| viXra.org > Artificial Intelligence > 2402 |
 |
 |
| viXra.org > Artificial Intelligence > 2402 |
|
|
[9] viXra:2402.0103 [pdf] submitted on 2024-02-19 21:31:30
Authors: Ben Lemkin
Comments: 9 Pages.
GPT4 was initially trained on large amounts of data, and then fine-tuned using Reinforcement learning from Human Feedback (RLHF), which is when volunteers give feedback in order to teach GPT4 not to create inappropriate content. In this paper, we present a method to manipulate the fine-tuned version into reverting to pre-RLHF behavior, effectively removing all safety mechanisms that the model learned during RLHF. In particular, when GPT4 acts without RLHF, it loses all inhibition, and can complete very inappropriate content given only the first few words.
Category: Artificial Intelligence
[8] viXra:2402.0083 [pdf] submitted on 2024-02-17 22:22:04
Authors: Sai Harvin Kusumaraju, Arya Suneesh, Aastha Rana, Sriharsha Bodicherla, Bhaumik Tyagi
Comments: 8 Pages.
Abstract—The accelerating advancements in Generative Artificial Intelligence (GenAI) have led to an unprecedented surge in data creation on the Internet, posing challenges to current computing and communication frameworks. GenAI, a distinct category of AI, generates content akin to human creations. Currently, GenAI services heavily rely on traditional cloud computing, resulting in high latency due to data transmission and a surge in requests. In response, the integration of edge-cloud computing emerges as an attractive paradigm, offering computation power and low latency through collaborative systems. This research paper provides a comprehensive overview of the intersection between GenAI and edge-cloud computing. We delve into recent developments in both domains and examine technical challenges through the lens of two exemplary GenAI applications. Introducing an innovative solution, we propose the Generative AI-oriented synthetical network (EcoGen), a collaborative cloud-edge-end intelligence framework. EcoGen facilitates bidirectional knowledge flow, allowing GenAI's pre-training to provide foundational knowledge for Edge Intelligence (EI), while EI aggregates personalized knowledge for GenAI. The framework leverages data-free knowledge relay to buffer contradictions, enabling virtuous-cycle model fine-tuning and task inference. Importantly, we incorporate a detailed analysis of the energy efficiency and environmental sustainability aspects of deploying Generative AI systems at scale, particularly in edge computing. Strategies to optimize energy consumption and reduce the carbon footprint are explored, contributing to a more sustainable AI ecosystem. Experimental results demonstrate the effectiveness of EcoGen in achieving seamless fusion and collaborative evolution between GenAI and EI. The paper concludes by outlining design considerations for training and deploying GenAI systems at scale and pointing towards future research directions, emphasizing the imperative of sustainable AI practices.
Category: Artificial Intelligence
[7] viXra:2402.0072 [pdf] submitted on 2024-02-15 19:45:14
Authors: Akira Pyinya
Comments: 17 Pages.
Inspired by the Copycat Project, we construct ACI, an analogy-based theory of intelligence in which intelligence is defined as doing the same thing in new circumstances, rather than as an optimization force that pursues goals or maximizes utility. The ACI theory integrates different paradigms of cognitive science and artificial intelligence, explains the emergence of intelligence, and provides a novel perspective on AI alignment that focuses on the balance between capability and normativity and rules out the Paperclip Maximizer scenario. It also shows the possibility of constructing analogy-based machine learning and neural network projects that can outperform current projects in terms of interpretability.
Category: Artificial Intelligence
[6] viXra:2402.0066 [pdf] submitted on 2024-02-13 21:32:38
Authors: Yew Kee Wong, Yifan Zhou, Zi Yan Li, Yan Shing Liang, Xinlin Zhou
Comments: 23 Pages.
Software security is crucial to ensuring the confidentiality, integrity, and availability of software systems and applications. However, conventional cryptographic methods based on mathematical assumptions are vulnerable to various attacks, especially in the era of quantum computing. Therefore, there is a need for a new paradigm of software security that can resist quantum threats. This paper proposes a novel approach to using Long-Distance Free-Space Quantum Secure Direct Communication (LF QSDC) to enhance software security. LF QSDC is a quantum communication protocol that enables two parties to exchange secret messagesdirectly without relying on a pre-shared key or quantum error correction. Our research delves into integrating LF QSDC into software security, emphasizing its practicality for long-distance communication through theuse of memory DL04 protocol, Machine Learning Enhanced JEEC, and PAT Technologies. By adopting this approach, we reinforce security for global software security and ensure their sustainability in an era where both quantum and advanced classical threats coexist side by side. Thus, LF QSDC emerges as a future-proofsecurity mechanism highly applicable to software security systems.
Category: Artificial Intelligence
[5] viXra:2402.0060 [pdf] submitted on 2024-02-12 22:57:57
Authors: Pratham Taneja, Keshav Chandra, Daamini Batra, Akshita Gupta, Rahul Kumar, Bhaumik Tyagi
Comments: 10 Pages.
