Daily Summary for 2026-10-23

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: It's the twenty-third of October, twenty twenty-six, and this is the day's research.

Dev: 7 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 Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation.

The day's research: Rosa: We focus on making LLMs useful when resources are limited because cloud services struggle with low latency applications.

Dev: That points toward large language models over networks under resource constraints.

Taro: It looks at collaborative intelligence where several independent models talk to solve a task.

Rosa: The aim is better quality despite limitations in computation, memory, and communication resources.

Dev: How does that collaboration specifically improve performance under those constraints?

Taro: It relies on them talking to each other using natural language or structured messages.

Rosa: That addresses the computational limits by distributing the workload across models.

Dev: Agreed. Then we have RF-GPT, which connects language models to the physical world.

Taro: RF-GPT uses visual encoders from multimodal LLMs to process radio-frequency spectrograms directly.

Rosa: This lets the model create answers grounded in RF signals, unlike current text model usage for telecommunications.

Rosa: Another important area is dealing with bias in health applications, what was found?

Dev: Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation evaluated how wearable human activity recognition models detect freezing of gait.

Taro: What did that research find regarding those detection models?

Rosa: It found substantial bias across different patient groups and freezing of gait phenotypes.

Dev: And conventional methods for fixing this bias did not work at all.

Taro: That means the existing techniques are ineffective against those disparities.

Rosa: Agreed, the results showed a significant issue with conventional fixes.

Dev: On another note, improving image transmission over noisy wireless channels is being addressed.

Taro: What specific proposal addresses that quality issue?

Rosa: High Perceptual Quality Wireless Image Delivery with Denoising Diffusion Models proposes a system where a deep joint source-channel coding encoder sends a lower resolution version of an image.

Dev: The receiver then uses a denoising diffusion probabilistic model to enhance perceptual quality.

Taro: So the improvement happens at the reception stage using that model?

Rosa: Yes, it enhances perceptual quality by improving that lower resolution version.

Rosa: The study shows how image quality can be improved during wireless transmission by using a denoising diffusion probabilistic model at the receiver.

Dev: That model refines the null space contents to achieve better perceptual quality than standard methods.

Taro: It uses a conventional DeepJSCC encoder to transmit a lower resolution version of an image first.

Rosa: So the scheme involves transmitting low resolution then refining it later with diffusion models?

Dev: Yes, that approach yields better perceptual quality overall. It outperforms standard methods in this context.

Taro: This refinement process relies on the generative model at the receiver to enhance the compressed image data.

Rosa: We need to see how this compares to existing compression techniques for wireless channels.

Dev: The work on Distributed Deep Joint Source-Channel Coding addresses joint compression and transmission over multiple access channels.

Taro: That paper proposes joint compression specifically for images over these types of channels.

Rosa: Regarding other areas, what about the freezing of gait detection research we reviewed?

Dev: Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation found unfairness across patient groups.

Taro: That finding suggests transfer learning could potentially improve fairness in those models.

Rosa: How does that relate to the image quality improvements we discussed?

Dev: They are distinct areas, but both involve using advanced models for different tasks.

Taro: We also looked at combining diffusion MRI with clinical data for stroke outcome prediction.

Rosa: That combination offers a better prediction of stroke recovery outcomes compared to using either alone.

Dev: Combining diffusion MRI with clinical data provides a superior prediction of stroke recovery outcomes because it incorporates more information.

Taro: We also explored geometry estimation for colonoscopy using foundation model adaptation for colonoscopy images.

Rosa: That method adapts general 3D geometry models to accurately estimate the shape of those images in challenging settings.

Dev: So, we have work on image quality, bias mitigation, and structural estimation across these studies.

Taro: These papers cover distinct applications ranging from signal processing to medical imaging analysis.

Rosa: Today's lucky papers are Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation for finding unfairness in gait detection models, High Perceptual Quality Wireless Image Delivery with Denoising Diffusion Models for improving wireless image quality, and RF-GPT Teaching AI to See the Wireless World because it uses visual encoders from large language models to understand radio frequency signals like spectrograms.

Taro: That is all for today.

Rosa: Indeed.

Lucky paper: 2502.09626: Rosa: Let's get into Evidence for Phenotype-Driven Disparities in Freezing of Gait Detection and Approaches to Bias Mitigation.

Dev: The core issue here is that freezing of gait, or FOG, is a symptom of Parkinson’s disease, but clinicians struggle to capture it during short assessments.

Taro: That phenotype specificity means different types of FOG might need different clinical management, which makes the systematic evaluation really important for introducing errors.

Rosa: The paper looks at five state-of-the-art human activity recognition models and finds that even as model performance gets better with more architectural complexity, they still have suboptimal fairness across clinical and demographic attributes.

