Chih-Fan Hsu

Kuan-Yu Chen, Yi-Cheng Lin, Po-Chung Hsieh et al.

Current bias evaluations in Instruction Text-to-Speech (ITTS) often rely on univariate testing, overlooking the compositional structure of social cues. In this work, we investigate gender bias by modeling prompts as combinations of Social Status, Career stereotypes, and Persona descriptors. Analyzing open-source ITTS models, we uncover systematic interaction effects where social dimensions modulate one another, creating complex bias patterns missed by univariate baselines. Crucially, our findings indicate that these biases extend beyond surface-level artifacts, demonstrating strong associations with the semantic priors of pre-trained text encoders and the skewed distributions inherent in training data. We further demonstrate that generic diversity prompting is insufficient to override these entrenched patterns, underscoring the need for compositional analysis to diagnose latent risks in generative speech.

Chih-Fan Hsu, Ming-Ching Chang, Wei-Chao Chen

The absence of training data and their distribution changes in federated learning (FL) can significantly undermine model performance, especially in cross-silo scenarios. To address this challenge, we introduce the Federated Learning with Data Digest (FedDig) framework. FedDig manages unexpected distribution changes using a novel privacy-controllable data digest representation. This framework allows FL users to adjust the protection levels of the digest by manipulating hyperparameters that control the mixing of multiple low-dimensional features and applying differential privacy perturbation to these mixed features. Evaluation of FedDig across four diverse public datasets shows that it consistently outperforms five baseline algorithms by substantial margins in various data absence scenarios. We also thoroughly explored FedDig's hyperparameters, demonstrating its adaptability. Notably, the FedDig plugin design is inherently extensible and compatible with existing FL algorithms.

Po-Hsuan Huang, Chia-Ching Lin, Chih-Fan Hsu et al.

Learning from noisy-labeled data is crucial for real-world applications. Traditional Noisy-Label Learning (NLL) methods categorize training data into clean and noisy sets based on the loss distribution of training samples. However, they often neglect that clean samples, especially those with intricate visual patterns, may also yield substantial losses. This oversight is particularly significant in datasets with Instance-Dependent Noise (IDN), where mislabeling probabilities correlate with visual appearance. Our approach explicitly distinguishes between clean vs.noisy and easy vs. hard samples. We identify training samples with small losses, assuming they have simple patterns and correct labels. Utilizing these easy samples, we hallucinate multiple anchors to select hard samples for label correction. Corrected hard samples, along with the easy samples, are used as labeled data in subsequent semi-supervised training. Experiments on synthetic and real-world IDN datasets demonstrate the superior performance of our method over other state-of-the-art NLL methods.

Jeng-Lin Li, Chih-Fan Hsu, Ming-Ching Chang et al.

Recent artificial intelligence (AI) technologies show remarkable evolution in various academic fields and industries. However, in the real world, dynamic data lead to principal challenges for deploying AI models. An unexpected data change brings about severe performance degradation in AI models. We identify two major related research fields, domain shift and concept drift according to the setting of the data change. Although these two popular research fields aim to solve distribution shift and non-stationary data stream problems, the underlying properties remain similar which also encourages similar technical approaches. In this review, we regroup domain shift and concept drift into a single research problem, namely the data change problem, with a systematic overview of state-of-the-art methods in the two research fields. We propose a three-phase problem categorization scheme to link the key ideas in the two technical fields. We thus provide a novel scope for researchers to explore contemporary technical strategies, learn industrial applications, and identify future directions for addressing data change challenges.

Chih-Fan Hsu, Ming-Ching Chang, Wei-Chao Chen

Continual learning (CL) seeks to mitigate catastrophic forgetting when models are trained with sequential tasks. A common approach, experience replay (ER), stores past exemplars but only sparsely approximates the data distribution, yielding fragile and oversimplified decision boundaries. We address this limitation by introducing synthetic boundary data (SBD), generated via differential privacy: inspired noise into latent features to create boundary-adjacent representations that implicitly regularize decision boundaries. Building on this idea, we propose Experience Blending (EB), a framework that jointly trains on exemplars and SBD through a dual-model aggregation strategy. EB has two components: (1) latent-space noise injection to synthesize boundary data, and (2) end-to-end training that jointly leverages exemplars and SBD. Unlike standard experience replay, SBD enriches the feature space near decision boundaries, leading to more stable and robust continual learning. Extensive experiments on CIFAR-10, CIFAR-100, and Tiny ImageNet demonstrate consistent accuracy improvements of 10%, 6%, and 13%, respectively, over strong baselines.

Jyun-Ze Tang, Chih-Fan Hsu, Jeng-Lin Li et al.

Flow matching and diffusion models have shown impressive results in text-to-image generation, producing photorealistic images through an iterative denoising process. A common strategy to speed up synthesis is to perform early denoising at lower resolutions. However, traditional methods that downscale and upscale in pixel space often introduce artifacts and distortions. These issues arise when the upscaled images are re-encoded into the latent space, leading to degraded final image quality. To address this, we propose {\bf Latent Space Scaling Generation (LSSGen)}, a framework that performs resolution scaling directly in the latent space using a lightweight latent upsampler. Without altering the Transformer or U-Net architecture, LSSGen improves both efficiency and visual quality while supporting flexible multi-resolution generation. Our comprehensive evaluation covering text-image alignment and perceptual quality shows that LSSGen significantly outperforms conventional scaling approaches. When generating $1024^2$ images at similar speeds, it achieves up to 246\% TOPIQ score improvement.

Chih-Fan Hsu, Chia-Ching Lin, Ting-Yang Hung et al.

Facial landmark detection has been studied over decades. Numerous neural network (NN)-based approaches have been proposed for detecting landmarks, especially the convolutional neural network (CNN)-based approaches. In general, CNN-based approaches can be divided into regression and heatmap approaches. However, no research systematically studies the characteristics of different approaches. In this paper, we investigate both CNN-based approaches, generalize their advantages and disadvantages, and introduce a variation of the heatmap approach, a pixel-wise classification (PWC) model. To the best of our knowledge, using the PWC model to detect facial landmarks have not been comprehensively studied. We further design a hybrid loss function and a discrimination network for strengthening the landmarks' interrelationship implied in the PWC model to improve the detection accuracy without modifying the original model architecture. Six common facial landmark datasets, AFW, Helen, LFPW, 300-W, IBUG, and COFW are adopted to train or evaluate our model. A comprehensive evaluation is conducted and the result shows that the proposed model outperforms other models in all tested datasets.