Again and again, people like Bernie have reminded me that success or failure in an entire field often depends on whether we have enough people who REMEMBER, who fully appreciate and apply, a few very basic principles they thought they already knew,.
1. The Basic idea and Overview
For example, in late 1979, when I just started my first tenured job, at the US Department of Energy, I soon resolved never to forget a great little poster I saw hanging on someone's wall, illustrating that old adage (which we need to remember and apply):"When you are up to your ... in alligators, it's hard to remember you came here to drain the swamp."
IMPLICIT in that important old saying is the idea that "you" were trying to accomplish something, and in danger of forgetting what it was. That "something" could be formulated as a goal or as a target variable. It turns out that TARGET VARIABLES are an important mathematical concept which we need to fully conscious of, as we approach many challenges, such as modeling coronavirus, economies, design of future internet, and increasing the degree of sanity of ourselves and even of the noosphere. MOST modern apps and integration platforms for the internet do not give target variables the importance they deserve, important to building connections between the low practical input variable level and the globalutility functions U which I have written about before here.
2. Target Variables in Neuroscience and Neural Networks
(True "AGI" is a type of neural network design.)
Since the brain science aspects are part of what key people in this discussion study, let me mention how target variables show up there.
In 2005, the IJCNN (International Joint Conference on Neural Networks) in Montreal that year, invited a leading-edge neuroscience researcher whose talk I summarize: "I have studied the great work of Goldman-Rakic on what the higher cerebral cortex of humans really does. I have fund that she was missing something, in trying to understand the real purpose and function of the very highest regions of the frontal lobes that cortex, the dorsolateral and... These two regions represent the cutting edge of human biological evolution, what our species has REALLY been working on, from a biological point of view. We are struggling to evolve a brain which begins to address the two greatest questions of human mental life: (1) where did I park my car this time in the parking lot?; and (2) what was I trying to do anyway?"
To be honest, there were a lot of people who did not really understand that talk, as the speaker was not experienced in talking to an engineering audience or in hooking up to their equipment. My explanation propagated in the discussion, and I was delighted a few months later to see neuroscience researchers actually using the parking lot EXAMPLE to illustrate the more basic and general issue of the type of memory (unobserved state variables) it illustrated. (It fits my mathematics of how brains handle spatial complexity, part of the papers I cite sometimes, but not for the post today). THE SECOND QUESTION, "What was I trying to do anyway?", directly addresses the TYPE OF OPTIMIZATION ARCHITECTURE USED IN ALL MAMMAL BRAINS, which is much more powerful than baby systems like alpha Go. There is a system of target variables or "goals" which lives in the middle between the raw input/output neurons of the brain (like senses and muscle control cells) and the ultimate global utility functions (the J and U I have discussed here before). Ancient forms of AI certainly knew that there are things called "goals," but how to embody such things firmly and smoothly and under control in a system based totally on learning, capable of learning and unlearning and putting different weights on things? The open access paper by Yeshua and myself, https://www.
includes a review of what we know and how it fits real-time brain data.
There are times and places when the higher cortex sends orders directly to muscle control systems, which generate action quickly. But more often, it sends signals to the BASAL GANGLIA, a MIDDLE level of the brain, where target variables are engaged and pursued, sometimes over long time intervals. This geometry reminds me of the Sian Kaan, the scared tree of the mayans, where the CONNECTION between the mud (muscles) and the cloud level (higher cortex) is essential to the life of the whole system.
3. REMEMBERING target variables in coronavirus analysis and climate
Yesterday I had a great experience in feeling stupid, the kind of experience we should all remember and treasure in memory. Having felt stupid about coronavirus, I am better equipped to understand and empathize with those well-meaning and powerful people who have been so utterly, placably stupid in policies to prevent human extinction by climate change (more precisely, by euxinia, a search term you can use in youtube if you care about whether the human species goes extinct in the next century +/- a few decades). In both cases, the key need is to LEARN AND REMEMBER the key target variables.
Lessons I learned about how to MISUNDERSTAND coronavirus spill over to climate change policy.
A group of highly intelligent friends challenged me to a Zoom discussion yesterday: "Let us compare the models of coronavirus and discuss which ones we should believe." BETWEEN us, we have had access to models and data and academic papers from EVERYWHERE.