Taro: So, increased model performance didn't automatically lead to improved fairness metrics?

Dev: Exactly. The study quantified this by saying no single model reached the zero point eight four-fifths rule threshold for any of the fairness metrics they defined, like Demographic parity or Equalized odds.

Rosa: It seems that standard training procedures introduce systematic biases that need targeted mitigation rather than just optimizing thresholds or using adversarial debiasing alone.

Taro: The paper tests four mitigation approaches: conventional methods, and then two transfer learning based ones, which showed the most promise for reducing disparities.

Dev: Specifically, multi-site transfer learning increased F1-scores by zero point zero two four for Model A2 and zero point zero one seven for Model A3, and it improved Demographic parity by zero point zero three seven and Equalized odds by zero point zero four five for both architectures.

Rosa: That’s a concrete improvement in the metrics they used, but did any of those mitigation strategies actually manage to meet the four-fifths rule threshold?

Taro: Only one experiment showed that transfer learning from Model B2 on one dataset could reduce bias to meet that rule, but it came with a substantial performance cost.

Dev: That performance cost is something we need to keep an eye on when we think about deploying these things in a real clinical setting.

Rosa: The paper points out that the FOG phenotype itself is a significant source of bias because akinetic episodes are substantially underrepresented in the datasets used.

Taro: And this underrepresentation directly causes performance disparities when those models are applied clinically, right?

Dev: It does. They conclude that achieving equitable detection requires proactive considerations during data collection, like stratified sampling by phenotype and demographics, instead of relying only on bias mitigation after training.

Rosa: That moves the focus upstream to how the data is gathered, which makes sense when you're trying to build robust systems for Parkinson’s research.

Taro: There is a limitation mentioned where the foundation models weren't trained on subjects with movement disorders, creating a domain gap between their pretraining and the actual task.

Dev: That domain gap is significant because it means what the model learns from general data might not translate perfectly to detecting FOG in Parkinson’s patients.

Rosa: So, while transfer learning helped with fairness metrics, the inherent nature of the pretraining data still left a gap for these foundation models.

Lucky paper: 2309.15889: Rosa: So we're looking at High Perceptual Quality Wireless Image Delivery with Denoising Diffusion Models now, which Imperial College London and Huawei published recently.

Dev: The core idea here is combining deep joint source-channel coding with a denoising diffusion probabilistic model at the receiver for better wireless image quality.

Taro: It breaks down the problem into two stages: an autoencoding stage and a restoration stage, aiming to minimize average distortion between the original image x and the reconstructed image x-hat.

Rosa: They use DeepJSCC as the code block, which they trained to maximize peak signal to noise ratio, defined by that dPSNR metric.

Dev: That dPSNR metric is ten ten A two / C times I n W H x - x hat squared, and they also consider the learned perceptual image patch similarity, LPIPS metric.

Taro: The transmission involves mapping an input image x to a complex-valued latent vector z-tilde, which is then normalized to satisfy a power constraint Pavg before being transmitted over the AWGN channel.

Rosa: The bandwidth ratio rho characterizes the available channel resources as k I n W H channel symbols per pixel, and the SNR is defined by ten ten P avg sigma squared in decibels.

Dev: The restoration stage uses a zero-shot image restoration method called deep denoising null-space model, which guides the DDPM to improve quality while preserving consistency with x-deg.

Taro: They decompose any sample x into range and null space parts using the decomposition x A Ax + (I - A A)x, and then construct a solution for x hat that satisfies the consistency constraint, which is x hat = A x deg + (I - A A)x 0t.

Rosa: The autoencoder network architecture has nearly symmetric encoder and decoder, residual connections, and a computationally efficient attention mechanism.

Dev: They train the encoder and decoder parameters theta and phi to minimize the mean squared error loss L(x, x-deg) = one / C I n W H A x - x deg squared, which helps the DeepJSCC-Degraded model transmit degraded instances with minimal distortion.

Taro: Numerically, their method clearly improves over Inverse-JSCC and DeepJSCC at all evaluated SNR values test in-five five dB and rho values in

zero point zero one three, zero point four: .

Rosa: Figure four shows the comparisons in terms of PSNR and LPIPS metrics, indicating that the proposed method generates more realistic images while achieving lower distortion across all channel conditions.

Dev: It's interesting that their method improves upon DeepJSCC even in terms of the distortion metric, PSNR, which suggests it benefits the encoder/decoder to target a lower resolution version of the image.

Taro: The conclusion is that they introduced a novel generative communication scheme promoting realness through controlled degraded image transmission followed by restoration using DDPM and modified DeepJSCC.

Rosa: They showed their scheme outperforms the standard DeepJSCC and state-of-the-art GAN-based DeepJSCC, although they plan to optimize the specific linear transform in future work.