Abstracts Track 2024

Area 1 - Theory and Methods
Area 2 - Applications

Area 1 - Theory and Methods

Nr: 92
Title:

Deep Learning-Powered Infrared Thermography for Rapid Diagnosis in Emergency Care Unit
Authors:

Akam Petersen, Mikkel Brabrand and Sergey Kucheryavski
Abstract: Infrared thermography (IRT) has emerged as an affordable, rapid and non-invasive complement to widely adopted yet resource demanding medical imaging techniques such as MRI, CT-scan and X-ray, offering diverse applications in medical field. While the IRT hardware is well established and capable of providing high-quality thermographic images, the analysis of such images often requires well-trained experts. Current state-of-the art methods for computer-aided IRT analysis rely on statistical tests of temperature gradients between control points, which is suboptimal as it does not fully exploit the available information regarding spatial temperature distribution. This paper addresses this issue by incorporating artificial neural networks (ANN) into the IRT analysis workflow. We focus on a particular case, where IRT was utilized in the emergency department (ED) for predicting 30-day mortality, thereby contributing to improved diagnosis and patient care in emergency medicine. IRT images of 626 patients were analysed in total. Various ANN based approaches were considered in this study, the best results were obtained using an anomaly detection model based on Variational Autoencoder (VAE), achieving an accuracy of 75.8% and an F1 score of 71.4%. The use of VAE models offer an additional advantage as they can visualize the sources of abnormality, thereby providing supplementary diagnostic information. The paper comprehensively presents all analysis details as well as recommendations regarding image pre-processing, augmentation and potential enhancements of the models.
Nr: 162
Title:

A Norm-Contrastive Representation Learning for Near and Far OOD Detection
Authors:

Takeshi Konoike and Shin Ando
Abstract: Out-of-Distribution (OOD) detection is an important problem for real-world applications of machine learning algorithms. One categorization of OOD detection method can be made by the input data for learning representation. The first category trains exclusively on in-distribution (ID) data, which is the primary input for training an OOD detection model. The second category additionally exploits augmented or generated data, which can aquire enriched representation, with auxiliary features not present in in-distribution data. Intuitively, it is benefitial for detecting near OODs, but we found it can detrimentally affect the performances on far OOD benchmarks, on which high performances are achieve without such a practice. In this paper, we propose a norm-based scoring function and a contrastive representation learning for enhancing the proposed score for near OOD detection. Furthermore. we propose an ensemble score to take advantage of the proposed model and an ID trained models with superior far OOD detection performance. Our empirical study using a collection of image benchmarks shows the advantage of the proposed ensemble score over state-of-the-arts for both near- and far-OOD detection benchmarks.
Nr: 163
Title:

Behaviour Monitoring and Assignment of Water Usage by Monitoring the Water Flow Rate in a Smart Home Environment
Authors:

Christoph Schultes, Justin Baudisch and Thorsten Jungeblut
Abstract: Due to the demographic change, the number of elderly people increases continuously, resulting in more people with a higher risk for chronical illnesses and multimorbidity. Thus, more people need (medical) care which is often provided by relatives and nursing services while the person still lives at home. To help elderly people, smart home technologies can be used to support them in their daily lives. Additionally, these technologies open up the possibility of using artificial intelligence (AI) to monitor behaviour, advice caregivers about the behaviour or detect anomalies to support the resident or call for help if needed. This contribution focuses on the water consumption in a household. With the help of a central water flow sensor, it should be identified which device is used. With this information caregivers get insights of the hygiene behaviour and do not have to ask unpleasant questions regarding the hygiene too frequently. For the research we collaborate with the non-profit association KogniHome e.V. KogniHome develops user-oriented Ambient Assisted Living (AAL) solutions for the demographic change and provides a research apartment where a flow sensor is installed to measure the water usage in the apartment. Each use of a device that consumes water will be detected by the sensor. It logs the liter per minute (l/m) consumption for the duration of the use of the device, which we consider as one event. The key research question is how we can use this data to monitor the user behaviour, predict future events or detect anomalies in the daily live of the resident. To assign events to the different devices, it is important to consider multiple challenges. For one, there can be overlapping water consumptions from different devices (e.g. the refill after a toilet flush and the use of the water tap for hand washing). The probability of this issue increases with multiple persons living in the smart home environment. Because nobody lives in the research apartment, the events must be simulated. Thus, in the dataset, simulated events like showering could last not as long as they would in a real situation. Dependent on the specific activity, the water tap may be opened to different degrees leading to a varying water consumption in the use of devices with adjustable water output. To deal with these challenges we will focus on one person households, but nevertheless, overlapping events should be included in the data. We can generate the dataset by simulating the use of the devices and taking track of when which device has been used to label the events afterwards. To accurately replicate the duration of water consumptions, it needs the knowledge of how long these events take. Additionally to the water flow data, responses from other sensors, like motion sensors, can be logged in the dataset. Is the data labelled and cleaned, a machine learning model can be trained on the dataset to cluster the data or assign an event to a device or to the room in which the device is located. Not only the strength and the duration of the water flow, but also the response of the motion sensors are input parameters for the model, that should be considered to assign the events to different rooms and devices. Depending on the realism of the generated data, it would be interesting whether correlations can be found between devices and rooms and between different devices like the toilet flush and the water tap.
Nr: 164
Title:

