Seong-Whan Lee

Jung-Ho Hong, Ho-Joong Kim, Kyu-Sung Jeon et al.

The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA.

Jeong-Woo Park, Young-Eun Kim, Seong-Whan Lee

Composed image retrieval (CIR) is a vision language task that retrieves a target image using a reference image and modification text, enabling intuitive specification of desired changes. While effectively fusing visual and textual modalities is crucial, existing methods typically adopt either early or late fusion. Early fusion tends to excessively focus on explicitly mentioned textual details and neglect visual context, whereas late fusion struggles to capture fine-grained semantic alignments between image regions and textual tokens. To address these issues, we propose FAR-Net, a multi-stage fusion framework designed with enhanced semantic alignment and adaptive reconciliation, integrating two complementary modules. The enhanced semantic alignment module (ESAM) employs late fusion with cross-attention to capture fine-grained semantic relationships, while the adaptive reconciliation module (ARM) applies early fusion with uncertainty embeddings to enhance robustness and adaptability. Experiments on CIRR and FashionIQ show consistent performance gains, improving Recall@1 by up to 2.4% and Recall@50 by 1.04% over existing state-of-the-art methods, empirically demonstrating that FAR Net provides a robust and scalable solution to CIR tasks.

Jeong-Woo Park, Seong-Whan Lee

Composed Image Retrieval (CIR) is the task of retrieving a target image from a gallery using a composed query consisting of a reference image and a modification text. Among various CIR approaches, training-free zero-shot methods based on pre-trained models are cost-effective but still face notable limitations. For example, sequential VLM-LLM pipelines process each modality independently, which often results in information loss and limits cross-modal interaction. In contrast, methods based on multimodal large language models (MLLMs) often focus exclusively on applying changes indicated by the text, without fully utilizing the contextual visual information from the reference image. To address these issues, we propose multi-faceted Chain-of-Thought with re-ranking (MCoT-RE), a training-free zero-shot CIR framework. MCoT-RE utilizes multi-faceted Chain-of-Thought to guide the MLLM to balance explicit modifications and contextual visual cues, generating two distinct captions: one focused on modification and the other integrating comprehensive visual-textual context. The first caption is used to filter candidate images. Subsequently, we combine these two captions and the reference image to perform multi-grained re-ranking. This two-stage approach facilitates precise retrieval by aligning with the textual modification instructions while preserving the visual context of the reference image. Through extensive experiments, MCoT-RE achieves state-of-the-art results among training-free methods, yielding improvements of up to 6.24% in Recall@10 on FashionIQ and 8.58% in Recall@1 on CIRR.

Woo-Chan Kim, Ji-Hoon Park, Seong-Whan Lee

Large language models (LLMs) have demonstrated state-of-the-art performance across a wide range of natural language processing tasks. However, high-performing models are typically accessible only via APIs, incurring substantial inference costs. Cascade methods address this by initially employing a cheaper model and escalating to a stronger one only when necessary. Nevertheless, existing cascade approaches struggle to select a reliable representative response and assess the overall reliability of free-form outputs, as they rely on exact text matching. To overcome these limitations, we propose Keyword-inspired Cascade (KiC), a novel framework for cost-efficient free-form text generation. KiC identifies the most representative answer among multiple outputs from a weaker model and evaluates the semantic alignment of other responses with it. Based on the degree of alignment, KiC determines whether to accept the weaker model's output or escalate to a stronger model. Experiments on three free-form text generation benchmarks show that KiC achieves 97.53 percent of GPT-4's accuracy while reducing API costs by 28.81 percent on average, and even outperforms GPT-4 in a specific benchmark.

Ju-Young Oh, Ho-Joong Kim, Seong-Whan Lee

Video question answering (VQA) is a multimodal task that requires the interpretation of a video to answer a given question. Existing VQA methods primarily utilize question and answer (Q&A) pairs to learn the spatio-temporal characteristics of video content. However, these annotations are typically event-centric, which is not enough to capture the broader context of each video. The absence of essential details such as object types, spatial layouts, and descriptive attributes restricts the model to learning only a fragmented scene representation. This issue limits the model's capacity for generalization and higher-level reasoning. In this paper, we propose a fundamental question generation with the integration of question embeddings for video question answering (FIQ), a novel approach designed to strengthen the reasoning ability of the model by enhancing the fundamental understanding of videos. FIQ generates Q&A pairs based on descriptions extracted from videos, enriching the training data with fundamental scene information. Generated Q&A pairs enable the model to understand the primary context, leading to enhanced generalizability and reasoning ability. Furthermore, we incorporate a VQ-CAlign module that assists task-specific question embeddings with visual features, ensuring that essential domain-specific details are preserved to increase the adaptability of downstream tasks. Experiments on SUTD-TrafficQA demonstrate that our FIQ achieves state-of-the-art performance compared to existing baseline methods.

