Jim These are just my thoughts for now. They may not necessarily be correct.
By CAS-based SDLC I mean a specific systems development life cycle that is geared towards complex-adaptive systems and being supported in engineering terms by a method and a management system specific to purposes of regenerative systems. From this, you gather my perspective of what AI should be. I accept that the definition of AI has somewhat fallen into ever-broader categories. However, XAI seems most specific in wanting to be able to trace and even control (in a sense) a degree of autonomous-machine outcomes all the way back and forward within machine logic. My contention is that this calls for a specialized, systems-engineering approach, not smart programming alone. I think it safe to say that we probably do live in a binary universe. I find your comments in this regard illuminating. It would follow that the engineering approach I'm referring to would be able to seamlessly interact with mathematical models. To satisfy such an objective in an "explainable" manner, would at least require auto-converting contextual information into binary constructs. In other words, it might be feasible to state that any image of reality should convert to a testable, objectively-manageable binary construct of the same reality. Human-driven machine programming is a typical way of such conversion, but machine-to-machine programming is another. In trying to understand the XAI question, my view is; the XAI stakeholders are searching for mature products of extensive R&D into non-human translation from X-input to binary-driven regenerative AI. Functional, yet structured in such a manner as to retain human control over it. When AI regulation eventually arrives, as it surely must, this would give the holders of such technology a distinctive, market advantage in that they would be able to offer transparency (in the sense of governance). To pay a few million dollars to a winning submission, and collecting all the other submissions for "free" is a quick and cheap alternative to trying to develop it yourself over the next 5-10 years. I would think that the race against time is truly on now. I'm grateful for this initiative. It reminds me of the value of the work I have already concluded (as R&D persons tend to think they exist in isolation with their passionate ideas in the making). Further, it inspires me to open my work more to the levels of scientific adaptation as implied by you. In the end there's only 1 qualification a product needs to satisfy: Does it work? As you quite rightly said; How can anything be explained to work if there is no demonstrable knowledge as to how it works, or not work? Further to that, how does anyone demonstrate exact knowledge of how-to anything not designed by self, or somehow replicated? The engineering approach I'm advancing here should satisfy the aforementioned knowledge criteria. Regards, Robert Benjamin ________________________________ From: Jim Bromer <jimbro...@gmail.com> Sent: 03 February 2017 09:40 PM To: AGI Subject: Re: [agi] Re: Digest for AGI I have been working on a lot of other things so I would, at this point, would be interested in working on a common AI project (in other words XAI.) I personally do not think that neural nets per se or drawing results from mathematical measures (again per se) have the potential to solve this problem. I do agree (I realize) that some kind of network of related idea-like data has to be used to achieve stronger AI (like XAI) but I just don't feel that creating recognition nets that could recognize the components of other learning nets has much potential in the short term. I've been wrong about a lot of things before so I could be wrong about this. I am not sure what Nanograte Knowledge Technologies was talking about when he mentioned CAS-based SDLC. There is an interesting instructional site on machine learning by Andrew Ng at http://openclassroom.stanford.edu/MainFolder/CoursePage.php?course=MachineLearning Machine Learning - OpenClassroom<http://openclassroom.stanford.edu/MainFolder/CoursePage.php?course=MachineLearning> openclassroom.stanford.edu Course Description. In this course, you'll learn about some of the most widely used and successful machine learning techniques. You'll have the opportunity to ... if you want to get started in ML I had a problem with it because there is something wrong with me. Andrew Ng seems to want the student to be able to make a few inferences which he would have to make if he was figuring it out for himself. So it is as if he leaves out a few key steps so you can have the fun of feeling like you are missing something but not quite getting it.The reason I am not going through the angst of figuring it all out for myself is because I want to save time by getting someone else to teach the key steps to me. I guess I have an emotional issue with his style of teaching. When I got stuck I was, after a week or two, able to guess what he did not specifically state, but then I had so much other stuff to do I never got back to it after that. It was like - OK, he is taking some particular (relatively simple) mathematical theory and then using approximation steps to write a simple ML application to solve it. After dealing with the annoyance of wondering why he did not specify a simple key step that he left out I ended up thinking that I've done things like that enough times that I don't have to learn what he is teaching. Why should I spend more time on it? However I do appreciate what he is doing in making that study freely available (regardless of my complaints) and if I do want to continue online instruction - perhaps when I am more mature emotionally. Jim Bromer On Thu, Feb 2, 2017 at 12:39 PM, Rustam Eynaliyev <rustam.eynali...@gmail.com<mailto:rustam.eynali...