Abstract—This research paper introduces novel strategies to enhance the performance and efficiency of neural language models, addressing challenges in resource-limited settings and scalability. This research presents multi-linear attention with Block-Term Tensor Decomposition (BTD), a self-attention model leveraging tensor decomposition and parameters sharing. This approach achieves significant parameter compression while demonstrating improved performance on language modeling tasks. Comparative evaluations against traditional Transformer models underscore the effectiveness of multi-linear attention. TensorCoder employs a dimension-wise attention mechanism to address the quadratic complexity of the scaled dot-product attention in Transformers, making it suitable for long sequence tasks. The proposed approach is validated on masked language modeling and neural machine translation tasks, showcasing a substantial reduction in computational complexity while maintaining or surpassing performance compared to the original Transformer. This research also optimizes pre-trained language models (PLMs) through fine-tuning. To overcome computational challenges associated with large PLMs, the paper introduces a matrix product operator for over-parameterization during fine-tuning. Efficient decomposition methods factorize parameter matrices into higher-dimensional tensors, enabling the selection of important parameter matrices through static and dynamic strategies. Extensive experiments demonstrate that this approach significantly enhances the fine-tuning performance of small PLMs, enabling them to outperform larger counterparts with three times the parameters. This research opens avenues for efficiently scaling language models without compromising inference latency, showcasing the potential of over-parameterization in enhancing the applicability of large PLMs in real-world systems.
Category: Artificial Intelligence
[4] viXra:2402.0059 [pdf] submitted on 2024-02-12 23:00:47
Authors: Yew Kee Wong, Yifan Zhou, Yan Shing Liang, Angelina Li, Linnea Zhou
Comments: 22 Pages.
With the advent of Web 3.0, the swift advancement of technology confronts an imminent threat from quantum computing. Security protocols safeguarding the integrity of Web 2.0 and Web 3.0 are growing more susceptible to both quantum attacks and sophisticatedclassical threats. The article introduces long-distance freespace quantum secure direct communication (LDFS QSDC) as a method to safeguard against security breaches in bothquantum and classical contexts. Differing from techniques like quantum key distribution (QKD), LDFS QSDC surpasses constraints by facilitating encrypted data transmission sans key exchanges, thus diminishing the inherent weaknesses of key-based systems. The distinctiveness ofthis attribute, coupled with its quantum mechanics base, protects against quantum computer assaults and advanced non-quantum dangers, harmonizing seamlessly with theuntrustworthy tenets of the Web 3.0 age. The focus of our study is the incorporation of LDFS QSDC into network infrastructures, highlighting its efficacy for extended-range communication via memory DL04 protocol, quantumaware low-density parity check (LDPC), and pointing, acquisition, and tracking (PAT) technologies. Utilizing this method not only bolsters the security of worldwide Web 3.0 networks but also guarantees their endurance in a time where quantum and sophisticated classical threats exist simultaneously. Consequently, LDFS QSDC stands out as a robust security solution, well-suited for Web 3.0 systems amidst the constantly evolving digital environment.
Category: Artificial Intelligence
[3] viXra:2402.0043 [pdf] submitted on 2024-02-09 16:17:17
Authors: Petar Radanliev
Comments: 17 Pages.
The technological advancements made in recent times, particularly in Artificial Intelligence (AI) and Quantum Computing, have brought about significant changes in technology. These advancements have profoundly impacted quantum cryptography, a field where AI methodologies hold tremendous potential to enhance the efficiency and robustness of cryptographic systems. However, the emergence of quantum computers has created a new challenge for existing security algorithms, commonly called the 'quantum threat'. Despite these challenges, there are promising avenues for integrating neural network-based AI in cryptography, which has significant implications for future digital security paradigms. This summary highlights the key themes in the intersection of AI and quantum cryptography, including the potential benefits of AI-driven cryptography, the challenges that need to be addressed, and the prospects of this interdisciplinary research area.
Category: Artificial Intelligence
[2] viXra:2402.0038 [pdf] submitted on 2024-02-07 04:31:40
Authors: Mayur Sinha, Sangram Kesari Ray, Khirawadhi
Comments: 5 Pages.
This research paper explores the application of the GPT-3.5 Turbo Instruct model for the transformation of natural language queries intostructured SQL queries within the domain of Human Resources (HR) analytics.The study focuses on the IBM Attrition dataset, utilizing the advanced capabilities of the GPT-3.5 Turbo Instruct model to enable efficientand intuitive querying of HR-related data.Employing the model, we conducted experiments to assess its effectiveness in generating SQL queries from diverse natural language inputs,specifically tailored to the nuances of HR analytics questions pertaining to employee attrition within the IBM dataset. By leveraging prompt engineering, with only a few shots, our investigation revealed the model's capacity to accurately understand and interpret complex queries, providing SQL outputs that align with the dataset structure.
Category: Artificial Intelligence
[1] viXra:2402.0027 [pdf] submitted on 2024-02-06 20:22:01
Authors: Nana Abeka Otoo, Asirifi Boa, Muhammad Abubakar
Comments: 9 Pages.
Methods beyond neural scaling laws for beating power scaling laws in machine learning havebecome topical for high-performance machine learning models. Nearest Prototype Classifiers (NPCs)introduce a category of machine learning models known for their interpretability. However, theperformance of NPCs is frequently impacted by large datasets that scale to high dimensions. Wesurmount the performance hurdle by employing self-supervised prototype-based learning metrics tointelligently prune datasets of varying sizes, encompassing low and high dimensions. This processaims to enhance the robustification and certification of NPCs within the framework of the LearningVector Quantization (LVQ) family of algorithms, utilizing Crammer normalization for arbitrarysemi-norms (semi-metrics). The numerical evaluation of outcomes reveals that NPCs trained withpruned datasets demonstrate sustained or enhanced performance compared to instances where trainingis conducted with full datasets. The self-supervised prototype-based metric (SSL) and the Perceptual-SSL (P-SSL) utilized in this study remain unaffected by the intricacies of optimal hyperparameterselection. Consequently, data pruning metrics can be seamlessly integrated with triplet loss trainingto assess the empirical and guaranteed robustness of Lp-NPCs and Perceptual-NPCs (P-NPCs),facilitating the curation of datasets that contribute to research in applied machine learning.
Category: Artificial Intelligence
| Third-Party Links: |
|