Taro: The independence from the labeling sounds like a very powerful feature for generalizing results across different signal processing contexts.
Dev: It means that no matter how you label those energy levels in the signal, the underlying scaling behavior remains consistent with that dimension D equals log two over log three.
Rosa: So we have these four distinct regimes: indicator, smooth, atomic, and self-similar in the Spectral Regimes of Truncated sinc-Power and Pulse-Train Kernels paper.
Dev: Right; the whole point of that paper is to give us a language for describing how signal regularity changes under different mathematical deformations.
Taro: It really shows how fundamental signal properties can be described using concepts from fractal geometry.
Lucky paper: 2611.07171: Rosa: We’re looking at A Particle-Based Multiuser Communication with a Budget Limit, which is paper number two of five on this show.
Dev: So, the core question here is how much particle budget you need to recover the exact sum of binary inputs in a multiuser network.
Rosa: The authors found that for simultaneous on-off keying or OOK, the minimum particle budget grows quadratically with the number of transmitters when we keep symbol duration and target error probability fixed.
Dev: That quadratic scaling is what makes it challenging at scale; how does that relate to what we're seeing in beam selection for embodied AI?
Taro: From a communication perspective, that quadratic scaling means if you double the number of transmitters, you need four times the particles just to maintain the same error target.
Rosa: Right, and they note that any budget growth slower than quadratic causes the error probability to approach one at a fixed symbol duration, no matter what your transmission or decoding scheme is.
Dev: That sounds like a hard limit; so if we can't get faster than quadratic scaling, that constrains how much data we can reliably process.
Taro: The paper also looks at groupwise on-off keying or GOC schemes to try and reduce that budget requirement by allowing longer symbol durations and more counting intervals.
Rosa: That’s where they show some improvement; when the number of counting intervals grows proportionally to the network size, the minimum budget becomes proportional to the number of transmitters times its logarithm.
Dev: So, GOC helps tame that quadratic growth by introducing those extra parameters, but it still has a dependency on network size scaling.
Taro: The paper also mentions that for fixed symbol duration and error target, even when you observe all arrival times, the minimum budget is still quadratic in network size.
Rosa: That’s a key caveat; observing every particle arrival time doesn't change the fundamental quadratic lower bound under those specific conditions.
Dev: So, the benefit of GOC isn't eliminating the quadratic scaling entirely when duration and intervals are fixed, but it changes the scaling behavior if you let those things grow with network size.
Taro: That means for a massive system, you have to trade off between symbol length and how often you count particles.
Rosa: Exactly; the numerical results show that GOC does provide achievable budget savings because the selected group count actually depends on the symbol duration used in the scheme.
Dev: It’s interesting that they found those achievable savings even with this complex structure, which is a bit surprising given how tight those lower bounds are.
Taro: We should remember that A Particle-Based Multiuser Communication with a Budget Limit explores resource constraints in communication networks, which informs our overall system design.
Rosa: That's right; it reminds us that resource management, like particle budget here or beam selection here, is always a trade-off against error probability.
Lucky paper: 2607.26426: Rosa: Alright, let's look at paper number three of five from this session, it’s Decoder-Guided Lossy Contour Coding Via Anchor Refinement from Elmore Family School of Electrical and Computer Engineering at Purdue University.
Dev: This paper deals with using object contours as compact structural priors for receiver-side vision tasks, like image super-resolution or edge-conditioned generation.
Rosa: So the core asymmetry is that the sender transmits a high-quality contour, but the decoder only gets a free coarse version of it, which this framework models as a structured geometric refinement.
Dev: The methodology involves extracting ordered anchors from that high quality contour and transmitting only their local refinements along the coarse chain.
Rosa: That process has several key steps, starting with extracting the high quality contour C H and then downsampling it into a coarse occupancy map to simulate the decoder side extraction.
Dev: Then they form a coarse block B by keeping just the outer boundary of that active block region and tracing it into an ordered block chain representing the coarse contour C L.
Rosa: After that, they select an anchor point at Bt which is defined as the fine contour pixel closest to the centroid of S(Bt), which are those fine contour pixels falling inside that specific block.
Dev: They also apply dynamic anchor skipping based on a user controlled threshold tau, removing an anchor if its local approximation error d(ai, ai minus i ai plus i) is less than or equal to tau.