A few days before, one of you challenged us: "You have seen those models used and debated by CDC. One of us may well be able to do better. What are your ideas?"
My response: "This is just like my first tenured job, at DOE/EIA in the 1980s: first do a deep analysis of all the existing models, over a few years, and then redo it all and build them right. BUT THESE models now are far more primitive. More important, there is a "mother of all missing data problems." My work (e.g. see www.werbos.com/Erdos.pdf or https://www.sciencedirect.
OOPS. I should say more about the antibody testing. Coronavirus was very much a part of my FIRST PERSON EXPERIENCE but NOT of my third person scientific efforts. A few weeks ago, I caught something from a fellow Quaker just back from China, and had exactly what my medical ex-wife described as classic coronavirus symptoms. I really wanted to know: am I ALREADY immune now, or could I die in a week if I go to the wrong store? I realized that these issues apply to MANY people, and to the issue of reopening the world economy (an issue worth many trillions of dollars), and even sent bits of information to chosen people here and there. I became to coronavirus what a lot of those pseudoexperts on climate are to climate.
I never INTENDED to become a world expert on coronavirus! I am already struggling with other larger, more important issues. Yet as I decide where to go on a day of lockdown, in a neighborhood which has one of the highest death rates per capita, I do think about my personal survival. PERHAPS THIS IS HOW many people think about climate change, taking it quite seriously, but now knowing JUST HOW SERIOUS it is or isn't, and not letting it EVER become a major focus of their highest level enlightened strategic thinking even for an hour. Like them, I read lots of things, but never let it get to me at a deep level.
BUT THEN YESTERDAY my friends challenged me to a sketch a real mathematical model. They would prefer a good old fashioned econometric model, like the one I cited above. (Really. The Stanford Energy Modeling Forum EMF found that the econometric model of industrial energy demand which I built back then was the most reliable such ever built, and that lead to the publication above in their special issue of the journal Energy.) BUT: I said no, BEFORE one can build a quality, reliable econometric model, one must first explore the territory, considering alternative Conceptual mathematical models, rather like the seven equation Keynesian model I learned about decades ago from Ackley's classic text on macroeconomics.
So there I was yesterday, at kitchen table and then outdoor patio table, trying to scribble just a few RELEVANT macroeconomic type equations to give MY view of what a correct model would look like. Oh, did I feel stupid! I consoled myself with the memory that it took three solid months of DOE library research and focused thinking to come up with even first draft of my "seven equation" model of industrial energy demand, but this time I HAD already seen the major branches of what is known today. I started to list a few key variables...
Of course, percentage of antibodies to resist the flu, BY group of people, is one crucial variable. There was a great study in Germany, followed by a new one from Stanford (which the drug companies love about as much as they love Dean Radin's work), suggesting that the death rate for people who catch coronavirus is only about 0.1%, FAR less than CDC was telling us just two weeks ago, because so many of us SURVIVED the virus and never got entered into CDC data. (I deeply respect Fauci and Cuomo for working hard now to fill this crucial hole in the data.)
But what about OTHER sources of immunity? The German study, based on solid statistical sampling in a prime "hot zone," SUGGESTED that only 20% of the people had the anti bodies but that 90% or so HAD BEEN exposed fully but never even needed antibodies.
So yesterday, feeling like a fool, trying to sketch a few equations on the back of a sudoku printout on a patio table outdoors, I did at least list those two time-series variables, percent antibodies by group and percent other immunity (for which we have essentially no direct data at all). And then...?
Certainly it is huge issue WHERE to restart the world economy (even before we get into the coupling with macroeconomics, which is what I WANTED to think about). As with physics... I had to remember again the lesson about CRAWL BEFORE YOU WALK. My understanding of coronavirus had been blocked by my interest in the economics issues, such that I did not really pay full attention to the virus policy issues, even for an hour... just like those COP25 and Obama'Biden bill people, distracted by "bigger priorities." JUST ONE HOUR of true focused attention, with strategic thinking... can bbe worth so much...
--
Anyway, after my mental failures yesterday, I reviewed this (among other things) at samadhi time early this morning, and see much better how to model this thing.
THE KEY TO CORRECT modeling of coronavirus epidemiology is, first, to remember the policy context. This is about doing the best we can within a limited time frame. Actions are needed NOW. More precisely, actions taken NOW will be important, and should be well informed, regardless of whether we will later learn better. Classic Raiffa style decision analysis, at root.
