How to Implement Common Sense for AI

This video introduces the KEY To Knowledge! How knowledge is stored in your brain OR in a computerized graph data structure. How such graph is created and searched is fundamental to understanding your mind and how we can implement it in the Brain Simulator III.
________________________________________
For more information, visit:
futureaisociety.org/
Facebook:
/ futureaisociety
Twitter:
/ futureaisociety
________________________________________
The Future AI Society is a member community dedicated to adding Common Sense to Artificial Intelligence.
The Society’s open source software includes the Brain Simulator II neuron simulator and the Brain Simulator III Common Sense AI research system which is a graph-based real-world representation system. BrainSimIII can create connections on its own between different types of real-world sensory input (such as sight, sound, and touch). This represents the next generation of software beyond today’s machine learning, generative AI, and knowledge systems.
Founder Charles J. Simon, BSEE, MSCS, is a nationally recognized entrepreneur, software developer and manager. With a broad management and technical expertise and degrees in both Electrical Engineering and Computer Science, Mr. Simon has many years of computer experience in industry including pioneering work on numerous neurological test system.
#commonsense #artificialintelligence #technologiesthatthink

[@EvelineFlowercrown]

Good Video! It's kinda funny that I just learned all this graph inheritance stuff in uni for the database relation model xD

[@FutureAISociety]

That's awesome! Be sure to stay tuned for the next video which will include some exciting ways that graph relationships are different from DB relationship.

[@christopherfry3126]

Charles, another great video. Superb graphics make this complex topic crystal clear. High schools everywhere should show your videos.

[@FutureAISociety]

Thanks.

[@matejmelek3371]

Awesome video. I can imagine how this fits very well with Hebbian learning. Many Thanks Charles.

[@FutureAISociety]

Glad you liked it!

[@judgeomega]

i tried to build a similar model decades ago but when it came time to add exceptions and negation it became computationally inefficient as suddenly you had to traverse much more of the graph in order to resolve anything. do we really need to accept that huge portions of the graph need to be traversed or is there some trick to make things more efficient?

[@FutureAISociety]

Thanks for this comment... What a difference a few decades make! Firstly, when we think parallel, now our search time only goes up with the number of "hops" not the number of nodes to be searched. As I said in the video, the search depth seems to be a relatively small number. Secondly, searching a list of 1M nodes used to be a big deal. I clocked adding nodes to the graph at over 100,000/sec single-threaded but haven't clocked the search times yet. Next, buried in my "Filter" comment is the idea that there are numerous ways of curtailing the search. We'll have to see what the balance is.

[@sgrimm7346]

Good video. This is more along the lines of what I had mentioned previously about adding videos that describe what the system is doing and how it works. Continued success! I'll be watching for more information. Thank you.

[@FutureAISociety]

Thanks for sticking with it! It will be a few more episodes before we get to the learning process.

[@domyforsale7497]

Hi Charles, Thanks for the video. I appreciate the ideas and the way you presented them. The model you introduced suggests, at least at this stage, a binary approach to the problem of cognition. However, research on natural language and non-linguistic expressions (e.g., Raykowski 2014, 2018, 2019, 2022) suggests the critical importance of nested and concatenated ways of conceptualizing in human cognition. I'm not referring to the use of weights and the ability to vary them in describing relationships but rather to the actual experience of nested and concatenated structures and their connection to sensations and their respective maps. As far as I understand, all neural networks are based on ON-OFF binary relations (essentially non-additive) between objects and their features. I don't believe this mirrors the way humans think. Taking the example of Fido the dog, the tails of different dogs exist on a nested scale from no tail (0) … an average tail … to some maximum (1), which is expressed in the real world by the extent of the tail. Note that I'm not discussing probabilities of having a tail or a binary relationship between the object and its features. Additionally, nested relationships already describe the inheritance of properties. The problem comes down to finding a way to implement such a non-binary system (e.g., expressions) on binary computers. What are your thoughts? Wes R.

[@FutureAISociety]

I agree and am working on the video which expands on the idea that relationships not only connect nodes but indicate the confidence that the relationship is actually true, the time at which it was thought to be true, the time that it is likely to remain true, etc. All this makes relationships spongy and varying over time. This is distinct to whether Fido has a long tail or a short tail which would be an attribute of the tail with a value like the Has3 in the video. I think that you can only internally represent a limited number of different tail lengths just as you can only perceive a limited number of different shades of gray.