Rosa: When it comes to coding and rate distortion control, they transmit three syntax elements: first the number of retained anchors M, second the differences between the original coarse chain indices delta i one through delta i M, and third for each retained anchor its local position l k inside its corresponding s by s coarse block.
Dev: That explicit syntax element transmission is what allows for precise rate distortion control over those structural refinements.
Rosa: They compare this method against two baselines: the HQ-only anchor coding baseline which transmits the absolute first anchor using its absolute image coordinate, and then subsequent anchors via a relative coordinate chain.
Dev: The proposed decoder guided codec only transmits those sparse local refinements along the shared coarse contour order instead of transmitting the full set of anchors.
Rosa: The experimental results on SportsSloMo show that this proposed method substantially reduces bitrate compared to both the HQ-only anchor coding and JBIG2 while maintaining high contour accuracy.
Dev: They report significant bitrate reductions, specifically achieving a fifty-four point five percent to sixty-six point nine percent reduction over methods without decoder side guidance.
Rosa: For example, at tau equals one point zero, the proposed codec reduces bitrate from an initial value of zero point three nine eight six bits per pixel per second to just about zero point two four four eight bits per pixel per second compared to JBIG2.
Dev: That direct comparison against JBIG2 shows a clear coding gain provided by having that decoder available coarse contour information.
Rosa: Ablation studies confirm that both the coarse block size and the entropy coding contribute positively to performance, though a smaller block size s equals three lowers the bitrate by about six point one percent compared to s equals four.
Dev: However, they also noted that a smaller block size is more sensitive to aggressive skipping, which leads to faster IoU degradation at large tau values.
Rosa: Rice coding provides the dominant gain because those retained index differences tend to concentrate near small positive values in the sequence.
Dev: The full entropy coding scheme further reduces bits per pixel per second by up to thirty-nine point five percent and by thirty-three point seven percent on average over fixed length syntax.
Rosa: Future work plans are to integrate this proposed compressed-contour representation into end-to-end super resolution and vision synthesis frameworks.
Dev: So, the paper shows how using decoder side information can drastically cut bitrate while preserving image structure fidelity in vision tasks.
Lucky paper: 2611.07562: Rosa: We're looking at FUSEMOS today, which is about perceptual evaluation of text-to-music generation using a dual-encoder fusion framework with ranking aware loss.
Dev: That paper by Jing Yang and Haoyu Wang proposes combining representations from CLAP and MERT to capture both high level semantic alignment and detailed musical characteristics.
Rosa: So it uses the CLAP branch to produce two L2-normalized embeddings, E Text CLAP and E Music CLAP, which live in a common semantic space for direct comparison via dot product.
Dev: And then they feed the music embedding into a dedicated MLP for the coarse overall musical impression estimation, while the MERT branch extracts hidden states from the final layer of MERT to get EMERT.
Rosa: They then concatenate these features to form a joint feature EMC and pass them through dedicated MLPs to produce intermediate estimations of OMI and TA scores.
Dev: The late fusion strategy they use combines predictions from both encoders at the score level, not fusing raw features early in the network.
Rosa: They use a ranking aware composite loss function defined as Lsum equals beta times Lcon plus gamma times Ltrun, where Lcon models relative quality ranking relationships and truncates regression loss to prevent overfitting.
Dev: That means they are training the model to understand the relative quality of musical samples while keeping things stable by ignoring errors below a certain threshold tau.
Rosa: The experimental validation on the MusicEval benchmark shows that FUSEMOS outperforms existing SOTA methods in both mean squared error and rank correlation metrics by a large margin.
Dev: Specifically, they reduce OMI, U-MSE, and SMSE by fifty-three point six percent and eighty-seven point six percent, with U-SRCC reaching zero point eight one one for OMI and S-SRCC of zero point nine seven seven for TA scores on the benchmark.
Rosa: That is a substantial improvement in the metrics they are trying to optimize, but I wonder if that large margin holds up when we move to more complex music styles or longer compositions.
Dev: It certainly shows strong performance on the benchmark, but those benchmarks are often limited, and real-world music evaluation might require different metrics.
Rosa: Taro, from a representation learning standpoint, how does the combination of CLAP and MERT change how we think about fusing different modalities?
Taro: The CLAP branch provides a strong semantic alignment space that lets us directly compare text and music features using dot product before they even hit the main prediction path.
