| viXra.org > Artificial Intelligence > 2403 |
 |
 |
| viXra.org > Artificial Intelligence > 2403 |
|
|
[14] viXra:2403.0140 [pdf] submitted on 2024-03-29 02:30:59
Authors: Mohammadjavad Maheronnaghsh, Mohammad Hossein Rohban
Comments: 7 Pages.
Edge machine learning (Edge ML) offers solutions for deploying ML models directly on resource-constrained edge devices. However, ensuring adversarial robustness remains a challenge. This paper presents an accessible approach for adversarial robust distillation (ARD) based in the limited confines of Google Colab.Our goal is enabling fast yet robust knowledge transfer to student models suited for edge devices. Extensive experiments are conducted distilling from a WideResNet34 teacher to MobileNetV2 student using limited computational resources. The efficacy of ARD is evaluated under settings with only 1 GPU (T4 GPU) and 13GB RAM for up to 6 hours a day.Notably, competitive adversarial robustness is attained using very few gradient attack steps. This improves training efficiency crucial for edge ML. Appropriately balancing hyperparameters also allows robust accuracy over 50% using just 1 attack step. Overall, the presented approach advances the feasibility of performing robust distillation effectively even with accessibility constraints.The democratized and reproducible method on Google Colab serves as a launchpad for those aiming to reap the advantages of edge intelligence. By sharing models protected against adversarial threats, this work propels broader adoption of trustworthy ML at society’s technological edges.
Category: Artificial Intelligence
[13] viXra:2403.0119 [pdf] submitted on 2024-03-25 19:56:36
Authors: Keith D. Foote
Comments: 13 Pages. (Correction made by viXra Admin to conform with the requirements of viXra.org - Future non-compliant submission will not be accepted!)
The concept of AI governance has been developed to promote responsible behavior in the use of artificial intelligence. Artificial intelligence can be used for the betterment of mankind, and has proven itself to be very useful in completing a large number of tasks both quickly and efficiently. Sadly, AI can also be used in support of criminal behavior, ranging from the creation and distribution of misinformation to audio and video impersonations. AI governance can be described as a philosophy developed to minimize the misuse of artificial intelligence for unethical and criminal behavior.
Category: Artificial Intelligence
[12] viXra:2403.0112 [pdf] submitted on 2024-03-22 20:35:03
Authors: Ki Song Kim, UiSong Hwang, SongHak Hong, HyonSok Han, YongChol Jang
Comments: 8 Pages.
Recently, the Quantum Neural Network(QNN) is the newly appeared discipline by combining the quantum computing theory and neural network attracts attention. As a matter of fact, the quantum artificial intelligence is no more than the beginning, however the theoretical research and analysis have already been developed for the quantum associative storage, quantum state superposition and quantum parallel learning, etc, in the quantum computing ranges in the world, so the theoretical basis has been laid for development of the quantum neural computing. In this paper, we described a simulation method of quantum BP neural network constructed with multiple Control-NOT(CNOT) gates in the "Jupyter lab" using python language. This QNN consist of the multiple CNOT gates and phase control gates, and is emulated with the sequence quantum steps in the emulator. In this work, we simulated this QNN using MNIST database, and have got the same results in accuracy as the classical neural network.
Category: Artificial Intelligence
[11] viXra:2403.0107 [pdf] submitted on 2024-03-22 14:35:57
Authors: Yemi Adetuwo
Comments: 18 Pages.
As organizations increasingly adopt cloud services for storing and processing sensitive data, the need for robust cloud security threat detection mechanisms becomes paramount. This research paper explores the application of large language models (LLMs) in the context of cloud security threat detection. Building upon the growing demand for robust cybersecurity measures in cloud environments, this study investigates the use-cases and practical implications of integrating LLMs to support threat detection capabilities. Log analysis, natural language processing (NLP) for security alerts, threat intelligence analysis, and social engineering detection were identified as key areas where LLMs can significantly enhance cloud security threat detection. While acknowledging the potential of LLMs to enhance threat detection, this paperemphasizes their role as complementary tools toexisting techniques, such as cloud SOC (securityoperations center), anomaly detection, networkmonitoring, and user behaviour analytics.Considerations pertaining to ethics, data privacy, and transparency are also discussed to ensure responsible deployment and usage of LLMs in cybersecurity.Through an extensive review of relevant literature,providing practical examples, and offering expertanalysis, this research paper not only sheds light on the potential of LLMs for cloud security threat detection but also delivers actionable recommendations for practitioners and organizations seeking to integrate LLMs effectively into their existing security infrastructure. The findings presented in this study contribute to the advancement of AI-driven cybersecurity and lay the groundwork for further research and development in this critical domain.
Category: Artificial Intelligence
[10] viXra:2403.0105 [pdf] submitted on 2024-03-22 20:46:45
Authors: Eliza Kosloff
Comments: 3 Pages.