Blended Modality-Based Learning Representations for Visible-Infrared Person Re-Identification
Authors:

Soonyong Gwon and Kisung Seo
Abstract: Retrieving and matching individual images for Visible-Infrared Person Re-identification is a challenging task due to the huge modality gap between daytime color and nighttime infrared images. It is a cross-modality problem that involves matching between different modalities. Most of existing approaches rely on modality-shared methods that extract modality-invariant features in a common embedding space. Although these approaches have achieved performance improvement, the models tend to focus only on the shared features, neglecting individual-specific discriminative information. To solve these problems, we propose an integrated approach between Modality-Specific and Modality-Shared method. First, introducing the Customized Modality-Specific Enhanced Module provides enhanced feature maps using attention mechanism between pre-pooling and post-pooling features and reinforces specific features for the modality-shared based network. Second, we introduce the Pseudo Label-oriented Modality-Specific Loss, which provides effective representation learning explicitly to reduce the modality gap using pseudo labels as anchors. We provide a blended Modality model based on ResNet50. We compare our and various existing methods for the mAP and Rank-1 performances on the SYSU-MM01 and RegDB datasets. Experimental results demonstrate that our proposed model outperforms the recent competitive methods.

Area 2 - Applications

Nr: 161
Title:

Fine-Tuning Large Language Models for Personalized Review-Generation
Authors:

Padipat Sitkrongwong and Atsuhiro Takasu
Abstract: User-generated reviews serve as valuable resources for analyzing the user sentiments across various product aspects, enhancing personalized product recommendations. However, due to a scarcity of user feedback, many e-commerce products received insufficient number of reviews, resulting in less effective recommendations. To address this limitation, many works adopt a data augmentation approach by generating synthetic reviews. This involves simulating textual preferences toward unseen products, providing additional information for user preference modeling. With the advancement of large language models (LLMs), generating human-like reviews has become more accessible through prompting strategies that leverage the LLM’s capabilities in reasoning and deep understanding in human natural language. Generally, there are two main approaches to generate reviews with prompting techniques: zero-shot and few-shot prompting. Zero-shot prompting involve designing a suitable prompt for LLMs to generate a synthetic review for a target product, without providing any examples. In contrast, few-shot prompting includes historical review samples from users or products in a prompt, enabling LLMs to better understand keywords and patterns in reviews, yielding more reliable synthetic reviews. Despite these advancements, neither approach incorporate user personal preferences into LLMs, leading to generic synthetic reviews that are not tailored to each individual user’s unique taste. For instance, using few-shot prompting to generate reviews for the movie “Titanic” often yields common keywords related to this movie (such as “romance”, “tragic”, or “James Cameron”), but lacks the user’s unique and sentimental opinions. Utilizing those synthetic reviews may generate indiscriminable user presentations, leading to less-personalized product recommendations. To address this, our objective is to propose a personalized review generation by incorporating user personal preferences to produce user-specific synthetic reviews using LLMs. We aim to fine-tine LLMs with a review-generation task, providing three types of input: review texts, their corresponding ratings, and user unique identifiers (IDs). For instance, to generate a synthetic review of the movie "Titanic," the review texts are sourced from the target user's historical reviews of other movies, preferably those similar to "Titanic," and/or historical reviews of "Titanic" from neighboring users who share similar interests. In addition to the review texts, their corresponding ratings are also incorporated as input. Including both review texts and ratings enhances LLM's reasoning capability by associating keywords and patterns in reviews with preference degrees toward a target product. Finally, to enable LLMs to generate personalized reviews, unique user IDs, specific to each individual user, are included as input. These IDs play a crucial role in capturing individual user sentimental opinions and preferences from the historical reviews and ratings in the fine-tuning process of LLMs. All types of input are mapped with corresponding embeddings and fine-tuned with the review-generation objective, using an actual user review on a target product as the desired output. By integrating user IDs and a fine-tuning strategy, we believe in LLM's ability to generate high-quality, personalized reviews—valuable for downstream tasks like data augmentation.