Controlled Application Spaces (TM) are a way of putting your application (system, robot, drone, machine, program, vehicle, etc.) into action under "controlled" conditions, but with full exposure to any interface that generates data, so that its background machine learning algorithm gets the most of it.
It´s all about data
Any application that has an intelligent algorithm at its core is likely to be deployed in an environment where it will continue to acquire data of its surroundings as well as any interaction, spatial and timely coordinates or behavioural characteristis it may be able to capture.
For example: a limb-prosthetics. There are four main types of prosthetics. These are known as transradial, transhumeral, transtibial, and transfemoral prosthetics. Each prosthetic serves a different function depending on what body part was amputated. Any AI supported prosthetics would be trained with the patient to yield the function necessary in the most convenient way to the patient. Yet, still it can improve. These prosthetics therefore not only perform a function, they also are tiny little applications that record signals from the nervous system, pressures and movements as well as the details of reactions to an executed function, like a push upwards.
Similarly, any surveillance system will have learned to identify body-parts of a person, recognize elements of your face and match movement, posture and general physical attributes to search terms for its database. Most importantly it will have learned, to recognize those combinations of beforementioned characteristics that are statistically coming with a high probability towards a certain outcome, for example a violant behaviour being shown. While these systems have basic "expericence", they will continue to learn on the job.
A widely known application is the self-driving car, although there are not too many already "on the road". Yet, the media has given this upcoming technology enough room, so that its use and usefullness can be considered public knowledge. Some people are eager to possess such a car. Others are sceptical in regards to safety, reliability and accountability for any possible lawsuit, should this car not function as promised. The self-driving car is a prominent example for a technology not fully trusted. Many people know and understand the decision making problem, that the underlying AI might face, should it have to evade in favor of either an elder couple or a woman with a child. It is also understood that these self-driving cars will learn how to react best, while they are "on the road". The product will mature while being used by the customer.
Any application that is based on machine learning or artificial intelligence of any kind will need to continue to learn, in order to reach its full potential. The question is: at what level of maturity will your product be ready to ship and safe enough to not cause any harm? The simple logic goes like this: the more data-sets have been used to train the algorithm, the higher its capabilities. That is not to say that anyone of us can even say, how many data-sets are needed in the first place. We, as humans, must admit, that whatever an algorithm learns, is at some point beyond our comprehension. We can only sit back and score the results
"Wouldn´t it be great to be able to generate enough data to train your algorithms so that you can increase your confidence in it to a more comfortable level?"
Controlled Application Spaces do just that.
Controlled Spaces to train your application
Imagine a drone manufacture whos intention is to deploy the drones as delivery drones in rural areas with people walking around, kids kiting, trees growing and at holiday season christmas lights spun across the street. One could easily think of several ways a desaster can happen, even with only so few "obstacles". Rest assured, there a plenty more.
A manufacturer of such an application, that can only be tested fully in an environment that presents "real challenges", needs a space for training, testing and calibrating. The same is true for any robot that is soon to be deployed in room together with humans. Co-Botics will soon be at a plant near you too. These robots have enough sensors to recognize the nearness of a human being. They also have enough programs to react to any approaching human. But any preset system is only as smart as its creator - and unfortunately we aren´t capable of preparing possible situation, let alone program this in some code. There a simply too many possible situations: humans coming from all sides, not just one, but many, with different speed and gesture - what should the robot do.
Here´s another example: Not long ago we were dealing with voice-controlled answering machines. Today we see chat bots on the rise in almost every situation, where communication between humans and a machine can be automated. Natural Language Processing, a specialisation of artificial intelligence, has made tremendous progress and is now at a level, where it will soon be deployed widely. How did this happen? Well, in the beginning, there were simple chatbots, which only offered predefined answers to a set of selected questions - not very smart. But then, algorithms were introduced, that would learn from the communication. A question that was answered and subsequently evaluated as "helpful" was like a "good" neuron added to the neural network of the application - a useful experience, so to speak. After a while such systems had had numerous conversations and the amount of good experiences had increased. Today, the vast amount of experiences from wich any language processing application can draw is enough to train algorithms so they can almost communcicate like a human.
Therein lies the beauty of machine learning. Rather than programming a lengthy spaghetti-like code that aims to consider all possible situations, in machine learning and also in deep learning or reinforcement learning, the machine builds a neural network (think of it as a database of neurons) that stores each and every situational aspect as well as possible reactions and an evaluation of its outcome. If you train a system long enough in an environment that offers random situations, the system will learn by simply gaining "experience".
A Controlled Application Space aims to provide the necessary setting for an application to get the "needed" training. Whether it is a production plant with robots, a concealed area with routes for drones to fly or an interface for commmunicating with a human being, a Controlled Application Space will offer whatever it needs to generate the right data.
The value of Human Interaction
We can´t even begin to imagine the value of any interaction, when it is evaluated as a data-set for an AI to learn from. Yes, a self driving car does not need a lot of interaction with humans, but still, if it happens, it is either the officer who asks for license and identification or a human being in front of the car. These may not happen so often, but when it does, it is most critical what the AI will do. Bottomline, we may think that we can program all situation that pertain to the mechanical, procedural or transitional situations in our lives, but when it comes down to learning how to interact with humans, there is no way of programming that in a typical sw-development approach.
How to interact with humans can only be learned by the AI, it cannot be programmed. It is therefor inevitable that we create environments, where these machines can learn, but without any harm to any human being.
Controlled Application Spaces (CAS) are hard (if not impossible) to find. The specifications typically require machine-human interactions that could result in damage to the machine or harm to the human being. Clearly though, it is more likely that the needed data-sets can be generated in an envrionment that is openly designated as a Controlled Application Space, than in a reduced and covert scenario, where the minimization of risk cripples the training.
While it seems difficult to find such a Controlled Applications Space, it is always possible to build one from scratch. Too costly, too risky, too timely, ... you think? Well, with a traditional perspective this may be true. But let´s think outside the box. What if it were possible to combine several applications in one CAS. CAS don´t need to be limited to just one application. It is very well possible to train a robot that prepares hamburgers and at the same CAS train the robot that serves and communcates with the person ordering. It is well possible to create a CAS that offers a training area, just like a little village, to train drones and also self-driving cars. Who is to say, that both trainings are counterproductive? They migh just as well be adding up.
Think of a Controlled Application Space as your stage. You will design the stage, the choreographie, the story. You will decide on the players, the action and the timing. All constraints, all limits but also all open decisions will be yours to specify. Imagine that you can create the training stage for your application.
Get your Controlled Application Space up and running
Controlled Application Spaces wil be costly and most certainly unique. But given a bit of creativity and entrepreneurial wit, we might be able to devise a CAS project that not only creates highly usefull training data, but does not cost you any more than you would be willing to pay for the data-sets anyway.
The idea of a Controlled Application Space does not have to be confined to a high security set, where no-one has access to. On the contrary. The more average human beings are involved, the better. We like to propose, to think of Controlled Application Spaces as spaces that are integrated in everyday life, that are part of a community or even a routine stop for travellers.
Controlled Application Spaces might even be based on business that generate a profit.
If you are looking for a place to generate data-sets for your AI application, talk to us.
We will be happy to build your Controlled Application Space.