The recent success of large language models (LLMs) in artificial intelligence has drawn significant attention from the machine learning community. However, the theoretical foundations of these models remain poorly understood. In this paper, we explore the deep connections between LLMs and spin glass theory, a well-established framework in statistical physics. We show how key concepts from spin glasses, such as frustration, random interactions, and phase transitions, can provide a powerful lens for understanding the behavior of LLMs. We argue that this interdisciplinary perspective can facilitate knowledge transfer between the machine learning and physics communities, leading to novel insights and algorithmic improvements.
Category: Artificial Intelligence
[9] viXra:2403.0103 [pdf] submitted on 2024-03-21 02:28:00
Authors: Xiangjun Mi, Chongru Huang, Bingyi Kang
Comments: 15 Pages.
In fuzzy systems, how to represent uncertainty is a crucial research topic. Negation is an inherent characteristic of knowledge, and it provides a brand-new perspective of solving problems from the opposite of the events. Intuitionistic fuzzy sets (IFSs), as a generalization of the fuzzy sets, have the ability to better express fuzzy information. However, since the existing methods have not completely broken through the constraints of the first (classical) negation and inconsistent calculation standards, IFSs still have limitations in expressing uncertainty. To address this issue, and strengthen the performance of fuzzy systems to represent uncertain information, this paper proposed a novel method to obtain the negation of the IFS from the perspective of maximum entropy. Some desired theorems and properties are investigated to denote the nature of the negative IFS. Moreover, entropy is used to describe the connection between the IFS and uncertainty in the negation process. Futhermore, based on the negation, this paper designed a new approach to measure the uncertainty of the IFS. Then, a new pattern classifi- cation algorithm is developed. Finally, the practical applications show the effectiveness of the negation method.
Category: Artificial Intelligence
[8] viXra:2403.0102 [pdf] submitted on 2024-03-21 02:31:57
Authors: Xiangjun Mi, Chongru Huang, Bingyi Kang
Comments: 11 Pages.
How to obtain negation knowledge is a crucial topic, especially in the field of artificial intelligence. Limited work has been done on the negation of a probability distribution, which has been studied in depth throughout the literature. However, the aspect of the intensity level of negation enforcement has not yet been investigated. Moreover, let us note that the main characteristic of intelligent systems is just the flexibility for the sake of being able to represent knowledge according to each situation. In general, researchers have a tendency to express the need for cognitive range in the negation. Thus, it would seem very useful to find a wide range of negations under intensity levels in a probability distribution. Based on these ideas, this paper first proposes a new approach of finding a probability distribution negation and gives a domain of intensity in which the negation is executed, which is called the negation space. Then, we investigate a number of desirable properties and explore their correlation with entropy. Numerical examples show the characteristics of the proposed negation solution. Finally, we validate the efficiency of the proposed method from the point of view of the Dempster- Shafer belief structure.
Category: Artificial Intelligence
[7] viXra:2403.0101 [pdf] submitted on 2024-03-21 02:44:05
Authors: Xiangjun Mi, Ye Tian, Bingyi Kang
Comments: 40 Pages.
Information fusion is an important topic in scientific research. Soft likelihood function is a common method of fusing evidence from multiple sources. However, when the combined evidence contains equally important decision information, the fusion results obtained using existing methods do not reflect the attitudinal characteristics of decision makers. To address this problem, a novel generalised soft likelihood function is developed in this paper. First, a new notion of decision maker (DM) pair is defined, which is used to char- acterise the outcome of the decision as well as the reliability of the evidence. Then, a series of algorithms for correcting the initial evidence set data are formulated. Eventually, a generic soft likelihood function for fusing com- patible evidence information is proposed. Numerical examples are used to illustrate the effectiveness of the proposed methodology.
Category: Artificial Intelligence
[6] viXra:2403.0100 [pdf] submitted on 2024-03-21 02:46:50
Authors: Xiangjun Mi, Pengdan Zhang, Bingyi Kang
Comments: 24 Pages.
In real criminal cases, the decision outcome is often influenced by many complex factors, such as the importance of initial evidence and the prioritization of evidence. How to model these information in an integrated manner to provide technical tools for case detection so as to find the real suspect is of great importance for social security and stability. To address the above issues, this paper proposes a novel soft likelihood function based on the Decision Making Trial and Evaluation Laboratory (DEMATEL) method. Firstly, the proposed method well preserves the preference of decision-maker (DM) in the soft likelihood function proposed by Yager et al. Secondly, the method takes into account the modeling of associated information. In addition, it also extends the soft likelihood function to reflect the preferences of DMs through the importance of evidence. Finally, based on these designed algorithms, a decision processing model for criminal cases is constructed, which systematically provides a guiding process for case detection. Numerical examples and applications show the practicality as well as effectiveness of the proposed method.
Category: Artificial Intelligence
[5] viXra:2403.0094 [pdf] submitted on 2024-03-19 19:47:25
Authors: Budee U. Zaman
Comments: 15 Pages.
