Re: a "separate non-trivial task of converting a gigantic stream of numbers into a reasonably compact structured description", this is called subitizing. It is an operation that can be performed by imposing a known or discovered structural pattern on an aggregation of objects. It can be implemented in computers at a level that far exceeds human capability.
The difficulty is that real patterns are part of the environment, but the system receives a set of pixels that play the role of an intermediate representation. And this intermediate representation as a set of numbers does not contain the desired patterns.
Discovery and representation of natural structure is my main current interest. I'm playing with measures of information entropy as a simple and single criterion for hierarchical 'chunking' the firehose of sensory input into static and dynamic representations (objects and events).
Things like information, knowledge and entropy are constructs that cannot be directly found in the physical world, existing separately as patterns of structures in a meta-world (i.e. the Mental World). They can be encoded by patterns in the physical world. We describe these patterns using external 'languages' to these worlds. We have to go to higher levels of abstraction (meta-meta worlds) to describe the functions that structure (form) patterns, then use 'languages' to describe objects, structures, behaviors and morphism in those worlds as well.
But these concepts are needed for the exchange of knowledge between people.
In the AI/AGI system, algorithms operate with numbers, functions, vectors, geometric and physical concepts that reflect the structure of the material environment.
Just for my understanding. The pixels are the intermediate mapping/ representation of the numbers, functions, vectors, etc. used in models and the challenge you are describing is how to optimize the mapping/ efficiency of the process for the datageneration?
The process of building a model using the data stream from the camera is based on the search (selection) of parameterized invariants and uses our knowledge of the physical world (geometry, time, kinematics). This is briefly described in
and will be described in more detail in some future chapter.
The difficulty lies not in optimizing the known ways of working with data from a video camera, but in building algorithms that can do what people and animals can do: detect objects unknown in advance and track their movement.
Re: a "separate non-trivial task of converting a gigantic stream of numbers into a reasonably compact structured description", this is called subitizing. It is an operation that can be performed by imposing a known or discovered structural pattern on an aggregation of objects. It can be implemented in computers at a level that far exceeds human capability.
Right.
The difficulty is that real patterns are part of the environment, but the system receives a set of pixels that play the role of an intermediate representation. And this intermediate representation as a set of numbers does not contain the desired patterns.
Discovery and representation of natural structure is my main current interest. I'm playing with measures of information entropy as a simple and single criterion for hierarchical 'chunking' the firehose of sensory input into static and dynamic representations (objects and events).
Things like information, knowledge and entropy are constructs that cannot be directly found in the physical world, existing separately as patterns of structures in a meta-world (i.e. the Mental World). They can be encoded by patterns in the physical world. We describe these patterns using external 'languages' to these worlds. We have to go to higher levels of abstraction (meta-meta worlds) to describe the functions that structure (form) patterns, then use 'languages' to describe objects, structures, behaviors and morphism in those worlds as well.
Right.
But these concepts are needed for the exchange of knowledge between people.
In the AI/AGI system, algorithms operate with numbers, functions, vectors, geometric and physical concepts that reflect the structure of the material environment.
Just for my understanding. The pixels are the intermediate mapping/ representation of the numbers, functions, vectors, etc. used in models and the challenge you are describing is how to optimize the mapping/ efficiency of the process for the datageneration?
The process of building a model using the data stream from the camera is based on the search (selection) of parameterized invariants and uses our knowledge of the physical world (geometry, time, kinematics). This is briefly described in
https://agieng.substack.com/p/agi-structuring-the-observable,
and will be described in more detail in some future chapter.
The difficulty lies not in optimizing the known ways of working with data from a video camera, but in building algorithms that can do what people and animals can do: detect objects unknown in advance and track their movement.