This is a summary of the paper "Collective intelligence for deep learning: A survey of recent developments," which explores the relationship between deep learning and collective intelligence.
- Introduction
- Collective intelligence for deep learning: A survey of recent developments
- Conclusion
- References
Introduction
The paper summarized here is:
Collective intelligence for deep learning: A survey of recent developments
- Published in September 2022
- A review paper by David Ha, who later founded Sakana AI
This is also the paper that Sakana AI cites when explaining their philosophy.
※ All figures in this article are cited from the above paper.
Note: This article was translated from my original post.
Collective intelligence for deep learning: A survey of recent developments
Overview
- Background
- Over the past decade, deep learning has achieved remarkable results across a wide range of domains including computer vision, natural language processing, and reinforcement learning
- Challenges in Deep Learning
- Low robustness
- Limited adaptability to new tasks
- Requires rigid prerequisites
- In short, exhibits "rigid" behavior
- Collective Intelligence
- Intelligence that emerges from interactions among many individuals
- A widespread phenomenon in nature: ant colonies, brain cells, human society, etc.
- Key concepts: self-organization, emergent behavior, swarm optimization, cellular automata
- Deep Learning and Collective Intelligence
- These two fields have influenced each other
- This paper introduces compelling research at the intersection of these domains
History and Characteristics of Deep Learning
Deep learning has a long history
- Neural networks using backpropagation date back to the 1980s
- They received little attention for many years afterward
In 2012, deep learning suddenly gained prominence
- AlexNet won an image recognition competition
- Won by a significant margin over second place
- Trained neural networks using GPUs
- Deep learning then began being applied across various domains:
- Computer vision: image recognition, image generation
- Natural language processing: text generation
- Reinforcement learning
- Speech recognition and synthesis
Deep Learning Challenges: Behavioral Rigidity
- Not robust: vulnerable to change
- Sensitive to changes in input data
- Vulnerable to unexpected tasks
- Input format and settings must be fixed
- Requires human "engineering" effort
- Specific architectures must be designed for specific tasks
- Example: Consider AlexNet's complex architecture
Engineering vs. Adaptive Approaches
- The limitations of existing deep learning can be viewed as challenges inherent to the engineering approach
- Building neural networks is conceptually similar to "building a bridge"
- Bridges and other engineered structures are built to "withstand" rather than "adapt to" external changes
- Adaptive Approach
- Adaptive approaches are common in nature
- Examples: collective intelligence, self-organization, swarm optimization, etc.
- Bridges built by ants adapt to their environment rather than resisting it
Deep Learning and Adaptive Approaches
- It's not inevitable but rather coincidental that deep learning has followed an engineering approach
- Recently, research incorporating adaptive approaches into deep learning has emerged
- Image processing neural networks applying cellular automata
- Reinforcement learning using self-organizing agents
- Hardware advances, including GPUs, are also driving the convergence of deep learning and collective intelligence
History and Characteristics of Collective Intelligence
What is Collective Intelligence (CI)?
- Distributed intelligence emerging from interactions among individual entities
- Key concepts: diversity, independence, decentralization
- A universal phenomenon in nature
Simulations of Collective Intelligence
- In 1993, research on social interactions using robots was conducted
- In 2000, simulation research inspired by ant colonies emerged
- In 2003, research on Brownian agent models was published
Research Directions in Collective Intelligence
- Has been studied more for understanding principles than for solving problems
- This differs from the history of artificial intelligence
- AI has primarily aimed to solve problems such as classification and optimization
History of Cellular Neural Networks
In the 1980s, Leon Chua's group created Cellular Neural Networks
- Cellular Neural Network: cellular automata utilizing neural networks
- Cells (or neurons) interact with adjacent cells
- A cell's state is continuously updated by a nonlinear function taking its own state and neighboring cells' states as input
- Operates in continuous time, unlike discrete-time deep learning
- Often implemented as nonlinear analog electronic components
Rise and Fall of Cellular Neural Networks
- From the 1990s to the mid-2000s, Cellular Neural Networks flourished as an AI subfield
- Applications: image processing, texture analysis, partial differential equations, modeling biological systems, etc.
