Slashdot Mirror
Artificial General Intelligence is Nowhere Close To Being a Reality (venturebeat.com)
Three decades ago, David Rumelhart, Geoffrey Hinton, and Ronald Williams wrote about a foundational weight-calculating technique -- backpropagation -- in a monumental paper titled "Learning Representations by Back-propagating Errors." Backpropagation, aided by increasingly cheaper, more robust computer hardware, has enabled monumental leaps in computer vision, natural language processing, machine translation, drug design, and material inspection, where some deep neural networks (DNNs) have produced results superior to human experts. Looking at the advances we have made to date, can DNNs be the harbinger of superintelligent robots? From a report: Demis Hassabis doesn't believe so -- and he would know. He's the cofounder of DeepMind, a London-based machine learning startup founded with the mission of applying insights from neuroscience and computer science toward the creation of artificial general intelligence (AGI) -- in other words, systems that could successfully perform any intellectual task that a human can. "There's still much further to go," he told VentureBeat at the NeurIPS 2018 conference in Montreal in early December. "Games or board games are quite easy in some ways because the transition model between states is very well-specified and easy to learn. Real-world 3D environments and the real world itself is much more tricky to figure out ... but it's important if you want to do planning."
Most AI systems today also don't scale very well. AlphaZero, AlphaGo, and OpenAI Five leverage a type of programming known as reinforcement learning, in which an AI-controlled software agent learns to take actions in an environment -- a board game, for example, or a MOBA -- to maximize a reward. It's helpful to imagine a system of Skinner boxes, said Hinton in an interview with VentureBeat. Skinner boxes -- which derive their name from pioneering Harvard psychologist B. F. Skinner -- make use of operant conditioning to train subject animals to perform actions, such as pressing a lever, in response to stimuli, like a light or sound. When the subject performs a behavior correctly, they receive some form of reward, often in the form of food or water. The problem with reinforcement learning methods in AI research is that the reward signals tend to be "wimpy," Hinton said. In some environments, agents become stuck looking for patterns in random data -- the so-called "noisy TV problem."
Most AI systems today also don't scale very well. AlphaZero, AlphaGo, and OpenAI Five leverage a type of programming known as reinforcement learning, in which an AI-controlled software agent learns to take actions in an environment -- a board game, for example, or a MOBA -- to maximize a reward. It's helpful to imagine a system of Skinner boxes, said Hinton in an interview with VentureBeat. Skinner boxes -- which derive their name from pioneering Harvard psychologist B. F. Skinner -- make use of operant conditioning to train subject animals to perform actions, such as pressing a lever, in response to stimuli, like a light or sound. When the subject performs a behavior correctly, they receive some form of reward, often in the form of food or water. The problem with reinforcement learning methods in AI research is that the reward signals tend to be "wimpy," Hinton said. In some environments, agents become stuck looking for patterns in random data -- the so-called "noisy TV problem."
Intelligence does not exist in a vacuum. In order for intelligence to develop, system needs motivation to do so. (An engineer saying "you must be intelligent" is not sufficient, by the very nature of intelligence).
Basic motivation for all life on this planet is 1. avoidance of death, 2. self preservation and 3. continuation of own kind.
1. Avoidance of death and self-preservation require "pain" - this is a signal to the organism that something is happening that is hurting it and may result in death (hence - avoid)
2. Self-preservation and continuation of own kind require "pleasure" caused by consumption of food (thus extending own life) and procreation.
These stimuli and search for optimization thereof is what causes all development of thought and intelligence. By the very nature computer systems lack either. They cannot "die", nor "procreate". Thus they cannot even in principle have motivation to learn. A first step to a true AI would be a system that is actual danger of destruction in a hostile environment. Do that (10^very large value times) and may be we'll have a working cockroach.
Put simply - most of the "Artificial Intelligence" you hear about in the news is really fancy pattern matching. So you can have software that can recognize voice commands, or faces in pictures, or general patterns in data.
What you don't have, and aren't even close to, are computers that can "think." That is, put different sets of data together in arbitrary ways and make sense of it. You can't feed in a bunch of musical information to a computer and have it spontaneously generate music. You can't feed in a bunch of economic data and have it decide that certain regulations are required to achieve some economic goal - unless someone specifically programs it to do so.
The underlying reason is computers lack any way of attaining "common sense." If you tell a computer a person is in a room, the computer has no concept of what you are talking about but will dutifully note that a person is in a room. To a computer that could mean the person is occupying all the space in the room, that the person is in every room that exists, that the person is in the room AND outside the room, or that a person IS a room. In actuality, the computer makes no inference beyond "something called a person is in something called a room, whatever that means."
My Other Computer Is A Data General Nova III.
That is why I allow cats to walk over my keyboard. I just open up a Hex editor let my cats do the work. So far they have trained me to feed them, keep their food dishes full, and sit perfectly still on cold days.
I expect in 50 years, I will be coded to a level where I could sit there and watch the events judging everyone with disapproval.
If something is so important that you feel the need to post it on the internet... It probably isn't that important.
AI is here now. How many Chess and Grandmaster Go players are out of a job because of AI? All of them.
