Thread by @bayesianboy, Thread. Should you work on or with the FEP? The answer depends [...]

Thread. Should you work on or with the FEP? The answer depends on a number of factors, and I believe that what discipline you are situated in is not one of them. The FEP is an interdisciplinary project. It rests on input from mathematicians, physicists, machine learning https://twitter.com/LitvakVladimir/status/1334236348912427014

https://twitter.com/LitvakVladimir/status/1334236348912427014

architects, psychologists, psychiatrists, cognitive scientists, neuroscientists, and biologists, among others.

There are three essential modes of involvement with the FEP:

1. Development. Uncovering interesting formal relations and making the fundamental formal architecture

of the framework water-tight. Could we render the FEP without the assumption of ergodicity? Could we adapt the Markov blanket formalism to encompass something like Birkhoff's wandering sets? Does the condition of time-synchronous conditional independence make sense for the sort

of systems we are interested in?

2. Application. Conceptual investigation of new problems and domains by way of the FEP. How can we encapsulate schizophrenia under the FEP? What might we learn about the evolution of multicellularity if we rendered it as a matter of stabilising

informational dynamics across hierarchical levels?

3. Implementation. Implementation of the FEP has two components: one in terms of technological development, the other in terms of empirical work, although these are tightly interwoven. At the moment, Alec Tschantz is working on

a python implementation. This is where we put the FEP to the test as an approach to machine learning, or in simulation work, which ultimately will be put into contact with data.

I have three cautionary remarks for those who are interested in the FEP:

A. If you are doing (1),

that is, tinkering with the fundamental formalism of the FEP, you are working on the FEP. If you are doing (2) or (3), that is, applications or implementations, you are working with the FEP. In other words, you are employing the FEP in the generation of a final product that is

distinct from the FEP. A model of attention or symbiosis constructed with the formal tools of the FEP is not identical to the FEP. When we implement the FEP down into a programme or simulation, the programme or simulation we produce will not be identical to the FEP. This is

related to the normative model/process model distinction from Luce (1995). Once we stipulate the form of a generative model leveraged by some system, we have switched from the domain of the FEP, a normative model, to that of an associated process model. There has been a

widespread conflation of the FEP with its associated process models. People who are doing conceptual work with the FEP often impute the FEP itself with the theoretical implications of whatever sort of theories or models they can construct with it. The FEP itself has no

theoretical implications.

B. The FEP is a formal framework. You should not delve into the FEP if you are unwilling or unable to grapple with the maths. This does not mean that you need to be a mathematician or a physicist to work on the FEP. But you should have an interest in,

and dedication to, getting into the weeds of the maths. A lot of people these days seem to think that they can pull philosophical (or otherwise) conclusions from the FEP without grappling with the underlying mathematics. This has resulted in a literature in which elaborate claims

are made about the FEP—or about perennial empirical or philosophical questions, in virtue of the FEP—which have nothing whatsoever to do with the FEP. This is deeply unfortunate, and should be avoided at all costs.

C. The value of the FEP will ultimately be determined when we

toss it in the water and see if it floats; that is, in putting it to use. In its rather short lifespan, the FEP has followed a trajectory that, from the standpoint of sociology of science, is really quite remarkable. A great deal of pageantry has accompanied the development and

exposition of the framework. This has had the regrettable effect of obfuscating what exactly it is that the FEP is or does or is intended to do.

Ultimately, you should pursue a model if it inspires you, if you see potential in it, if it calls to you as a way to relate to the

world, as a way to relate to data, even as a sort of scaffold for your own private theorising. What generic, targetless formal models like the FEP offer us is an inroad to otherwise prohibitively complex natural phenomena or theoretical problems.

Problems we've been beating our

heads against for decades, or even centuries, like the evolution of multicellularity, attention, senescence, autism spectrum disorder, post-traumatic stress disorder, morphogenesis, or consciousness. There's no tackling these problems head-on, so our best shot at unraveling the

mystery is really just repeated experimentation with the form that a solution might take.

I came at the FEP wanting to debunk it. And I did debunk it; at least, what I had believed it to be. But in doing so, in rolling up my sleeves and pulling it apart to see how it worked, I

found that I had adopted it as a tool for generating insight into the problems that had long evaded me.

Autism Spectrum Disorder became a miscalibrated mechanism for weighting error signals against top-down predictions. Post-Traumatic Stress Disorder became an irreconcilable

mismatch between some horrifying piece of evidence and our deeply-ingrained models of who we are and how the world works. Multicellular senescence became a necessary tradeoff between the precision of evidencing behaviours at cellular and hologenic levels. Cancer became the

kickback at the cellular level. Biological functions became emergent directives for multiscale error minimisation.

I am excited to see these ideas played out in computational models. Down the road, I hope to see the results of such models weighed against empirical data.

As the vehicle of science comes to aim the beams of its headlights at ever more complex phenomena, and as the raw power of our computational machinery increases, I think methods like the FEP are going to become ever more prevalent.

But we don't have much of a precedent to go on

in evaluating these frameworks. The philosophy of science is going to have to pick up the pace if it hopes to catch up.

Then again, with methods like the FEP, you're seeing disciplinary boundaries break down. The boundary between science and philosophy, between industry and

scholarship, between hard and soft sciences becomes meaningless for this kind of work. Indeed, the success of the method rests fundamentally on its practitioners wearing multiple hats and working in tight collaboration with individuals from other fields and sectors.

And I hold this to be a necessary development.

Thank you for coming to my ted talk.

@threader_app compile

Latest Threads Unrolled: