Just a quick dump of thoughts I can refer to and reduce repetition in conversation. The books: Scale, EoA, Pib.

Sometime by the end of 2017 I picked up a tweet pushing Geoffrey West’s 2017 book: Scale - the universal laws of life and death in organisms, cities, and companies. This immediately rang with me by promising hot food regarding complexity, long-tailed distributions, failure and growth. By that time I was already hyper-alert about growth and pieces of growth related theory, based on its role in development and learning (Refs: Pask, 1959, Physical analogues to the growth of a concept). Current working hypothesis is, that open-ended learning processes can be seen as growth processes. A theory of growth should allow predictions about the preconditions, expected learning speed, expected learning time and upper bounds of learning. This is of course highly relevant when building curious, self-aware, exploration agents and robots. Overt physical growth appears to occur in the three-dimensional space of intuitive experience. Alas, intuitive experience is itself an emergent phenomenon of accumulated growth, and thus is probably significantly mistaken about the character of the natural information space of any given instance of a growth process. Many of you will recognize this as another way of stating the manifold hypothesis. In summary, this allows to model learning as the growth of explanation, obtained through the accumulated information “precipitating” from an active search process (aka stochastic optimization). Growth is provided energy in terms of different resources from an underlying implementation fabric (power, memory, time constraints); it is driven by a need of explanation of experience, sufficient with respect to a given motivation; implemented by equating explanation of experience with actions, and by observing their adequacy; adequacy is estimated via different exploration stratgies; finally growth is limited by finite time horizons, uncertainty, inherent total complexity, and again, resources. The explanation-action equality is one of several useful consequences of the predictive brain hypothesis.

The book has a lot of fundamentals, examples, and pointers to more detailed material on all of this. Main research questions I’m taking home from the angle of developmental learning and robots are: how can the properties of space-filling-ness and impedance be brought to work on information flow in dynamic inferential circuits (aka adaptive models); and what is the relation of these properties to modularity in model space. Obviously module types and available sizes should support impedance matching, coverage, learning speed, and scaling, besides reuse.

Apart from this specific research perspective, on the basis of my individual knowledge the book was smooth to read, and provided a huge amount of highly informative reaffirmations, additional perspectives and entirely novel concepts with immediate practical applicability in everyday social, ecological, and ultimately psychological questions, e.g. urban life, mass phenomena, one’s own resources, teams and management, communication and marketing, attention, information spread, opinions, and so on, scaling.

Very happy having finally ingested a long time resident of my to-read stack, Georg Franck, 1998, Die Ökonomie der Aufmerksamkeit, Hanser, originally recommended to me by Otto Rössler in 2012. The book seems to be available only in German on paper, but googling the ad-hoc translation of the title should reveal a few english language articles, which are much more compact than the book, so you’re lucky. Interestingly, the meaning of the German “Aufmerksamkeit” has no direct correspondence in English and falls apart into the two distinct terms of attention and awareness. Attention emphasizes the outward direction and the objects of the process, while awareness encodes the inward direction and the subject of the process.

Franck is a rare proponent of the radical privateness assumption about subjective experience, probably necessary for effective self-models and theories of mind in embodied agents and real world robots. This is based on an extended version of Leibniz’s monads and implies, that there is no way in principle for any agent to obtain the ground truth about any other agent’s experience, ever. All information about the experience of another has to be channelled through behaviour.

After getting attuned to the idiosyncracies, the book really takes off and turns out to be written in beautifully precise and funny German prose, imho. Again for me, it provided two essential assets to take immediately. The first one is a thourough and explicit analysis of the flow of attention in agent interaction, creating a powerful explanation that simplifies several classically disparate phenomena by the unifying them in a single framework by leveraging economical methods of the analysis of behaviour. A welcome side effect of this is a reference refutation of rational agents in classical game theory and mainstream economic theory, that continues to fail so superbly over the consequences of ignoring issues of the rational agent assumption. Rationality can only be optimal in bounded linear domains. In all others, where observability, dimensionality and non-linearity combine to set a bound on the principal predictability, irrational moves, more often called randomness, are more often than not beneficial the agent’s long-term well-being. For in-depth disussions of the limits of naive application of the optimality principle, see for example Loeb, 2012, Optimal isn’t good enough; Iigaya et al, 2017, Learning fast and slow: deviations from the matching law can reflect an optimal strategy under uncertainty; tbc.

A lot of practical advice, guidance and inspiration can be found on the tactical and strategic level valid in almost all domains requiring systematic communication on any level. In particular, it includes an analysis of the self-referential level of motivation and self-storytelling, and extends upwards to all levels of social relationships from dyadic, families, extended families, teams, teams of teams, to supercritical phenomena like mass media and large scale social networks.

A recurring example and workhorse for the theory is the science it self. I would expect that anyone involved with science and its orthodox implementations on a professional level will shop one or more insights into the processes of the science market, allowing new kinds of power moves. Or so. Anyway, need to run, tbc.

Comments

Footnotes

1fn.1

tag cloud

robotics music research books AI sound intelligence feed ethology computational psychology startups jcl audio models brain organization motivation micro management jetpack funding dsp algorithms testing test synthesis sonfication smp scope risk principles musician motion mapping gt generative fail exploration evolution epistemology digital decision datadriven computing computer complexity aesthetics wyafinding visualization tools theory temporal sustainability stuff sonic-art sonic-ambience society signal-processing sequence self score robots robot-learning robot python pxp priors predictive policies philosophy perception organization-of-behavior open-world open-culture neuroscience networking network navigation minecraft midi measures math locomotion linux learning kpi internet init health hacker growth grounding graphical gaming games explanation event-representation embedding economy discrete development definitions cyberspace culture creativity computer-music compmus cognition business birds biology bio-inspiration android agents action