Baddley's model of working memory describes short-term memory as being federated into the sensory modalities that humans normally interact with: visual spatial, phonological, and episodic. However, if an organism possessed a very different sensory modality, it would expect the working memories to be federated into that additional sensory modality category as well. This suggests that reinforcement learning agents might structure their internal information processing after the sensory modality is and structure that the environment provides of the external information processing. It's suggests that the environment itself acts as a structural constraint to the nature of information processing realized in a reinforcement learning agent cognitive system.
Information processing mechanisms in the brain interplay is over in a spectra frequencies, from the slow drifting genetic and anatomical changes, to faster synaptic modification, to the very fast spike frequency spectrum of activations. Cross frequency phase correlations exist through the frequency product domain
I feel it is important to explain what you mean by learning. Biologicalinformation processing - weather realized realized in genetic, anatomical, or physiological mechanisms - interplay over a spectra of frequencies. Taking the dynamical system perspective, we might say these mechanisms exhibit cross-frequency phase modulation.
Towards principle of artificial intelligence
The problem is generally impossible. Our universe can only display a subset of all conceivable problem domains.
This problem is not just theoretical
Still, The historical response to this problem has not been to simply give up.
In this paper, I want to exemplify a general methodology that may be applied to tackling complex problem domains using artificial systems: 1) starting by taking inspiration from existing natural and artificial intelligence, 2)
What is intelligence? "Intelligence makes definitions dissolve into descriptions." In the section we identify some of its characteristic descriptions - principles of intelligence - and discuss how they may be applied in an artificial system.
There are many other principles of intelligence that guide in the development which I am too afraid to speculate on.
In the conclusion of the paper,
We began this paper by briefly explaining that it serves as an example of a general methodology to tackle complex problem domains. (Sec:1) The human author first considered existing natural and artificial intelligence and extracted principles of intelligence. (Sec:1) These principles were then encoded in a computational system while leaving many variables open for Machine optimization. Individual artificial systems were then progressively integrated into an increasingly complex composite system as guided by experimental results. (Sec:3)
That is the principle algorithm behind general intelligence, and this paper and pai-0 are concrete realizations of the process and product respectively. This algorithm may be expressed in biological, human, or computational processes (in this paper by a combination of human and computational processes), but in any case, the instrumental result is a critically complex artificial system that serves as a stepping stone to continue solving the Problem.
In the principles of intelligence section, I should use a lot of black quotes and then make some brief summaries. Maybe actually only keep a few short quotes and then move more longer quotes to an appendix
I need to find good resources to make the point that complexity is not just spreading out but going deep. Shallow versus deep networks. Instinctual versus deliberated response. In humans the correlatable response trajectory is too long to attribute its too external stimulus
The brain and entropy during aging
The entropic Brian (2014) maximize entropy during conscious states
The brain from inside out 2019