@gmail.com>> wrote: Hi guys, I'm looking to participate in the General AI challenge. My background: CS degree from Indiana University, currently a full-stack web developer, working on government projects. Looking to learn AI and machine learning in the near future and apply my new-learned knowledge to the General AI competition and may be, making a career switch to a Machine Learning engineer/researcher. Please email me if you're interested in joining me. Rustam On Mon, Jan 16, 2017 at 6:31 AM, <a...@listbox.com<mailto:a...@listbox.com>> wrote: This is a digest of messages to AGI. Digest Contents 1. Re: [agi] Explainable Artificial Intelligence 2. Re: [agi] Explainable Artificial Intelligence 3. Unsharp Mask? 4. Re: Unsharp Mask? Re: [agi] Explainable Artificial Intelligence<https://www.listbox.com/member/archive/303/2017/01/20170109075356:AC4083B6-D66A-11E6-9365-BC080493EEC8> Sent by Jim Bromer <jimbro...@gmail.com<mailto:jimbro...@gmail.com>> at Mon, 9 Jan 2017 07:53:49 -0500 The explanation (sorry for the duality of that term in this discussion) seems to imply that the explanatory AI would either have to be significantly better than anything contemporary Deep Learning or it would have to work with Deep Learning so that it could be used to explain what features the DL network was using to make its categorical selections. (DL is the dominant AI method right now.) Old AI did not really get to the point where it could extend the ability to draw good inferences on a sufficiently wide spectrum of data. So as long as the features in the data were strongly consistent with the features in the training, or strongly related to a simple defined range, old AI could work. DL is important just because it has noticeably extended AI capabilities. I do think DL is a kind of hybrid of old AI and NNs. Jim Bromer [ trailing quoted section removed ] Re: [agi] Explainable Artificial Intelligence<https://www.listbox.com/member/archive/303/2017/01/20170110091535:3E847E9C-D73F-11E6-8CFF-5667BCDDB970> Sent by Jim Bromer <jimbro...@gmail.com<mailto:jimbro...@gmail.com>> at Tue, 10 Jan 2017 09:15:26 -0500 Explainable AI may start out as being as difficult as np-complete. There are some cases where it will work out well, but there are going to be a lot of other cases where it won't. Human thinking has a lot of unfilled spaces and it may be that is a key to solving problems that are np-hard. By substituting good estimates and approximations *at the right time* we can *begin* to explain why we make some decision (once we have had some practice). A discrete-based system (like Watson) can give you a response about why a poor alternative was disqualified but it cannot explain that decision in any depth beyond noting that it got some kind of score and how that score might have been arrived at. That is the problem with combinations of lossy decision methods (including discrete decision processes where the store of the precise steps taken are disposed of once the product of the method is computed). And systems where the reasons are distributed and combined across the memory (like neural nets) are probably as difficult as np-complete problems. As a neural net grows larger it will be more difficult to keep track of the all the embedded 'feature detection objects' and if each one of them has to be related to all the others (to be able to use them in explanations) then this will result in a combinatoric explosion. Maybe a deep learning net (a hybrid system) could be used to analyze a deep learning net, but if so why isn't this a straightforward project? Unsharp Mask?<https://www.listbox.com/member/archive/303/2017/01/20170111101243:6266B664-D810-11E6-82C1-D511BDDDB970> Sent by Jim Bromer <jimbro...@gmail.com<mailto:jimbro...@gmail.com>> at Wed, 11 Jan 2017 10:12:33 -0500 Unsharp Mask? Jim Bromer Re: Unsharp Mask?<https://www.listbox.com/member/archive/303/2017/01/20170111112210:186A357C-D81A-11E6-B416-CFA9F54E02C3> Sent by Jim Bromer <jimbro...@gmail.com<mailto:jimbro...@gmail.com>> at Wed, 11 Jan 2017 11:22:03 -0500 You need a primary subject against a background for an unsharp mask to work on an out-of-focus image. Jim Bromer [ trailing quoted section removed ] AGI | Archives<https://www.listbox.com/member/archive/303/=now> [https://www.listbox.com/images/feed-icon-10x10.jpgecd5649.jpg?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2ZlZWQtaWNvbi0xMHgxMC5qcGc] <https://www.listbox.com/member/archive/rss/303/25423763-04a01972> | Modify<https://www.listbox.com/member/?&;> Your Subscription [https://www.listbox.com/images/listbox-logo-small.pngecd5649.png?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2xpc3Rib3gtbG9nby1zbWFsbC5wbmc] <http://www.listbox.com> AGI | Archives<https://www.listbox.com/member/archive/303/=now> [https://www.listbox.com/images/feed-icon-10x10.jpgecd5649.jpg?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2ZlZWQtaWNvbi0xMHgxMC5qcGc] <https://www.listbox.com/member/archive/rss/303/24379807-653794b5> | Modify<https://www.listbox.com/member/?&;> Your Subscription [https://www.listbox.com/images/listbox-logo-small.pngecd5649.png?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2xpc3Rib3gtbG9nby1zbWFsbC5wbmc] <http://www.listbox.com> AGI | Archives<https://www.listbox.com/member/archive/303/=now> [https://www.listbox.com/images/feed-icon-10x10.jpgecd5649.jpg?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2ZlZWQtaWNvbi0xMHgxMC5qcGc] <https://www.listbox.com/member/archive/rss/303/26941503-0abb15dc> | Modify<https://www.listbox.com/member/?&;> Your Subscription [https://www.listbox.com/images/listbox-logo-small.pngecd5649.png?uri=aHR0cHM6Ly93d3cubGlzdGJveC5jb20vaW1hZ2VzL2xpc3Rib3gtbG9nby1zbWFsbC5wbmc] <http://www.listbox.com> ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