Who dominates the destiny of the world, humans or artificial intelligence (AI)? This question strikes at the very heart of contemporaryhumanity’s existential anxieties about its future. If we want to seriouslyconsider whether or not unfriendly AI ‘neurons’ pose any threat to humancivilisation and humanity’s continual existence and evolution in the Universe, we need to know as much as possible about the Universe in whichwe find ourselves, our place in it, and what cognition, consciousness andmentality really are.How might we combine philosophical, cognitive science and technological perspectives, to explore the evolving relationship between humansand AI, in order to engage and address the questions at the core of thishuman-AI complex, namely the future of civilisation — what will it looklike, who can claim to be our successors, towards what goals and ends?The evolution and development of human cognition as well as the emergence of AI can help us define these potential paths of future development.Where do we stand today, in relation to our own history and developmentand to the possibilities that artificial intelligence can offer us? The essayexplores the ethical, social and existential questions that arise from theincreasing automation of artificial intelligence and how it relates to thestory of humanity, from its origins to its contemporary cultural expression.
Category: Artificial Intelligence
[4] viXra:2403.0063 [pdf] submitted on 2024-03-14 02:09:56
Authors: Philip Naveen
Comments: 6 Pages.
A learning rate scheduler is a predefined set of instructions for varying search stepsizes during model training processes. This paper introduces a new logarithmic method using harsh restarting of step sizes through stochastic gradient descent. Cyclical log annealing implements the restart pattern more aggressively to maybe allow the usage of more greedy algorithms on the online convex optimization framework. The algorithm was tested on the CIFAR-10 image datasets, and seemed to perform analogously with cosine annealing on large transformer-enhanced residual neural networks. Future experiments would involve testing the scheduler in generative adversarial networks and finding the best parameters for the scheduler with more experiments.
Category: Artificial Intelligence
[3] viXra:2403.0060 [pdf] submitted on 2024-03-14 21:08:03
Authors: J. G. Wolff
Comments: 143 Pages.
As the title of this book suggests, it is about how intelligence may be understood as information compression (IC). More specifically, the book is about the {em SP Theory of Intelligenc} (SPTI) and its realisation in the {em SP Computer Model}---and their potential applications, benefits, and associated ideas. The SPTI draws on substantial evidence for the importance of IC in human learning, perception, and cognition. Since the SPTI also has much to say about issues in artificial intelligence (AI), it is a theory of both natural and artificial intelligence. In the SPTI, IC is achieved largely via the powerful concept of {em SP-Multiple-Alignment}, a major discovery which is largely responsible for the versatility of the SPTI in aspects of human intelligence and beyond. Strengths of the SPTI include: the modelling of several kinds of intelligent behaviour, including several kinds of probabilistic reasoning; the representation and processing of several kinds of intelligence-related knowledge; and the seamless integration of diverse aspects of intelligence, and diverse kinds of knowledge, in any combination. That seamless integration appears to be {em essential} in any AI system that aspires to the fluidity and versatility of human-level intelligence. Related to the SPTI is another major discovery: {em that mathematics may be seen as a set of techniques for IC, and their application}. This suggests the creation of a {em New Mathematics} via the integration of mathematics with the SPTI, combining the strengths of both. The SPTI also suggests new thinking in concepts of probability and new thinking about `computation’, with potential benefits in both areas. The SPTI has been shown in peer-reviewed papers to be relevant to areas not closely associated with AI. These include: the management of `big data'; the development of autonomous robots; medical databases; sustainability of computing; transparency in computing; and computer vision.
Category: Artificial Intelligence
[2] viXra:2403.0026 [pdf] submitted on 2024-03-06 21:36:57
Authors: Jinho Kim, Jooney Han
Comments: 10 Pages.
In this work, we aim to solve the problem of unauthorized learning of works arising from the process of collecting large amounts of data from Text to Image (TTI) AI models represented by Stable Diffusion. The TTI model performs indiscriminate web data crawling to collect a substantial number of images, and these images are used for model learning without the consent of the original author. The TTI model is capable of learning the drawing style of an image, which undermines the value of the original work. Therefore, we suggest a method of transforming images to deteriorate the learning accuracy of TTI models. Then, we compare the quality of original images to images processed by the modification method presented in this study, using both quantitative measurement and qualitative measurement. Thus, we confirm that the image modification method we propose prevents AI models from learning literary works without permission.
Category: Artificial Intelligence
[1] viXra:2403.0021 [pdf] submitted on 2024-03-06 07:43:20
Authors: Satish Gajawada
Comments: 2 Pages.
Data Science and Artificial Intelligence are popular fields of research. A significant contribution was made to Artificial Intelligence in the recent past by defining branches like "Artificial Intelligence Plus Plus (AI++)", "The Interesting and Complete Artificial Intelligence (ICAI)", "Out of the Box Artificial Intelligence (OBAI)", "Twenty Second Century Artificial Intelligence (TSCAI)". A similar significant contribution can be made to Data Science by defining branches like "Data Science Plus Plus (DS++)", "The Interesting and Complete Data Science (ICDS)", "Out of the Box Data Science (OBDS)", "Twenty Second Century Data Science (TSCDS)". This article is based on these research gaps. The primary focus of this work is to coin, define and invent a new Data Science field titled "Data Science Plus Plus (DS++)".
Category: Artificial Intelligence
| Third-Party Links: |
|