What is Self-Organisation?

• Self-organisation refers to the process where systems or organisms create complex structures without external guidance.
• It happens at all levels of biological life, from molecules forming proteins to cells forming complex systems like tissues and organs.
• The idea is that “the whole is greater than the sum of its parts,” meaning when parts come together, they create something more than what each part could do on its own.

Why is Self-Organisation Important?

• Self-organising systems are crucial for the functioning of all biological life, from single-celled organisms to complex human societies.
• These systems allow organisms to grow, adapt, and even repair themselves when damaged, all without direct external input.
• Self-organisation plays a key role in evolution, helping life forms respond to their environment and survive.

What is Differentiable Programming?

• Differentiable programming is a technique in machine learning where models are designed to learn over time through optimization.
• It allows systems to improve by adjusting their internal parameters to meet specific goals, like self-organising and adapting to changes in their environment.
• In this research, differentiable programming is used to learn agent-level policies that help achieve larger system-level objectives.

What is a Cellular Automaton?

• A Cellular Automaton (CA) is a model used in computing and biology that simulates how cells or agents interact to form complex patterns.
• It consists of a grid of cells, each of which can be in one of several states, and the state of each cell changes based on the states of its neighbors.
• This model is used to understand how complex behaviours can emerge from simple rules, much like how simple organisms can form complex life forms.

Growing Neural Cellular Automata (NCA)

• This study investigates morphogenesis, the process by which organisms grow and form their bodies.
• The authors propose using Neural Cellular Automata (NCA) to simulate this self-organising process, where a single cell can grow into complex structures.
• The model is designed to be differentiable, allowing it to learn and improve over time.
• The goal is for this model to be able to create any structure starting from a single cell, mimicking the way living organisms develop.

Self-Classifying MNIST Digits

• This follow-up study applies the NCA model to a new task: self-classifying digits from the MNIST dataset (a collection of handwritten digits).
• Instead of manually labeling the digits, the Cellular Automaton (CA) is taught to classify them on its own.
• The model adapts to the digits, learns the patterns, and can even self-correct if the input is changed or altered.

Self-Organising Textures

• This work uses NCA to generate textures that mimic real-world patterns, such as those found in nature.
• First, the system learns to reproduce textures from template images.
• Then, it creates new textures that “fool” a vision model, much like how camouflage works in nature.
• The textures that the model generates are surprising and often unexpected, demonstrating the robustness of NCA models.

Adversarial Reprogramming of Neural Cellular Automata

• This research shows how Neural CAs can be reprogrammed to perform tasks they were not initially designed for.
• The authors demonstrate how MNIST classification can be sabotaged, causing the CA to produce incorrect outputs.
• Similarly, the shapes and colors of the Growing CA patterns can be altered through adversarial manipulation.

Key Takeaways

• Self-organising systems, from simple cellular automata to complex human societies, are essential for life and adaptation.
• Using differentiable programming, we can create systems that learn and improve over time to meet specific goals.
• Cellular Automata can simulate processes like morphogenesis, the formation of complex structures, and even learn to perform tasks like digit classification and texture creation.
• Adversarial manipulation allows us to challenge and change the behavior of these self-organising systems, showing their flexibility and potential for diverse applications.

自组织是什么？

• 自组织是指系统或有机体在没有外部指导的情况下创建复杂结构的过程。
• 它在所有生物生命的各个层面上都存在，从分子形成蛋白质，到细胞通过合作和沟通形成像组织和器官这样的复杂系统。
• “整体大于部分之和”这一说法意味着，部件的结合创造了比单独部件能做的更多的东西。

自组织为什么重要？

• 自组织系统对所有生物生命的运作至关重要，从单细胞生物到复杂的人类社会。
• 这些系统使有机体能够成长、适应，甚至在受损时修复自己，而无需外部直接输入。
• 自组织在进化中起着关键作用，帮助生命形式响应其环境并生存下来。

什么是可微编程？

• 可微编程是一种机器学习技术，通过优化设计模型，让模型在时间中学习和改进。
• 它使得系统能够通过调整其内部参数来满足特定目标，如自组织和适应环境变化。
• 在这项研究中，可微编程用于学习代理级别的策略，从而帮助实现更大的系统级目标。

什么是细胞自动机？

• 细胞自动机（CA）是计算机和生物学中使用的一种模型，模拟细胞或代理如何互动形成复杂的图案。
• 它由一个细胞网格组成，每个细胞可以处于几种状态之一，且每个细胞的状态根据邻居的状态变化。
• 这个模型用于理解如何从简单规则中涌现复杂行为，就像简单的有机体如何形成复杂的生命形式。

成长的神经细胞自动机（NCA）

• 这项研究探讨了形态发生过程，生物体如何生长并形成其身体。
• 作者提出使用神经细胞自动机（NCA）模拟这一自组织过程，其中单个细胞可以生长成复杂的结构。
• 该模型是可微的，使其能够随着时间的推移学习和改进。
• 目标是让该模型从单个细胞开始，能够创建任何结构，模仿生物有机体如何发育。

自我分类的MNIST数字

• 这项后续研究将NCA模型应用于一个新任务：自我分类MNIST数据集中的数字（一个手写数字的集合）。
• 该细胞自动机不需要手动标记数字，而是通过自我学习来进行分类。
• 模型适应数字，学习模式，并且在输入发生变化或被修改时，可以自我纠正。

自组织纹理

• 这项工作使用NCA生成模拟现实世界图案的纹理，如自然界中常见的图案。
• 首先，系统学习如何复制来自模板图像的纹理。
• 然后，它创造出新的纹理，能够“迷惑”视觉模型，就像自然界中的伪装一样。
• 该模型生成的纹理令人惊讶且常常出乎意料，展示了NCA模型的强大。

神经细胞自动机的对抗性重编程

• 这项研究展示了如何将现有的神经NCA模型重编程，使其执行原本不为其设计的任务。
• 作者展示了如何欺骗MNIST分类，使细胞自动机输出错误的分类。
• 同样，成长的细胞自动机图案也可以通过对抗性操控改变其形状和颜色。

关键收获

• 从简单的细胞自动机到复杂的人类社会，自组织系统对生命和适应至关重要。
• 通过可微编程，我们可以创建在时间中学习并改进的系统，以实现特定目标。
• 细胞自动机可以模拟形态发生过程，创建复杂的结构，甚至学习执行任务，如数字分类和纹理创建。
• 对抗性操控让我们挑战和改变这些自组织系统的行为，展示了它们的灵活性和广泛的应用潜力。