They simply start the AI system, supply it a stream of incoming data, in the form of pixels, like an old Atari game. The AI system, is first confronted with a series of events. It soon learns the rules of the game. And then becomes a master at the game. Learning strategy, optimal methods to achieve goals faster and can outperform most humans.
It uses a simple set of rules. Observe a set of environmental behaviors, learn, determine optimal outcomes based on specific possible outcomes, and continue the cycle.
Observe + Learn + Decide + Action then repeat.
Some of the advancements recently have been to add read/write memory to Recurrent Neural Networks. What that means is the traditional Neural Networks could learn over time through algorithms, they now have the ability to remember prior knowledge over time, retained in memory, for future use.
This is known as the "Neural Turing Machine" in which an Artificial Neural Network can learn new algorithms by example. For instance, you show the system a routine to sort numbers, and than ask it to perform the function and it knows how simply by watching it previously, actually bypassing the need and time to train the model.
Here's a tweet I posted earlier today: "#AI is a self learning technique by observing, determining possible outcomes, selecting best alternative, then observing... #AGI holy grail."
And another: ""process that converts unstructured data or information into useful actionable knowledge"=AGI, tech is neutral, requires ethical constraints"
Most of this information was lifted from a YouTube video from the Deep Mind founder Demis Hassabis, you can watch the video here
I think the best part of the video is where he differentiates between true Artificial General Intelligence and Narrow Artificial Intelligence. Siri, Cortana and Big Blue are considered Narrow. What Deep Mind is focused on is AGI. A subject that's been around a very long time. With lots of smart people working towards the goal. Although it seems to be a hard nut to crack.
I recommend watching the full video to get insight into the happenings at Google Artificial Intelligence team. I followed Deep Mind before they were acquired by Google and they have a tremendous pool of talent in the space of AI.
The interesting thing about this field, is although it's deeply rooted in technology, they depend on outside science for assistance, like psychologist, neural research for brain activity to behavioral of humans and a bunch more.
This field is pretty exciting. Although most of my knowledge is self taught, I'm not an expert on the underlying technology that makes it work. I do understand the base concepts of what a Neural Network is, a Recurrent Neural Network (has memory) to Models and some of the different methods for training them.
It seems some of the latest challenges are applying domain knowledge across multiple domains without the need to train the models for long periods of time. And perhaps adding memory to the model is one way to accomplish this.
Another factor is the majority of the work thus far has been in a virtual environment. They have not adapted the AI technology to robots or external environments, where the number of potential conflicting input streams is potentially unknown.
How would a robot handle a barking dog, or a plane flying overhead or the mailman driving by, all unknown events that don't happen consistently over time. So the AGI system wouldn't have prior knowledge so it would take to to learn and adapt. Unless you could simply plug in an algorithm to the system so it understands and adapts immediately.
When we think of ethical concerns, Mr. Hassabis indicates that the technology is neutral or independent. It depends on how the software is used.
An analogy I think about it the speedometer in a car. Sure they could set a control to cap the maximum speed of a car. Because technically, the speed limit is typically 70 MPH in the United States, so why would a car have a need to exceed that limit? If the driver exceeds the limit, he or she is breaking the law technically. Why don't the car manufacturer's place a restriction on the vehicle to prevent us from speeding? Perhaps that's an ethical question.
Could the car send a signal to the patrol station, which electronically generates a citation sent in an email, where an ACH withdraws occurs the following business day, removing a $100 fine from your checking account? I'm sure that could be done. But it's not. Why is that.
How would you create the algorithm to select which drivers to give a citation, or would you apply a blanket rule that everyone exceeding 80 MPH would get a citation, if that were the case, why not place a cap on the accelerator of the car not to exceed 80 MPH. They could do that today without an AGI system.
I suppose there's some flexibility baked into the system. Perhaps you need to accelerate to 75 to overtake a car in the slow lane, multiple possible reasons. But the level of enforcement is by human interpretation and perhaps not applied evenly. In fact, I'm not sure how they know to pull over certain speeders and not others. A mystery. But it could be automated. But would we want it to be?
What level of automation do we want for our future AGI systems? Who gets to decide what is deemed ethical and what isn't. How do you handle the exceptions and the gray areas?
Stay tuned to find out.