[@domyforsale7497]

Hi Charles, Regarding your statement, 'I think that you can only internally represent a limited number of different tail lengths just as you can only perceive a limited number of different shades of gray.' The magic figure in psychology is seven, which is related to short-term memory. In practice, the number is around four, even though the brain can simulate millions of shades. However, this is not what I had in mind when referring to nesting. If the theory is correct, at the level of sensations, the process of accumulation (additivity of layers within a unit) along with the concatenation of unit extents represents the most fundamental cognitive processes already inherent in all sensory/motor maps. In other words, the molecular mechanism expressing gradation, duration, and spatial objects with their shape must exist before it can be utilized. Regarding your statement, 'I agree and am working on the video which expands on the idea that relationships not only connect nodes but indicate the confidence that the relationship is actually true, the time at which it was thought to be true, the time that it is likely to remain true, etc.' My impression is that the related notions of confidence and reliability are top-down activities, unlike the bottom-up sensory schema combining intensity/extent of sensations. What I am trying to convey is that all sensations are already preformatted by the way they are organized in the cortical maps. One can express a property in terms of any levels (e.g., light gray, mid gray, dark gray) of the million shades of gray possible to simulate/generate. It seems to me that I talk about PROPERTIES while you have ATTRIBUTES in mind. One issue with our discussion is the use of the terms. I understand property as a quality or characteristic inherent to an object or entity. In the case of having a tail, it is a defining feature of some entities, like animals, and is typically present or absent rather than existing in degrees. Properties, on the other hand, often refer to measurable qualities that can exist on a scale or be graded. For example, the length of a tail could be considered a property because it can vary in degrees. What are your thoughts on this issue? Wes R.

[@christopherfry3126]

Do you consider your post to be "merely binary"? If no, well, its implemented on a binary computer so obviously you can implement "non-binary" data on a binary computer. If not, well binary just isn't very restrictive!

[@domyforsale7497]

Dear Christopher, Your comment appears to me to be a playful and somewhat rhetorical way of challenging the idea that working with binary systems is inherently limiting. The issue is that, in my earlier comments, I used the term 'binary' in the context of linguistic relations rather than computer technology. I strongly believe that it is quite possible to implement both binary and non-binary linguistic relations on a computer, but it will not be a simple task. If you're curious, I will be happy to elaborate on my ideas and concerns in an email if possible. Best regards, Wes ​@@christopherfry3126

[@FutureAISociety]

@@domyforsale7497 Just a clarification, in my nomenclature, an Property relates to the internal handling of nodes and relationships (IsA hasproperty transitive) while attributes refer collectively to the external object that a node represents. Your properties and attributes are all lumped under attributes in my lingo. Let's talk more about sensory maps next week.

[@Djei3747ejd]

I see you liked spending time with ChatGPT 😁. Thanks for the video!

[@FutureAISociety]

...if you want a lot of answers and don't care if they are correct (like for a video montage), ChatGPT can't be beat!

[@sevret313]

The issue you're facing lies in how to train this AI as you're presenting a very trivial example here and you seem to forget why deeplearning became popular. Real world data isn't as clean as you present it as here, especially not the data we actually care about regarding machine learning.

[@FutureAISociety]

We'll get to how to train the system after a few more videos. I'm working on the next video which will show how the graph is modified to fit more real-world data...fuzziness, time-dependency, etc. Stay Tuned! And join the Future AI Society at futureaisociety.org for more in-depth discussions.

[@allyc0des972]

The human brain does not encode things this way. Why not jump the gun and implement the human brain algorithm, like more bare metal biological approach? This system is good in theory but in practice you will never get AI to integrate naturally with this system since the neural representations dont combine very well with this. Steven Wolfram has already tried this 30 years ago already.

[@FutureAISociety]

Stay tuned for even more details on how your brain works. We HAVE the Brain Simulator II neuron simulator which shows how the brain would encode things exactly as described. We've made a lot of progress both in hardware and our insights into the brain's operation since Wolfram.