Dev: So it’s leveraging pre-trained knowledge to create a shared ground for comparison, which is very powerful for capturing high level meaning.
Rosa: That leads me to think about the training process; since the encoders are trained independently, how do you ensure that the joint feature EMC isn't just a noisy concatenation?
Taro: They use dedicated MLPs after concatenation specifically to adapt those features into intermediate estimations of OMI and TA scores before they get fused at the score level.
Dev: So they are not relying on a single massive fusion layer early on, but rather separate pathways leading to a lightweight linear fusion head.
Rosa: And regarding the loss function, what does that ranking aware composite loss actually do for the model during training?
Taro: It explicitly models the relative quality of samples using Lcon, which constrains consistency between predicted score differences and true score differences of sample pairs.
Dev: That’s smart; it forces the model to learn preferences, not just absolute values, which is key for perceptual evaluation.
Rosa: And what about the truncation part of that loss function? How does ignoring errors below tau help keep the training stable?
Taro: The truncated regression loss prevents the model from overfitting to noisy labels or exact rating values by simply ignoring errors that are below a certain threshold tau.
Dev: So they get the benefits of learning relative quality without getting stuck trying to perfectly predict every single noisy label.
Rosa: That sounds like a very practical approach for handling the inherent noise in human judgment data.
Lucky paper: 2611.07311: Rosa: We're diving into A Novel BVH-accelerated Dynamic Ray-Tracing Channel Model for Vehicular Communication Systems today.
Dev: This paper tackles the bottleneck of dynamic ray tracing which usually relies on time-driven updates, and this one proposes a shift to a spatial-driven update mechanism.
Rosa: The core idea is using a bounding volume hierarchy, or BVH, as an environment pre-processing method to guide how often we update the ray information.
Dev: It links the channel coherence time, which is an intrinsic physical characteristic of the wireless channel, to a setting parameter called Tcs.
Rosa: That relationship gives us two main ways to accelerate dynamic ray tracing: shrinking extrapolation intervals and adaptive interval adjustment based on how long a dynamic object takes to cross a bounding box.
Dev: The paper explains that when small scatterers force the coherence time to become small, the coherence time is approximately Tcs equals Ls over v.
Rosa: And when we have dynamic objects, the BVH framework lets it adaptively adjust that extrapolation interval based on the time it takes for a dynamic object to cross a specific bounding box.
Dev: The SIBVH structure is developed as this key pre-processing component, extracting the size of bounding volumes by checking intersections along the trajectory of transceivers.
Rosa: It processes static and dynamic objects separately using different leaf node thresholds and partitioning strategies within that SIBVH structure.
Dev: They also improved the SAH cost function by considering both the surface area and spatial distance to reduce unnecessary intersection tests.
Rosa: The simplified cost function for this SIBVH is η of beta equals alpha times the sum of N one plus N two times beta plus four k m beta squared, minus alpha times d squared, where d is the spatial median plane.
Dev: That cost function guides the selection of a reduced candidate interval determined by that spatial median plane and the equivalent-element plane.
Rosa: The validation in vehicle-to-vehicle scenarios shows that BVH-DRT achieves spatialtemporal consistency with measurements and commercial software, while significantly enhancing computational efficiency.
Dev: Specifically, the results validate that ODDM can enhance robustness and strict orthogonality over OTFS in 6G high-dynamic vehicular communications.
Rosa: And the proposed BVH-DRT achieves a remarkable computational time reduction of seventy-five point six percent compared to previous methods.
Dev: That efficiency gain is significant when we consider the real-time demands of high dynamic channels.
More episodes
- Daily Summary for 2026-11-19
- 2111.07910-Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- 2203.02149-HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- 2205.10102-Degradation-Aware Unfolding Half-Shuffle Transformer for Spectral Compressive Imaging
- 2311.10959-Structure-Aware Sparse-View X-ray 3D Reconstruction
- 2403.04116-Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis
- 2410.02572-Combining Pre- and Post-Demosaicking Noise Removal for RAW Video
- 2604.17453-Learned Nonlocal Feature Matching and Filtering for RAW Image Denoising
- 2603.20387-End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
- 2507.20825-Chirp-Permuted AFDM: A Versatile and Multifunctional Waveform for 6G