Min-Jeong Lee, Hee-Dong Kim, Seong-Whan Lee

Stable diffusion is an outstanding image generation model for text-to-image, but its time-consuming generation process remains a challenge due to the quadratic complexity of attention operations. Recent token merging methods improve efficiency by reducing the number of tokens during attention operations, but often overlook the characteristics of attention-based image generation models, limiting their effectiveness. In this paper, we propose local representative token guided merging (ReToM), a novel token merging strategy applicable to any attention mechanism in image generation. To merge tokens based on various contextual information, ReToM defines local boundaries as windows within attention inputs and adjusts window sizes. Furthermore, we introduce a representative token, which represents the most representative token per window by computing similarity at a specific timestep and selecting the token with the highest average similarity. This approach preserves the most salient local features while minimizing computational overhead. Experimental results show that ReToM achieves a 6.2% improvement in FID and higher CLIP scores compared to the baseline, while maintaining comparable inference time. We empirically demonstrate that ReToM is effective in balancing visual quality and computational efficiency.

Aung Aung Phyo Wai, Yangsong Zhang, Heng Guo et al.

Steady-state visual evoked potentials (SSVEP) brain-computer interface (BCI) provides reliable responses leading to high accuracy and information throughput. But achieving high accuracy typically requires a relatively long time window of one second or more. Various methods were proposed to improve sub-second response accuracy through subject-specific training and calibration. Substantial performance improvements were achieved with tedious calibration and subject-specific training; resulting in the user's discomfort. So, we propose a training-free method by combining spatial-filtering and temporal alignment (CSTA) to recognize SSVEP responses in sub-second response time. CSTA exploits linear correlation and non-linear similarity between steady-state responses and stimulus templates with complementary fusion to achieve desirable performance improvements. We evaluated the performance of CSTA in terms of accuracy and Information Transfer Rate (ITR) in comparison with both training-based and training-free methods using two SSVEP data-sets. We observed that CSTA achieves the maximum mean accuracy of 97.43$\pm$2.26 % and 85.71$\pm$13.41 % with four-class and forty-class SSVEP data-sets respectively in sub-second response time in offline analysis. CSTA yields significantly higher mean performance (p<0.001) than the training-free method on both data-sets. Compared with training-based methods, CSTA shows 29.33$\pm$19.65 % higher mean accuracy with statistically significant differences in time window less than 0.5 s. In longer time windows, CSTA exhibits either better or comparable performance though not statistically significantly better than training-based methods. We show that the proposed method brings advantages of subject-independent SSVEP classification without requiring training while enabling high target recognition performance in sub-second response time.

Paul Bertens, Seong-Whan Lee

Although Deep Neural Networks have seen great success in recent years through various changes in overall architectures and optimization strategies, their fundamental underlying design remains largely unchanged. Computational neuroscience on the other hand provides more biologically realistic models of neural processing mechanisms, but they are still high level abstractions of the actual experimentally observed behaviour. Here a model is proposed that bridges Neuroscience, Machine Learning and Evolutionary Algorithms to evolve individual soma and synaptic compartment models of neurons in a scalable manner. Instead of attempting to manually derive models for all the observed complexity and diversity in neural processing, we propose an Evolvable Neural Unit (ENU) that can approximate the function of each individual neuron and synapse. We demonstrate that this type of unit can be evolved to mimic Integrate-And-Fire neurons and synaptic Spike-Timing-Dependent Plasticity. Additionally, by constructing a new type of neural network where each synapse and neuron is such an evolvable neural unit, we show it is possible to evolve an agent capable of learning to solve a T-maze environment task. This network independently discovers spiking dynamics and reinforcement type learning rules, opening up a new path towards biologically inspired artificial intelligence.