Welcome to the Cognitive AI CG

Dear Saurabh,

Welcome to the Cognitive AI CG. Could you please introduce yourself to group?  You can find most of the materials for the group on github at https://github.com/w3c/cogai.

In summary, we’ve worked on chunks & rules as a lightweight representation of facts and rules inspired by John Anderson’s work on ACT-R, and as a level above RDF’s triples.  I’ve been working on applying this to industrial use cases, e.g. for controlling robots, and as part of swarms of agents.  Other work covers plausible reasoning inspired by Alan Clark’s studies in the 1980s.  See the specification for PKN (plausible knowledge notation) and the associated web demo. We’re also looking at the future of AI with Sentient AI as a quantum leap forward from today’s Generative AI.

Generative AI is powerful but has known limitations, e.g. lacking the ability to acquire new skills once deployed, and is prone to making things up (hallucination). Sentient AI adds episodic memory along with continual learning through continual prediction and continual reasoning, enabling agents with awareness of their environment, goals and performance, along with the means to remember and reflect on their past experiences. Research is now needed on the technical challenges to realise this breakthrough.

Sentient AI is likely to require new approaches to training that don’t rely on biologically implausible back propagation with gradient descent. Instead, we can expect to exploit local learning signals at each layer for continual learning. We want to mimic human cognition more faithfully, for example, focusing attention on the unexpected when the data doesn’t match the predictions, and using System 2 cognition to guide learning. Something like emotion will be needed to focus cognition on what is currently important, akin to a cognitive operating system.

p.s. I’ve also started work on AI powered personal agents as a means to access services over the Web. Your Personal Zone includes your persona agent, your devices, your AI apps, your home server and your personal cloud, providing safe and secure access to services. You can think of AI apps as akin to mobile apps that run within your personal zone. One of the challenges is to combine this with zero-knowledge proofs as a basis for protecting personal data. Personal agents can also be used to monitor security and block harmful content according to your preferences and parental controls. In the short term, personal agents can built around large language models, but I expect this will evolve to use Sentient AI as work on that progresses.

Best regards,

Dave Raggett <dsr@w3.org>