Gang Fang's Blog

Bengio’s system 2 AI and GFlowNet took inspirations from Bernard Baars’ Global Workspace Theory. Mila researchers take the “limited capacity” element from GWT, and posit that high-level thoughts are a necessity and constructed in a sequential manner using a small number of discrete concepts due to the biological bottleneck.

However, such a position increasingly looks like a leap of faith to me because:

• GWT’s limited capacity seems to be more about the narrow focus of attention.
• GWT seems to study more about sensory consciousness than system 2 style reasoning.
• GWT seems to focus more on the integration and broadcasting aspect.
• GWT is at the end of the day a hypothesis, albeit with backing evidences.

I am now going through more literature, about System 2 AI and GWT both, to understand if my challenge is reasonable or simply due to a lack of deeper understanding.

Mila’s interpretation is not in conflict with GWT, albeit a different focus. The relationship between consciousness and system 2 seems to be consciousness, encompassing a range of phenomena, is a superset of system 2 cognitive abilities. The relationship between GWT and system 2 AI seems to be consciousness prior, which is inspired by GWT, helps design and build system 2 AI.

This interpretation can be easily validated, though I am not sure how scientific it is, by observing my own conscious experiences: when I reason, I focus on a narrow set of elements, construct the logic chain step by step.

When it was first conceived, the GWT is purely psychological and not neuroscientific. The integration and broadcasting aspect of the theory came later, when more evidences on the brain were observed. It makes sense to me now when Baars’ described integration and broadcasting are the most important prediction of GWT. The emphasis of the theory shifted from “limited capacity” to “integration and broadcasting” because those evidences provided a path forward for the theory.

The more recent discussions on GWT are all centered around the brain and the theory of connectome of cortex.

Appendix

Excerpts from Mila literature:

With the help of natural languages, we keep inventing new abstractions and definitions to better compress and make sense of the world.

https://mila.quebec/en/article/scaling-in-the-service-of-reasoning-model-based-ml/

Brain sciences show that conscious reasoning involves a sequential process of thought formation, where at each step a competition takes place among possible thought contents (and relevant parts of the brain with expertise on that content), and each thought involves very few symbolic elements (a handful). This is the heart of the Global Workspace Theory (GWT).

https://milayb.notion.site/95434ef0e2d94c24aab90e69b30be9b3

We closely associate conscious processing to Kahneman’s system 2 cognitive abilities.

https://arxiv.org/pdf/1709.08568.pdf%EF%BC%89%E3%80%82

Excerpts from Consciousness literature:

GWT started in the 1980s as a purely psychological theory of conscious cognition

https://www.frontiersin.org/articles/10.3389/fpsyg.2021.749868/full

Cortex is extraordinarily flexible in its dynamic recruitment
of different regions for different tasks. Therefore, an arbitrary
division between prefrontal and other neuronal regions tends
to be misleading. Consciousness requires a much broader, more
integrative view

https://www.frontiersin.org/articles/10.3389/fpsyg.2021.749868/full