ICOT 2023

The primary problem is that while the answers that ChatGPT produces have a high rate of being incorrect. How to overcome this serious problem becomes very essential. In addition, ChatGPT announced in Nov. 2022 has a poor human interface. Moreover, most enterprises like to build their own ChatGpt for the reason of privacy & security.
Therefore, in this talk, we will provide our solution & experiences for the above problems.
The outline of this talk is listed below: ⚫ 1. Introduction ⚫ 2. From ChatGPT to ChatGPT-4v ⚫ 3. How to process specific Data for ChatGPT ⚫ 4. How to Read Papers Using ChatGPT-4v effectively ⚫ 5. How to train your own LLM without ChatGPT ⚫ 6. Efficiency & Effectiveness Comparison Between Llama2 & ChatGpt ⚫ 7. Conclusion & Future Works

Understanding human emotions is one of the crucial studies that can be applied to various fields. Recently, various services can be provided through human emotions due to the advancement of AI technologies. Human emotions are mainly recognized through facial expressions. Facial expression recognition using images can cause privacy infringement issues. To overcome this problem, researchers have attempted to analyze emotions from images through the shapes of the face's eyes, nose, mouth, etc. It is called facial landmark detection technology. However, using only facial landmarks has the disadvantage of reducing performance compared to using the entire image.

A new attempt at emotion recognition is using facial images and landmarks simultaneously. Through this, the advantages of both methods can be utilized simultaneously. Recently, context-based emotion understanding techniques that use the environment by looking at objects other than facial expressions in the video are being studied.

In addition to using images, one way to recognize human emotions is to use voices. This talk will cover trends and the latest AI technologies in image-based emotion recognition.

Approximate computing enhances hardware efficiency for error-tolerant applications like deep neural networks (DNNs), making it suitable for DNN accelerators. However, a specialized software framework is crucial to analyze and optimize the impact of approximate arithmetic on both hardware efficiency and accuracy in DNN accelerators. In this talk, a simulation framework supporting various approximate operations on a DNN platform is presented. To expedite the emulation of approximate DNN hardware, the massive parallelism of GPUs is leveraged within the framework. Additionally, a scalable floating-point representation and its application to approximate arithmetic are discussed.

In the era of GPT, speech processing continues to play a pivotal role. With the recent integration of auditory functionalities into ChatGPT, the super assistant now can hear and speak. This natural user interface empowers such a super assistant to seamlessly operate from any location. In this presentation, I will delve into the emerging opportunities and challenges within the field of speech processing. Specifically, I will highlight the significance of speech front-end processing in mitigating interferences and enhancing system robustness. Furthermore, I will explore how the advent of GPT shapes the trajectory of speech processing tasks, especially speech recognition and generation, and address the imperative of safeguarding our speech in light of the imminent zero-shot voice cloning capability.

Electroencephalography (EEG) is the current reference standard for diagnosis of most of the mental and neurological disorders as it is inexpensive, non-invasive and portable compared to other tests. Measuring brain activity through EEG leads to the acquisition of a huge amount of data. In current practice, massive EEG data are visually analyzed to identify and determine abnormalities within the brain how they propagate and function. Analyzing huge volumes of dynamic data in this way is time-consuming, subject to human error, and reduces decision-making reliability. Therefore, there is an ever-increasing need for developing artificial intelligent (AI) techniques that will produce accurate, timely and robust scientific evidence for reliable decision-making in the mental health disorders.

In this talk, we will report some of our recent work on the detection of mental health and neurological disorders such as mild cognitive impairment, schizophrenia, sleep stages problems, and epilepsy using EEG signal data. We aim to develop reliable, robust and efficient analysis techniques and an integrated platform for the above-mentioned mental / neurological disorders from EEG signal data.

Incorporating physical-world objects and information into the virtual world is a key aspect of enhancing immersiveness in the metaverse. However, existing solutions are neither efficient nor scalable when projecting enormous objects due to their reliance on expensive specialized equipment. In this talk, I will introduce our proposed PolyVerse, an edge computing-powered metaverse platform that supports practical Physical-to-Virtual (P2V) projection while preserving low costs, scalability and state consistency. We have developed and evaluated a metaverse campus prototype where real-world pedestrians are projected into the virtual world in real-time.

Computational Intelligence includes three major fields: fuzzy logic, neural networks and evolutionary computing. Swarm intelligence is a branch of evolutionary computing，including ant algorithm, particle swarm algorithm, artificial bee colony algorithm, cat swarm algorithm and quasi-affine transformation evolutionary algorithm (QUATRE). Swarm intelligence has been widely used in various fields that require optimization, such as electronic and electrical engineering, transportation engineering, business management and mechanical processing. Parallel technology is also widely used in scheduling, cloud computing, software engineering, etc. This report will detail some core technologies of swarm intelligence, as well as the concept of parallel technology in the field of optimization, and explain how to combine parallel technology with swarm intelligence.

This talk will present the applications of various machine learning algorithms on the steel production data, to gain insights from past production behaviors/patterns, and to help the industry achieve high production.

In emotion recognition by using physiological signals, the subject’s affective responses are captured by sensors by presenting target stimuli. The emotional responses are inconsistent throughout the acquired signal, rather it arises at a certain duration with high prominence at a distinct part. However, existing studies ignored this vital issue and considered the entire acquired signal for processing leading to inaccurate results. This concern is very important to perform emotion recognition more accurately. This talk will be about how to find such specific parts of the physiological signal having high emotional content and make the emotion recognition systems more efficient. It will be discussed by considering an EEG-based emotion recognition system.

AI-based computer vision for industry refers to the application of artificial intelligence (AI) techniques in the field of computer vision to address various industrial challenges and tasks. Computer vision is a field of AI that trains computers to interpret and understand the visual world. By using digital images and deep learning models, machines can accurately identify and classify objects, and then react to what they "see." In this presentation, we introduce AI-based computer vision technology being applied to industry and share the results.