Robert Lange on NN Pruning and Collective Intelligence

Intrinsic Motivation

Intrinsic motivation plays a crucial role in both artificial and natural systems, as discussed by Robert Lange. He highlights the challenges in current approaches, such as the "television problem," where agents are rewarded for meaningless changes in their environment, like static noise on a screen 1. This issue underscores the need for more sophisticated models that can differentiate between genuine learning opportunities and trivial stimuli. Lange also emphasizes the potential of multi-agent scenarios, where agents can learn from each other without explicit communication, as seen in fish behavior 1.

There's a lot of work also, starting with Jorgen Schmitt Huber. Right on kind of surprise based intrinsic motivation and curiosity.

--- Robert Lange

He believes that understanding these dynamics could lead to more effective AI systems that mimic human-like learning processes.

   

System One and Two

The distinction between system one and system two thinking is pivotal in understanding artificial intelligence. Robert Lange discusses how current reinforcement learning models are largely correlational, akin to Pavlovian conditioning, and lack the higher cognitive functions associated with system two 2. He notes that while there are fragmented efforts to address these gaps, a comprehensive theory integrating these elements remains elusive.

I think when Joshua Bengo talks about moving from system one to system two, there's most definitely something there.

--- Robert Lange

Keith Duggar adds that human intelligence often involves abstract planning, a skill underutilized in current AI models 3. This suggests that future advancements in AI could benefit from incorporating more human-like cognitive strategies.