- Numerous papers and conference presentations; a vibrant academic field
- Multiple startups were founded
- However, in the late 2000s, interest suddenly declined
- Interest in specialized hardware for Cellular Neural Networks was displaced by GPUs
- Google Trends clearly shows this shift in interest between "Deep Learning" and "Cellular Neural Network" (see figure above)
- One likely reason for the decline was the difference in accessibility
- Deep learning's ecosystem developed so anyone with basic programming skills could experiment with it
- Cellular Neural Networks required knowledge of electronic engineering, including analog circuits, making them harder to master
- While Cellular Neural Networks declined, their concepts live on
- Complex systems, self-organization, emergent behavior
Deep Learning and Collective Intelligence
Collective intelligence is intelligence emerging from interactions among multiple independent entities. Let's examine this in connection with neural networks.
Image Processing
Neural Cellular Automata (neural CA)
- Treats pixels as cells in a neural network
- Trained to determine their own color based on neighboring cells' states
- Even though each cell only knows information from neighboring cells, they can collectively represent an image
- This method can also adapt to noise and image damage
- Can "regenerate" when damaged
- Application to text classification
- Applied to the MNIST classification task (see figure below)
- Each cell determines classification
Neural CA Regeneration in 3D
- The ability to regenerate/repair 3D models would be valuable for many applications
- Such as completing scan data
- Research using Minecraft to regenerate whole structures from partial blocks (see figure below)
- Regenerating objects composed of thousands of blocks
- Regenerating creatures
Deep Reinforcement Learning
Challenges in Deep Reinforcement Learning (Deep RL)
- Like deep learning generally, vulnerable to changes in assumed prerequisites
- Changes in input/output format
- Example: changing a four-legged agent to six legs
- Task changes
- Changes in input/output format
Modularized Agents: Soft-body Robot Simulation
- Compose soft-body robots from modules each controlled by neural networks (see figure below)
- Each module acts based on its own perception
- Result: learned walking behavior
- Research also exists on soft-body robots that regenerate their own bodies by combining neural CA
Neural Network Modules as Robot Actuators
- Each actuator module takes only information within its perception range as input and controls only what's in its control range
- Information is exchanged between adjacent modules
- Walking is possible with the same policy even when morphology changes
Extensions of Deep RL
- Agents that produce diverse morphologies
- Agents that handle noise in input information
- And many other developments
Multi-Agent Learning
Increasing the Number of Agents
- The number of agents in multi-agent reinforcement learning (MARL) is often just a few
- Collective intelligence observed in nature operates at a much larger scale
- Thousands or even millions
- There's value in increasing the number of agents
- Hardware advances are making this possible
MAgent: Simulation of Millions of Neural Network Agents
- Enables research on social phenomena, hierarchical structures, and language emergence
Neural MMO: Platform for Studying Multi-Agent Behavior
- Enables large-scale agent simulation
- Each agent can have its own neural network parameters
- This was previously difficult due to memory constraints
- When agents are composed of unique parameters, phenomena such as niche differentiation have been observed
Meta-Learning
What is Meta-Learning?
- Aims to train systems that can learn how to learn
- Creating a meta-learner
- Transfer learning can be considered a form of this
- Schmidhuber and others have emphasized the importance of meta-learning
Neurons as Reinforcement Learning Agents
- Treat neurons in a neural network as reinforcement learning agents
- Neuron's observation: current state
- Neuron's action: modifying neuronal connections
- In this framework, meta-learning can be understood as a reinforcement learning problem for neurons as agents
- While interesting, this approach has limitations in that it can only solve very simple problems
Learning Without Backpropagation
- Allow neurons and synapses to each have states
- Maintained as neural network states rather than just scalar values
- Like cellular automata, neurons and synapses update themselves based on surrounding states
- Inference and learning occur simultaneously
- These behavioral rules are parameterized as a "genome vector"
- Updating this genome vector constitutes meta-learning
- Using evolutionary algorithms and optimization methods
- Rules meta-learned this way showed better generalization performance
Application of RNNs
- Research also exists on composing neurons and synapses with RNNs
- Weights are represented as hidden states of RNNs
- RNNs can be made to emulate backpropagation
- Apparently able to perform backpropagation while simultaneously conducting inference and learning
Conclusion
Above is a summary of the paper "Collective intelligence for deep learning: A survey of recent developments."
Below are my personal notes:
- What cited papers should I read next?
- Thoughts
- The more I learn about deep learning, the more I realize its structure and design are more rigidly defined than I initially thought. I read this paper because I felt current deep learning lacks the mystique that comes from the complex systems found in nature. I didn't particularly relate to the concerns about flexibility in deep learning, given the recent impressive generalization performance of LLMs. However, I've long felt uneasy about the clear separation between inference and learning, so the examples of learning occurring during forward passes were particularly interesting.
[Related Articles]
