Sentient Web

[draggett]

This is a stub to be expanded

Introduction

There is increasing use of AI for a wide range of applications based upon deep learning with multi-layer neural networks for classification and control. These have drawbacks in requiring very large amounts of training data, and suffering from a lack transparency. This becomes an issue when AI is used for making decisions that effect people. If the decision is challenged people will demand but not receive an explanation as to why their loan request was reused, or why their insurance premium is so high.

Symbolic representations by contrast enable explanations based upon named relationships and inferences across them. For this reason it is interesting to explore the means to combine symbolic representations with the numerical weights featured in neural networks. The Cognitive Web is an extension of the Semantic Web that combines the Semantic Web with concepts and techniques from Cognitive Science. In particular, associating weights and time stamps with triples and nodes in RDF Graphs as a basis for cognitively plausible models of memory retrieval and rule activation, and based upon work on John R. Anderson on ACT-R.

This is relevant to W3C as an evolution of the Semantic Web and Linked Data to support cognitively plausible mechanisms of recall and reasoning as a basis for future services.

Architecture

The following is adopted from ACT-R by expanding ACT-R's chunks to RDF's triples.

• Cognitive Graph with a set of triples
• Conditon-Action rules
• Input and Output buffers
• Goal stacks

The weights and timestamps are updated to reflect a mathematical model of utility. This applies to both graphs and the rules that operate on them. The rule language would be based upon the shape rule language (SHRL).

Next steps

An initial investigation in the form of a proof of concept implementation and application to selected use cases. This would be followed by the launch of a W3C incubation group to take the work further.

[huchunming]

Seems this idea is more (graph-based, rule-based, and knowledge-based?), am I correct?

[draggett]

More than what? The sentient web builds upon Anderson's work at CMU on ACT-R as a widely acknowledged cognitive architecture combining semantic networks, sub-symbolic processing, production rules and reinforcement learning. As such it builds upon the extensive work in cognitive science. My insight has been to reinterpret this as an extension to RDF, and to consider how cognitive AI systems can be trained using natural language based teaching materials.

There are plenty of challenges to address to realise the full benefits, but incremental benefits are possible e.g.using manually programmed rule sets where human models of memory including priming effects and forgetfulness are valuable.

For natural language the emphasis is the mapping between language and semantic graphs rather than on parse trees. This involves a fresh approach to statistical natural language processing given that current work emphasises parsing and puts very little attention on semantic representations.

Common sense and emotional intelligence can be considered as a broad set of skills, which can be decomposed and arranged in a dependency graph. This allows such skills to be taught via a sequence of lessons, with an emphasis on counter factual reasoning and reinforcement learning.

ACT-R's semantic networks are closely related to RDF's triples and it should be noted that the explicit symbols in semantic networks have a dual representation as vector spaces and tensor expressions. Much of the current work on AI focuses on deep learning for neural networks, and works well for perceptual tasks. However, little work as yet has been done on production rules with the neural network paradigm. I anticipate that this gap will be filled, leading to a much more complete account of cognitive phenomena.

The above is probably rather cryptic as if covers a great deal of ground. I would be happy to explain further.

[draggett]

Since adding this issue, I have made considerable progress, launching the Cognitive AI CG, developing an open source library and a suite of demos, extensive documentation and thanks to François, we also have a formal specification the chunks and rules format. The work has also been published in ERCIM News n125 and presented externally on numerous occasions.

More recently the ideas have been incorporated into a narrative around simplifying digital transformation with web-based access to federated enterprise-wide knowledge graphs, along ideas for web-based editing of UML and ER diagrams, and cognitive agents to migrate Excel spreadsheets into knowledge sheets that are integrated as part of knowledge graphs. A workshop on this is under consideration.