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.
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.