--
Platform for AI
A modern approach to the AI problems is simplified, and as part of this simplification, the core resources of the further development progress already exhausted.
To give a substantial impetus to the movement, it needs a more complicated approach. There is a need to integrate scattered directions where some success has already been achieved and to add several components that are not being given attention.
The platform on the basis which we can find the links between disparate directions, I would define as follows.
1. The basic instincts of AI are not supposed to be the instincts of self-preservation and the continuation of the genus, and how the attention of Isaac Asimov, the basic instincts must be as priorities of obedience, protection of human, diligent foresight and precaution. If the machines think of themselves and their sexual behavior, they will become dangerous indeed for people.
2. The work and evolution of AI are focused on the real-time flow of real-world information and real-time management of real actuators. This work and management are based on the model of the surrounding space, including the robot itself, its actuators, sensors, computer, network with which the AI interacts, and the person who is authorized to manage this instance of AI.
3. The surrounding space model has four main sections:
3.1. Real-time objects.
3.2. Objects in the desired state (target designation).
3.3. Simulation section in accelerated time. In the future or in the past.
3.4. Section of decision making and management of executive devices, actuators.
4. Auxiliary sections and functions:
4.1. Meta-data generation by filtering, interpolation, scaling, neural network data transformations, spectral-temporal and other transformations, measurement of angles of orientation, illumination, distances and so on in order to abstract the discrete features of sensors associated with the sample rate, step size of the grid of pixels and the grid orientation, vibrating noises and that similar distortions, arising in an input stream of the information.
4.2. Recognition of meta-images for the real-time section maintenance.
4.3. A hierarchical system of priorities and principles on the top of which are the principles of Isaac Asimov: obedience, protection of human, diligent foresight and precaution. This subsystem is necessary for all the surrounding space model main sections management except real-time objects modeling. Real-time objects exist independent from anybody’s will.
4.4. Rendering metadata at the output of the decision-making section and other meta-data circulating in the system. This function is necessary for the monitoring of the system by the human supervisor. On the other hand, the rendering of meta-data with the deliberate addition of noise and distortion can be directed in the form of feedback on the input of the recognition subsystem in the course of testing or training. In this case, the real actuators and sensors are disconnected, and the system operates under the control of a supervising person in a state very similar to sleep. In a state of deep sleep, rendering generates detailed, long-time, picture-intensive images for detailed analysis and error correction. While waking, more simple “meta-renderings” will incur short pauses to figure out the next real command on a real actuator, otherwise, some small number of reflex motions are executed instantaneously without any modeling. Perhaps the reflex motions are good only for the wild jungle or in a situation of extreme danger.
4.5. Speech and text. Although some sources claim: “At first there was a word”, I believe that speech appeared much later than vision and only the words of people are convenient for us to understand. However, our machines should understand us and speak in a language that is understandable to us. Speech and text subsystems should be deeply integrated into the general AI system from the outset.
To summarize the vision of the platform, I want to pay attention to neural networks which I mentioned as an auxiliary DSP tool for meta-data generation. The self-organization of neural networks by evolutionary complication with respect to growing to some more substantial thing can take an unacceptably long time. At the same time, very powerfully developed graphics modeling systems presented in areas like VR, AR, CAD, and so on very poorly involved in the development of AI. This disproportion will be fixed on the proposed platform.
