TL;DR
This paper explores how network connectivity and prolonged evolution impact "general intelligence" in artificial neural networks. They found that denser connections and longer training times, even after initial performance plateaus, lead to better recognition of novel, related tasks, suggesting that general intelligence is shaped by specific evolutionary factors.
What Problem Does This Paper Address?
The paper tackles the question of what drives "general intelligence" – the ability of an agent to perform well on a variety of tasks, even those it hasn't been explicitly trained on. While many AI systems excel at specific, narrow tasks, achieving broader, adaptable intelligence remains a significant challenge. The authors draw an analogy to biological species, which exhibit varying degrees of domain-specific versus domain-general intelligence, and propose that understanding the evolutionary pressures that shape these abilities in artificial systems could offer insights into biological intelligence. They specifically investigate how network architecture (connectivity density) and the duration of evolutionary training influence this general intelligence.
Methods
The researchers employed the NEAT (NeuroEvolution of Augmenting Topologies) algorithm, a well-established method for evolving neural network structures and weights. Their experiments involved evolving simple, two-layer neural networks. The core of their methodology involved two key manipulations:
- Varying Connection Density: They initialized populations of networks with different densities of connections between neurons. This was done to mimic variations in the interconnectivity of brain modules observed in biological systems.
- Prolonged Evolution: After networks reached a plateau in performance on their training tasks, they were subjected to further evolutionary generations. This was to see if continued evolution, even without immediate fitness gains on the original tasks, could improve general intelligence.
The "general intelligence" of a network was measured by its performance on a novel, related task that was not part of its training set.
Datasets
The primary dataset used was the NIST handwritten digit recognition dataset. This dataset was adapted for the experiment by including not only the original digit images but also their mirror images. A specific digit (the number 7) and its mirror image were intentionally excluded from the training set and reserved solely for testing the networks' ability to generalize.
Key Insights and Ideas
- Connectivity Matters: Denser neural network connections, up to a certain point, significantly improve a network's ability to generalize to new, unseen tasks. This suggests that a richer internal structure facilitates more robust learning.
- Evolutionary Time is Crucial: Prolonged evolution, even after performance on the original training tasks has stabilized, can continue to enhance general intelligence. This implies that evolution might be refining underlying mechanisms for adaptability rather than just optimizing for the immediate task.
- General Intelligence is Shaped by Specifics: The findings suggest that general intelligence isn't an abstract, emergent property but is influenced by specific "anatomical" (network structure) and "environmental" (evolutionary history) factors.
- Complexity vs. Overfitting: The paper notes that while increased connectivity is beneficial, excessive complexity can lead to decreased performance, possibly due to the difficulty of evolving very dense networks or issues akin to overfitting in recurrent networks.
Key Results
- Connection Density: Test performance (recognition of mirror images) increased with connection density, peaking around 65% connectivity. Beyond this point, performance began to decrease.
- Prolonged Evolution: Networks with optimal connection density (0.65) showed improved test performance after prolonged evolution, even when their performance on the original training inputs had plateaued. This improvement was observed across various durations of prolonged evolution (50, 100, 200, and 300 generations).
Limitations and Future Work
Limitations: * The study used simple, two-layer neural networks. The findings might differ for more complex network architectures. * The "general intelligence" task was specific (digit mirror images). More diverse and complex tasks would be needed to fully assess general intelligence. * The paper notes that excessive connectivity beyond 65% led to decreased performance, suggesting an optimal range rather than a simple linear increase.
Future Work: The authors suggest extending this research to more realistic tasks and environments that more closely resemble those in which real-life animal species evolve. They also propose that this neuroevolutionary approach could be valuable for creating intelligent behaviors in artificial agents for applications like video games and robotics.
Summary updated