Subgraphs of Yann LeCun's Path Towards Autonomous Machine Intelligence
Subgraphs of Yann LeCun's Path Towards Autonomous Machine Intelligence
Introduction
Over 40 years of research have resulted in hundreds of cognitive architectures that describe abilities like perception, attention mechanisms and selection, memory, learning, and reasoning.1
Last week, Yann LeCun shared an extensive, humbly-expressed vision2 for intelligent machines: A Path Towards Autonomous Machine Intelligence
Subgraphs
My goal is to construct meaning by representing portions of Yann’s architecture as semantic triples expressed at a computational level of analysis. Two subgraphs are provided:
Challenges that AI research must address
Architecture modules
Subgraphs present information at a higher level of abstraction than factor graphs.
Questions are offered to spur the explication of AI research challenges.
Learning
As implicit learning is made explicit by graphing, it is practical to enhance knowledge representation by interleaving model constructs and architectural diagrams. The goal is to integrate gradient-based methods and symbolic methods and to strengthen hypotheses about all forms of reasoning.
If cognitive architects model their designs as concept maps, then experiential learning will likely be generalizable to enlarge action spaces for any task at hand. If these steps are instantiated in software, then a form of intelligence amplification is realized.
Challenges that AI research must address
Subgraph: AI research challenges
Source: Section 2 - Introduction of A Path Towards Autonomous Machine Intelligence (Version 0.9.2, 2022-06-27)
Questions about AI research challenges
Regarding the challenges that AI must address, please consider:
Should decomposing be added to observing, representing, predicting, and acting? Or should decomposing be applicable to all?
Is predicting the same as estimating?
How are percepts related to action plans?
How might reasoning and planning be mapped to observing, representing, predicting, acting, and interacting?
Should learning tasks (observing, representing, predicting, acting, interacting) be differentiated from real-world tasks (tasks at hand)?
Triples: AI research challenges
Architecture modules
Subgraph: Architecture modules
Triples: Configurator module
Triples: Perception module
Triples: World model module
Triples: Cost module
Triples: Short-term memory module
Triples: Actor module
Please share your feedback and make use of this material in accordance with the Creative Commons license. 3
Kotseruba, Iuliia and John K. Tsotsos. “40 years of cognitive architectures: core cognitive abilities and practical applications.” Artificial Intelligence Review 53 (2018): 17-94.
… in keeping with Meta AI’s open-science approach, we are taking this opportunity to preview our research vision and ideas in the hope that it spurs discussion and collaboration among AI researchers. The simple fact is that we will need to work together to solve these extraordinarily challenging, exciting problems.
I hope that this piece will help contextualize some of the research in AI whose relevance is sometimes difficult to see.
This is something that is going to take a lot of effort from a lot of people. I’m putting this out there because I think ultimately this is the way to go. … I hate to see people wasting their time.