自己組織化による学習：人工知能の新たな基礎を開拓

通过自组织学习：开创人工智能的新基础

• 批准号: 22K19814
• 负责人: VARGAS DANILO
• 金额: $ 4.16万
• 依托单位: Kyushu University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Challenging Research (Exploratory)
• 财政年份: 2022
• 资助国家: 日本
• 起止时间: 2022-06-30 至 2024-03-31
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-22K19814/
• 关键词: SyncMap; Self-organization; Novel AI Paradigm; Dynamical systems; 自己組織化; Dynamical Equations; New Learning Paradigm

In this research we further developed the novel machine learning paradigm based solely on dynamical equations and no optimization. SyncMap was further improved in stability (Symmetrical SyncMap), to deal with high dimensionality (Magnum) and also to work with changing hierarchies (Sigma). Symmetrical SyncMap improves the equations of SyncMap to deal with long-term stability of the initial equations. Magnum uses multiple Inertia-SyncMaps in subsets of variables, enabling it to deal with high dimensional problems. Sigma achieves the astonishing feat of both being accurate enough to predict the structure of hierarchical input structures at the same time as being adaptive enough to deal with changes in them throughout the experiments. All these methods were extensively tested and their results were submitted to top journals in the area. Hopefully, most of them will become available in the next few months.We also developed two key understanding of how input sequences can be used generally to deal with images and other non-sequential input. Based on these new findings, we have proceeded to develop a novel architecture for computer vision and two methodologies for supervised learning classification tasks.

在这项研究中，我们进一步开发了仅基于动态方程且没有优化的新颖机器学习范式。 SyncMap 的稳定性（Symmetrical SyncMap）得到了进一步提高，可以处理高维（Magnum），也可以处理不断变化的层次结构（Sigma）。对称SyncMap改进了SyncMap的方程，以处理初始方程的长期稳定性。 Magnum 在变量子集中使用多个 Inertia-SyncMap，使其能够处理高维问题。 Sigma 实现了惊人的壮举，既足够准确地预测分层输入结构的结构，同时又具有足够的适应性来处理整个实验过程中的变化。所有这些方法都经过了广泛的测试，其结果已提交给该领域的顶级期刊。希望它们中的大多数将在接下来的几个月内可用。我们还对输入序列如何通常用于处理图像和其他非序列输入有了两个关键的理解。基于这些新发现，我们着手开发一种新颖的计算机视觉架构和两种用于监督学习分类任务的方法。

项目成果

“Understanding SyncMap: Analyzing the components of Its Dynamical Equation”

“理解 SyncMap：分析其动态方程的组成部分”

• 发表时间: 2022
• 作者: T. Y. Foong; D. V. Vargas
• 通讯作者: D. V. Vargas

“Curious-II: A Multi/Many-Objective Optimization Algorithm with Subpopulations based on Multi-novelty Search”

“Curious-II：一种基于多新颖性搜索的子群体多/多目标优化算法”

• 发表时间: 2023
• 作者: Y. Jiang; D. V. Vargas
• 通讯作者: D. V. Vargas

“Generating oscillation activity with Echo State Network to mimic the behaviour of a simple central pattern generator”

“使用回声状态网络生成振荡活动以模仿简单的中央模式生成器的行为”

• 发表时间: 2023
• 作者: T. Y. Foong; D. V. Vargas
• 通讯作者: D. V. Vargas

“Understanding SyncMap’s Dynamics and Its Self-organization Properties: A Space-time Analysis”

“了解 SyncMap 的动力学及其自组织特性：时空分析”

• 发表时间: 2022
• 作者: H. Zhang; D. V. Vargas
• 通讯作者: D. V. Vargas

Adversarial Robustness Assessment: Why both L0 and L∞ Attacks Are Necessary

对抗性鲁棒性评估：为什么 L0 和 L∞ 攻击都是必要的

• DOI: 10.1371/journal.pone.0265723
• 发表时间: 2022
• 影响因子: 3.7
• 作者: Kotyan, Shashank;Vargas, Danilo Vasconcellos
• 通讯作者: Vargas, Danilo Vasconcellos