Re: [W3C Group Management] XAI Asset Management has nominated Aric Whitewood to the Cognitive AI Community Group

Thanks Dave - just briefly on circular convolution, this is achievable using the Fast Fourier transform (FFT) and the inverse FFT. It has been almost 20 years since I’ve had to calculate this though!

For two signals, call them s1 and s2, the convolution in this case is ifft ( fft(s1)*fft(s2) )

There is a numpy implementation of the fft. I believe it’s just np.fft.fft

Happy to talk more..

Regards,
Aric


________________________________
From: Dave Raggett <dsr@w3.org>
Sent: Thursday, February 1, 2024 6:34 pm
To: Aric Whitewood <aric.whitewood@xaimacro.com>
Cc: public-cogai@w3.org <public-cogai@w3.org>
Subject: Re: [W3C Group Management] XAI Asset Management has nominated Aric Whitewood to the Cognitive AI Community Group

Thanks for the background.  I am preparing a talk on defeasible reasoning and PKN for the 8th February, and today finally got the time to get started with the implementation work in Python for the neural-based approaches for cognitive agents. Do you know of any Python implementations for circular convolution?

In respect to use cases for balanced arguments, one source of inspiration is Aristotle’s work on rhetoric, which is as pertinent today as it was for Ancient Greece, see my email yesterday to Matteo.  SQL and SPARQL are so low level compared to considerations of ethos, pathos, logos and kairos!

With clever prompting LLMs do a decent job, but we can look forward to much better systems. AI is still at a very early stage of maturity, more like alchemy than science.

Let’s talk further about what you might want to do to support this CG.

Kind regards,
    Dave

On 1 Feb 2024, at 17:12, Aric Whitewood <aric.whitewood@xaimacro.com> wrote:

Hi Dave,
Thanks for the email.

In terms of common-sense reasoning research, just to illustrate the research flow so far. And by the way, this has been more in relation to the education domain.

As some context, some of my historical research topics have included rule databases, rule engines, graph networks and the relations between these. I’ve designed and run some fuzzy rule-based systems as well as graph-based systems in production.

More recently, I’ve done some work on integrating LLMs with Knowledge Graphs.

I became interested in the limitations of human-authored knowledge graphs, and how to mitigate this. For example, Symbolic Knowledge Distillation (there is a paper by Hinton et. al. from 2015 on Knowledge Distillation, and then a paper by West et. al. from 2022 using an actor/critic model). An interesting, related paper is by Choi, on Commonsense Intelligence, 2023. I think both are linked to the Allen Institute.

This crossed over with some of my past work on clustering and fuzzy clustering in particular, and the formation of effective rules from these structures. I was interested in how the strength of neural connections in human brains helps to reinforce particular memories and perhaps concepts, and was interested in combining that idea with the machine-corpus-machine model of the knowledge distillation above. There are some related elements here in terms of using sub-graphs rather than one large knowledge graph - so representing time-based sets of events, concepts, and grammatical structure separately. This is similar in some ways to recent research on using LLMs and KGs to solve multi-stage problems (will need to find the specific paper). Also, there is an element of creating a noisy graph with overlapping concepts but reinforcing the key elements - or adding a hierarchy, particularly when obtained from a trusted expert / source of facts.

And then as a final piece to the research I’ve been looking at feedback mechanisms (although this is a much earlier stage) for allowing the system to ask for more information, for example about a particular concept, and increase the richness of the graph(s) - impacting the strength of connections as mentioned in the paragraph above. I'm still thinking about the best objective function(s) here.  So effectively a feedback loop between the LLM and the Knowledge Graph. This could of course be extended to include elements of reinforcement learning -  but that would add quite a lot of complexity, and instead I wanted to focus more on the graph structure(s) and other parts.

I think your suggestions on use cases and continual learning centric synthetic data are interesting. Perhaps some overlap with the latter and the feedback mechanism above.

(Also - the above is basically some part-time research, and I would say a mixture of code, analysis, and ideas.)

Regards,
Aric




From: Dave Raggett <dsr@w3.org<mailto:dsr@w3.org>>
Sent: 01 February 2024 11:04
To: Aric Whitewood <aric.whitewood@xai-am.com<mailto:aric.whitewood@xai-am.com>>
Cc: public-cogai <public-cogai@w3.org<mailto:public-cogai@w3.org>>
Subject: Re: [W3C Group Management] XAI Asset Management has nominated Aric Whitewood to the Cognitive AI Community Group

Hi Aric,

Thanks for introducing yourself.   From my perspective, I’d welcome help with gathering use cases for effective arguments that balance the case for and against some supposition.  This would help with work on declarative approaches to describing strategies and tactics for developing arguments.

Another opportunity is around developing synthetic datasets for evaluation of different neural architectures for human-like cognition.  ChatGPT is a large language model trained for language prediction.  It would be both interesting and challenging to look at complementary approaches based upon continual learning.

Can you tell us about your research on knowledge representation and common sense reasoning?

Best regards,
    Dave

On 1 Feb 2024, at 10:54, Aric Whitewood <aric.whitewood@xai-am.com<mailto:aric.whitewood@xai-am.com>> wrote:

Hi - great to be a part of the group.

Brief background - I run an Artificial Intelligence (AI) focused hedge fund and am also co-founder of an AI centric Edutech firm. I’ve worked in the AI field for around 18 years, in the defence, finance, and education sectors.

In the hedge fund, we do quite a lot of research on topics like: fuzzy rules based systems, deep reinforcement learning, and unsupervised learning with multivariate time series. Some of that is in conjunction with UCL in London, but most is proprietary.

In the Edutech firm, the emphasis is on integrating knowledge graphs with LLMs to better surface learning objectives and relevant content for students..

I’ve also been doing some research on knowledge representation and common sense reasoning, which led me to this group.

Happy to contribute in whatever way I can, thanks.

Regards,
Aric
From: Dave Raggett <dsr@w3.org<mailto:dsr@w3.org>>
Sent: 31 January 2024 15:54
To: Aric Whitewood <aric.whitewood@xai-am.com<mailto:aric.whitewood@xai-am.com>>
Cc: public-cogai <public-cogai@w3.org<mailto:public-cogai@w3.org>>
Subject: Re: [W3C Group Management] XAI Asset Management has nominated Aric Whitewood to the Cognitive AI Community Group

Hi Aric,

Welcome to the W3C Cognitive AI CG.  I encourage you to email the group with a brief account of your interest and intentions.  So far we have mostly worked on chunks and rules, inspired by John Anderson’s ACT-R, and PKN (plausible knowledge notation) inspired by Alan Collins, with demos and draft specifications for both.

I see opportunities for further work on extending PKN to model tactics and strategies for defeasible reasoning. We’ve also started new work on neural network architectures to experiment with ideas for enabling continual learning, episodic memory and reflective cognition, see:

http://www.w3.org/2023/10/10-Raggett-AI.pdf

If you are a programmer, your help with the latter would be much appreciated.

Best regards,

Dave Raggett <dsr@w3.org<mailto:dsr@w3.org>>

Dave Raggett <dsr@w3.org>