Yijun Zhao

Akayou A. Kitessa, Yijun Zhao

Vision-language models map visual features into a shared embedding space through learned projection layers, yet it remains unclear how these transformations alter the structure of visual information. This study examines changes in representation through spatial-frequency accessibility, measured by the linear recoverability of band-limited Fourier energy from model representations. To isolate effects beyond dimensionality reduction, we introduce Residual Spectral Loss (RSL), which evaluates changes relative to a dimension-matched random projection baseline. To reduce confounding effects from optimization, the analysis uses pretrained models with all parameters frozen. The experimental results show consistent frequency-dependent changes in accessibility across CLIP and DINOv2 on ImageNet and MS-COCO datasets. Spectral accessibility follows a non-monotonic trajectory across depth, peaking at intermediate layers before decreasing toward the output representation. The final transformation differs across architectures: CLIP's learned projection is spectrally neutral, with changes explained by compression, whereas DINOv2's [CLS] pooling induces a structured loss across the spectrum. These findings identify intermediate layers and pooling mechanisms as primary drivers of spectral transformation in modern vision encoders.

Yunshuo Tian, Akayou Kitessa, Tanuja Chitnis et al.

Large language models (LLMs) and classical machine learning methods offer complementary strengths for predictive modeling, yet their fundamentally different representations and training paradigms hinder effective integration: LLMs rely on gradient-based optimization over textual data, whereas models such as Random Forests (RF) employ non-differentiable feature partitioning. This work introduces a reciprocal co-training framework that couples an LLM with an RF classifier via reinforcement learning, creating an iterative feedback loop in which each model improves using signals from the other. Tabular data are reformulated into standardized textual representations for the LLM, whose embeddings augment the RF feature space, while calibrated RF probability estimates provide feedback signals that guide reinforcement learning updates of the LLM. Experiments across three medical datasets demonstrate consistent performance gains for both models, with particularly strong effects for the LLM. Ablation analyses show that iterative refinement, hybrid reward design, and dimensionality control jointly contribute to these gains. The proposed framework provides a general mechanism that allows incompatible model families to leverage each other's strengths through bidirectional adaptation.

Yao Xiao, Yijun Zhao

Realistic age-progressed photos provide invaluable biometric information in a wide range of applications. In recent years, deep learning-based approaches have made remarkable progress in modeling the aging process of the human face. Nevertheless, it remains a challenging task to generate accurate age-progressed faces from infant or toddler photos. In particular, the lack of visually detectable gender characteristics and the drastic appearance changes in early life contribute to the difficulty of the task. We propose a new deep learning method inspired by the successful Conditional Adversarial Autoencoder (CAAE, 2017) model. In our approach, we extend the CAAE architecture to 1) incorporate gender information, and 2) augment the model's overall architecture with an identity-preserving component based on facial features. We trained our model using the publicly available UTKFace dataset and evaluated our model by simulating up to 100 years of aging on 1,156 male and 1,207 female infant and toddler face photos. Compared to the CAAE approach, our new model demonstrates noticeable visual improvements. Quantitatively, our model exhibits an overall gain of 77.0% (male) and 13.8% (female) in gender fidelity measured by a gender classifier for the simulated photos across the age spectrum. Our model also demonstrates a 22.4% gain in identity preservation measured by a facial recognition neural network.

Yijun Zhao, Jacek Ossowski, Xuming Wang et al.

In-scanner motion degrades the quality of magnetic resonance imaging (MRI) thereby reducing its utility in the detection of clinically relevant abnormalities. We introduce a deep learning-based MRI artifact reduction model (DMAR) to localize and correct head motion artifacts in brain MRI scans. Our approach integrates the latest advances in object detection and noise reduction in Computer Vision. Specifically, DMAR employs a two-stage approach: in the first, degraded regions are detected using the Single Shot Multibox Detector (SSD), and in the second, the artifacts within the found regions are reduced using a convolutional autoencoder (CAE). We further introduce a set of novel data augmentation techniques to address the high dimensionality of MRI images and the scarcity of available data. As a result, our model was trained on a large synthetic dataset of 225,000 images generated from 375 whole brain T1-weighted MRI scans. DMAR visibly reduces image artifacts when applied to both synthetic test images and 55 real-world motion-affected slices from 18 subjects from the multi-center Autism Brain Imaging Data Exchange (ABIDE) study. Quantitatively, depending on the level of degradation, our model achieves a 27.8%-48.1% reduction in RMSE and a 2.88--5.79 dB gain in PSNR on a 5000-sample set of synthetic images. For real-world artifact-affected scans from ABIDE, our model reduced the variance of image voxel intensity within artifact-affected brain regions (p = 0.014).