Read the headline. TFA is talking about General AI, which means broad human level capability in any field, not just in a single narrow field like Go.
We are no where near achieving General , or "strong", AI. Narrow, or "weak", AI is proving to be very useful for many tasks, but it is not clear if we are even on the right track to general AI. For instance, there is no evidence that the brain does "backprop", which is the core foundation of Deep Learning.
Put simply - most of the "Human Intelligence" you see is really fancy pattern matching as well.
That's a big part of it, but there's some "secret sauce" that lets organic brains combine patterns in new and different ways that AI researchers haven't been able to crack. Whatever it is, it's more than just matching patterns.
My Other Computer Is A Data General Nova III.
"In some environments, agents become stuck looking for patterns in random data -- the so-called 'noisy TV problem.'"
BF Skinner wrote another paper that might be relevant:
'SUPERSTITION' IN THE PIGEON
https://psychclassics.yorku.ca...
Until we have a proper definition for intelligence my pet rock qualifies.
Here is the proper definition of intelligence:
Intelligence: The ability to formulate an effective initial response to a novel situation.
Each word is important:
1. Intelligence is an "ability" not a mechanism. An entity that behaves intelligently is intelligent. internal mechanism is irrelevant.
2. Intelligence is the ability to "formulate" a plan, not to physically act on it.
3. A response is effective if meets an objective criteria.
4. It is the "initial" response that counts. Success achieved by a long term random process, including evolution over multiple generations, is not intelligence.
5. It is the response to "novel" situations that is the measure of intelligence. It is not just rote application of a solution that worked in the past. Memory and learning are important components of intelligence, but an intelligent entity can see how a past solution may or may not apply, and how to modify it for the new situation.
Your pet rock doesn't qualify.
Actual intelligence is pretty rare too.
Expert systems, deep learning, etc are all very useful tools and do work today.
“Common sense is not so common.” — Voltaire
Rocks tend to be a bit of a one trick pony. Their strategy is always to just sit and wait. It works pretty well for them though. They survive a long time. They're numerous, and not just on earth. Some of them even travel between star systems, something mankind has only dreamt of so far.
The primary problem is that we are unable to define what general intelligence is and therefore are unable to create it. We know it when we recognize it but we still can't define it.
The generic animal brain is composed of predefined structures which are all their own neural networks to it therefore it's fair to say that what is required is a neural network of specialized neural networks.
Anons need not reply. Questions end with a question mark.
You are a linguist but you don't know the word "novelty"?
Comment removed based on user account deletion
Trace-able machines may be more important than "instant" smart machines. If a bot decision is made that's wrong that has big consequences, society is going to want to know WHY the decision was made. Lawsuits will pile up if there's no trace-ability. This is both public lawsuits, and business-to-business lawsuits as claims made in contracts may be difficult to verify and/or quantify.
Trace-ability is why things like chains of Factor Tables (sig) appear more practical. DNN's are powerful, but are a dark grey box that's hard to dissect, debug, and understand. Factor tables may be harder to train, but offer better trace-ability and manual tuning by non-PhD's as a possible upside. And they are probably more modular than DNN's, as intermediate operations and templates can be plugged in as needed.
AI experts may set up the outline/framework, but "regular" office workers can study, trace, and tune the intermediate results using familiar tools that resemble and/or use spreadsheets, RDBMS, and statistical packages. Regiment-tize an otherwise dark grey art.
Table-ized A.I.
How often have you been a tree in order to test this? (I confess I often feel like a stump on Mondays.)
Anyhow, the entire "must mirror biology" is a dubious claim. It's like saying in order to make flying machines, you must mirror flapping wings.
There may be multiple paths to intelligence, not just Darwinian selection and survival-related emotions.
Table-ized A.I.
Sort of yes and sort of no. General AI is not a singular entity, just like the internet in toto is factually the most advanced AI on the planet, it works not on the basis of one program but many working together in their own speciality. As for what is though of as general AI, the error in design is a singular learning structure, rather than many interacting. So for language, not one AI but many working together each with specific roles and only those roles and an overarching AI that puts the solutions of each sub AI together. Any learning structure has to be broken down into it's smallest elements possible, with AI analysing and learning from just that element and passing them along for collation by the controlling AI. You always have to work to reduce the variables down to simplify and speed up the computational process. So for language, the initial analysis for one AI, how the words are arranged on the page (this is essential for sentences, paragraphs, first page or last page of a document or chapter, that is all that AI does and then is passes on that structure for more detail to be added in by other AI unites who process simultaneously, sentence structure, word definition (a cross correlation exercise based upon nearby words in the overall structure of words on the page). What makes it fun, speel cheker and grammor correktor because both having meaning and both r often wrong and we ruin internal korrecting to get. The not wrong meaning an AI that works with real world language has to deal with that.
So general AI is a misnomer, it will be always separate AIs working together, no different to our own brains, the perception of conscious is of one entity but that is entirely false, the point of collation is one overarching structure below which many AIs reside. Weird things are still required like a sense of humour, the means by which we break negative thought build up, a so called general AI will go insane without a sense of honour, it needs to be able to disrupt negatively reinforcing patterns.
Chaos - everything, everywhere, everywhen