Stephen Thaler made headlines last year when courts around the world began to rule on whether his AI system could be named as the inventor of a patent. (no.)
But as one of the pioneers of the AI industry, Thaler has had a fascinating and storied career. He recently shared some highlights in his talk online at his March 1st meetup for the Chicago and Washington, DC chapters of the Association for Computing Machinery (ACM). Thaler told the audience that for the first time in the 1970s he played with a neural network and created an alternate version of the “Have a nice day” smiley face.
By 1995 Thaler had founded Imagination Engines Incorporated. This is a pioneering AI company that claims to have mapped the act of thinking itself onto a system of neural networks.
Thaler’s company eventually patented the idea of using artificial neural networks for “noise-driven brainstorming sessions,” according to Thaler’s introduction to ACM. Even creative robots. But then the company tried to push the knowledge domains to join together.More specifically, they result It uses chains of interconnected neural modules to trigger simulations of the human mind’s positive responses, selectively enhancing the most influential ideas.
The ACM preface states that the end result is “arguably the subjective feelings (senses) of conscious machine intelligence,” and that the valuable ideas formed through the computational process are “a long-standing study of biological intelligence and personality.” It challenges beliefs,” he said.
Sure enough, by the end of the talk, Thaler was explaining how to build a Sentient artificial general intelligence, and it all sounded pretty simple. But along the way, his audience learned an awful lot about how the human brain works, about thought itself, and ultimately about ourselves.
sensory case
Thaler began by casually emphasizing that he does not believe the claims that our current large language model is sentient. “But what I’m going to talk about teeth “I make the right claims.”
Thaler’s company built a system called DABUS. DABUS is an acronym that stands for “Device for Integrated Sensory Autonomous Bootstrap”. Thaler says that the company’s DABUS system is not just one algorithm, but an entire system with both computerized and electro-optical components, “each subsystem being governed by its own specific algorithm. It was revealed that there is Thaler argued that “it’s like a brain, with subsystems and algorithms working within each.”
But first, Thaler describes some early experiments with neural networks in the 1970s. While fixing the input at a constant starting value, he also steadily increased the ‘injected perturbation’ (later called ‘internal noise’) and tracked the resulting increase in patterns or ‘concepts’. Did. As the noise increased, even the patterns tended to become “generic patterns, often nonsense”, and eventually the system was unable to even generate patterns. We call it the deer’ regime,” he said. “The whole neural network seems to be flooded with the equivalent of adrenaline in the cerebral cortex, and I can’t think of anything else.”
However, the experiment also identified the ‘Goldilocks Zone’. This zone provided the optimum level for optimal pattern generation. So Thaler hooked up his second neural network, making the first act as an “imagination engine” and the second to enhance the best ideas (both as a filter of idea utility and a kind of memory). function). Thaler called the resulting system a “creativity machine”. It’s an associative memory implemented with a kind of neural network that stores the ability to replicate a given set of input patterns. “What you’re doing here isn’t what I call creativity,” Thaler said later in his talk. “Essentially, it’s an ongoing optimization process to find the overall best solution, but we’re not really combining the conceptual space to come up with a whole new idea.”
Simulation of results
The shortcomings of this process prompted further experiments in trying to make neural networks aware of the ‘core idea’ and ‘result’ chains, as well as the so-called ‘hot buttons’ that affect the entire system. For humans, it may be “existential”, such as getting nourishment and survival. “We all have these hot buttons at birth,” Thaler told the audience, joking, “They’re factory-installed.” Importantly, however, the human perception of a “hot button” triggers the release of neurotransmitters.
To simulate these, Thaler used the same perturbations in his system. “When a hot button is activated, it essentially secretes a simulation of the appropriate neurotransmitters.” The idea is to reinforce the chain that triggered the hot button. And later you can reuse some of these result chains. ”
DABUS is a large array of such memories. “The training sessions basically read information, talk to them, show them pictures, show them all kinds of media,” he says.
Is it all really that easy? “What you end up with is a circus of a thousand wheels,” he told the audience.[with] Many memories are formed, latent ideas that are nothing more than temporary noise in the system. In addition to the old ideas that form the basic idea, there are result chains and hot buttons. And there are new ideas that are completely forming the resulting chain. ”
Clearly, DABUS wants to distinguish between new and old ideas. As Thaler explains, “Essentially the whole picture, the entire neural landscape, goes through a large auto-associative network that is constantly being trained.” We filter out past sequences that weren’t particularly novel, and highlight the new ones — we identify those that appear as anomalies when compared to the longer history. , basically you have to shoot different combinations of chains at once. The resulting image (in JPEG format) creates a useful “condensation of all this activity” that Thaler passes to other machines.
So where does this lead? The combinations just keep getting bigger and bigger. Some may be dead ends (or circular paths that just go back to where they came from) and some may be gradual growth like plants. Thaler says it looks like the way molecules form and even end up in specialized subgroups. But in a way, Thalar argues, things end up being defined by the functionality of each component, which in the end isn’t so abstract. It is, in essence, the formation of thoughts. “The entire conceptual space is compressed into associative memory…concepts and strategies are represented as forms of linked associative memory.”
And, as a key mimic of human reinforcement, “when the hot button is activated, a simulation of the appropriate neurotransmitters is essentially secreted.” Thaler argues that in the end it was “a crude simulation of consciousness—because the noise produced the diversion of various pattern-based ideas.” Thaler says it’s similar to how the human brain works. “We may not have any input from our environment, but we are relaxing and basically observing the flow of ideas that run through our heads. I would argue — I have many papers along those lines, but essentially the entire stream of consciousness has the same fractal distribution as the human stream of consciousness.
Thaler sees a pattern of “result” chains that activate hot buttons. and hot button). And they ultimately equate to sharp instincts, or synthetic emotions, he believes.
Thaler’s DABUS system, “assessing the stream of consciousness through such sensations,” explains one slide, “matures certain concepts while culling others.”
“All these results show that awareness, sensation, and cognition are all combined to do what they do,” Thaler told the audience. Feelings drive the formation of ideas.
“And it’s all done in the non-protoplasmic part,” Thaler adds as an aside. “Thus, it may turn out that sensation and consciousness are not confined to the protoplasm.”
beyond the senses
Thaler has worked on several of his other projects, including a financial application that monitors and classifies trajectories of various stock markets, and a 2006 AI-devised music CD titled “Song of the Neurons.” It started right away. Thaler also showed off a fractal-shaped beverage container devised by the system. (“For a cylindrical soup can, for example, it has about three times the area of its equivalent volume. It has fast heat transfer and is easy to grasp by a human hand or a robot his handler.”)
In 2022, his company will begin working on medical applications.
When asked what it would be like to port this system to a robot, Thaler said: However, for some reason, we were unable to raise funds beyond Phase 2. ”
“But yes, it is a fascinating area to get into. The problem is, I am drowning. I am drowning in technology. We have too many patents to get.”
Here is the full video:
