So the prophets scored One: https://web.archive.org/web/20160205202950/http://www.futuretech.ox.ac.uk/sites/futuretech.ox.ac.uk/files/The_Future_of_Employment_OMS_Working_Paper_1.pdf
On 20.06.2019 17:38, Stefan Reich via AGI wrote:
Prednet develops consciousness? On Wed, Jun 19, 2019, 06:51 Alan Grimes via AGI <[email protected] <mailto:[email protected]>> wrote: Yay, it seems peeps are finally ready to talk about this!! =P Lets see if I can fool anyone into thinking I'm actually making sense by starting with a first principles approach... On first approximation, yer computer's memory is just a giant string of charactors. Since everything you have ever done on your computer has been encoded into this string, we can be reasonably confident that this is a universal way to encode anything, so we go ahead and say that all the modalities (sight, sound, etc,) are mapped to this memory and continue on. [I'm maximally overtired right now, I didn't really get any sleep last night, should wait until tomorrow at least to write this but I want to get a head start...] If you prefer, you can think in terms of a multi-tape turing machine, where the program and working variables are stored on other tapes and the mind file is in main memory, but the point is that we presume there exists a computation that does intelligence using the main memory, which we will say is the complete state of the mind. So therefore what we are looking for is a set of principles and basic operations that, when applied to the main memory, intelligent behavior and subjective experience emerges. [next evening] The most important thing to realize is that the data is NOT RANDOM/stochastic/whatever, it is STRUCTURED,?? It has patterns, it has types, it may have meaning.?? This is true of all modalities. Therefore, in order to create a mind, we need the system to be able to capture that structure and be able to use it in computations. Lets say that we had some identified structure in the data which had some modest but non-trivial k-complexity. Replacing that data with the K-complexity description, even if if were a "good enough" approximation, would satisify most definitions of understanding that structure and creating some way to point to that description and incorporate it in other descriptions would satisfy the definition of abstraction. I hope you are all boned up on your study of non-standard computational models. What I'm getting at is that the expression of the K-complexity is a structured object just like the structured object which it expands to. So we can treat it the same way and approximate it using the basic functions we use to manipulate abstractions. So if you had a string/object that basically meant "make big" and another that expressed the shape, then you only need a way to apply one to the other to make a big shape. It is important that things are done like this because it is how general intelligence is achieved, because it allows the operation as well as the object to be learned by essentially the same mechanism. Now, let me shift gears here and go into a thread that I'd been meaning to start for the past week or so. DEEP LEARNING IS DEAD! If we don't have a viable replacement for it in the next 18 months or so, we're at risk of another AI winter. =0???? There are certainly some good and useful machine learning algorithms out there, some of which are just rediscovery of techniques from ages past, but well that's OK too. When you get down to it, DL is the realization of ideas dating back a very long time but only made feasible with gpGPU techniques, see Nvidia's website for a recounting of the history. [few days later] Have a look at this and count up the number of problems this field of study has from an AGI perspective: https://scholar.google.com/scholar?q=neural+architecture+search&hl=en&as_sdt=0&as_vis=1&oi=scholart Neural nets are brittile in that you can only select their topology when you are creating them. (iirc) The brain, by contrast is quite infamous for exhibiting plasticity in the face of dammage and/or changing requirements. For example, london taxi drivers develop an enlarged hypocampus to help deal with the maze-like roads in that city. https://www.youtube.com/watch?v=UoJf_tXU2Zk Ok, that video was ammusing and little else. ;) Also, consider what a "golden network" would look like. A golden network is a network like a universal turing machine that exhibits general AI instead of some special AI. The bottom line is that while we have come a long way from 2012, we are reaching the top of the S-curve with respect to deep learning. Once the RoI falls below some critical level, we will enter an AI winter. =| We do not have time for another winter. Deep learning is also hitting a computational complexity wall but I couldn't find good referances for that so I left it out of this post. Here's something else to think about. Consider the near photorealistic graphics that you got with Assassin's Creed on the Playstation 3. Consider the amount of power people are throwing at autonomous driving and just recognising the general outlines of visual scenes. One of the reasons I flog prednet so much, https://coxlab.github.io/prednet/ is not just because it yields consciousness, but it reduces perception to something a PS3 could do. You render what you think you are seeing, the LGN of the thalamus computes the error signal, and then you use your expensive complex computations or just that error signal! [midnight a day or two later, just jamming random related topics into the post to make myself look smart...] Here's another fascinating area of study: https://scholar.google.com/scholar?q=explainable+neural+networks&hl=en&as_sdt=0&as_vis=1&oi=scholart To me, it's just proof that deep learning, despite what can be accomplished with it, is NOT a form of intelligence. When you think about it, it doesn't learn in any true sense, in that it can recite facts, or recall events but it, over many thousands of iterations, can *adapt*. The result is a system that essentially operates as mindlessly as a reflex. The apparent complexity of the behavior not changing that fact. Going to a recurrent model does give the network a legit memory, after a fassion. Still, the prospects for further progress is dim. [next night around midnight again] Consider, by contrast a system that works by analyzing structure. There is no reason why it can't simply continuallly fill its memory, because each thing is recorded as a representation of it's structure, the storage requirement for even millenia of ordinary life experience is not at all implausible. Ok, the magic sauce here is an ability to organize and catalog representations. I think there exists a geometric technique that can yield pattern recognition. This would involve a way to normalize the input and then search the database, the transformation used to achieve the normalization also being a relevant structure that is also analyzed in the same manner. Other tricks such as the space-time rotation that I've described a number of times are also important techniques for detecting spatial patterns. I think that approach can be much more powerful than "convolution" which is just scanning a larger matrix with a smaller... I'm talking about taking the visual scene and turning the scan of it into a temporal signal like a sound. Just as a speech recognizer doesn't care when the speaking actually begins, treating spatial signals can work the same way. Anyway, what was this post supposed to be about again? bleh -> hitting send. -- Please report bounces from this address [email protected] <mailto:[email protected]> Powers are not rights. *Artificial General Intelligence List <https://agi.topicbox.com/latest>* / AGI / see discussions <https://agi.topicbox.com/groups/agi> + participants <https://agi.topicbox.com/groups/agi/members> + delivery options <https://agi.topicbox.com/groups/agi/subscription> Permalink <https://agi.topicbox.com/groups/agi/T395236743964cb4b-M686d9fcf7662ad8dc2fc1130>
------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T0a32a6b19608fcc4-M9bc17f509d61f7d68b6a6c26 Delivery options: https://agi.topicbox.com/groups/agi/subscription