Second: to identify not just causal variables but TARGET variables. What are we trying to accomplish anyway?
For a correct ballpark analysis (the best we can get), we should perhaps first assume that EVERYONE WILL BE EXPOSED TO THE VIRUS. Let us be clear that the target variable is not zero exposure.
Postponing the economic aspects: key target variables, "local utility functions," are the rate (per capita?) of AVOIDABLE DEATH and AVOIDABLE PERMANENT BODY DAMAGE, and DAMAGE TO HEALTH CARE SYSTEMS.
Those first two target variables suggest grounding a model CONCEPTUALLY on time series of INDIVIDUAL PEOPLE. The times series of the LEVEL of exposure to the virus will have crucial impacts on probability of death or permanent damage. Many deaths of health care workers may be due to TOO MUCH exposure too fast. (There is serious literature on these effects, but NOT enough to give is a reliable decisive feel for exacrtly how it cuts.) Low levels, like a vaccination, may give time for enough buildup of antibodies OR those other mysterious immunity mechanisms. OR people with those mechanisms from day one may simply not change at all, until those mechanisms weaken for some reason and gentle exposure works best.
However, level of exposure as a function of time is certainly not the only input variable to predict what happens to a person. There is the mystery immunity factor, and things like asthma, diabetes, hypertension, paranoid immune systems. At first, one might say that there is an optimal schedule of exposure as a function of those other attributes of people, and that it may be slower for some than for others, especially slow in nursing homes. HOWEVER: it is crucial that those other factors are VARIABLES AS WELL. Above all, there are new studies showing how aerobic exercise may RAISE immunity A LOT, raising suspicion that dietary variables and exercise (both mundane and esoteric, folks!) may change the level of "mystery immunity."
There are huge policy implications even to this simplified conceptual mathematical model. In restarting economies, it is important to push hardest to open up in areas where the PERCENTAGE OF ANTIBODIES is a high fraction of the equilibrium antibody percentage, as the spread of the virus reaches equilibrium statistics. And so, that hot zone in Germany and New York may be the FIRST places to ease up on restrictions (with exceptions for places like old age homes where the optimal schedule of exposure is over more time, OR EVEN crowded isolated places like ships at sea where the level of exposure per person tends to be higher in response to an initial injection of the virus to the place). God help the people of Texas if THEY decide to go first because they haven't got the antibodies yet! (I suppose it would eliminate some of the more starry eyed Trump believers, a bit like that town/church in South Korea where believers who refused to use masks died like flies.)
But for policy purposes, any such model should be supplemented by something like a county level model, to predict the risk of breakage to health care facilities. A key target variable is not ONLY total death rate summed over time but the gap between the need for hospital care AT ANY TIME and the facilities available THEN. Fauci righhtly reminds us not to forget what happened in Italy!!! But new data also suggests that shortage ofventilators is not as important as we thought a month ago, since.. well Cuomo tells us that people put on ventilators for corona virus mostly die anyway, and other data suggest ventilators THEMSELVES (e.g. complications of anesthesia) may do as much harm as good in such cases. Simple CPAP machines and oxygen tanks may be easier and better. Is it a matter of rational supply chain modeling and management rather than timing of patient load? Or should US extend lockdown more than it would IF the first two target variables were the only issue, JUST in order to accommodate supply chain management inferior to what should be possible in today's world (e.g. true AGI)? Are people dying today to pay for the sins of overconfident supply chain app developers? (That, in US AND China and EU.)
4. Various Extensions
(Google wanted to insert $ here. Maybe there Is a ghost in the machine,,,)
Having well-defined target variables and laser-focused strategic thinking is ESSENTIAL to clear thinking and even meager levels of competence in areas like climate policy and virus policy. but it is EQUALLY essential to be able to stand back later and account for the bigger picture. I OFTEN recommend Levitin's popular book, The Organized Mind, stressing the need for us to BALANCE extreme focus and relaxed greater openness. This is not just an old adage; it is a key principle in AGI and organization design. (For example, both NSF and NASA have often needed budget decisions to be based on a SUM of value to core mission -- target variables -- and broader impacts. And lots of intelligence to get both right.)
For virus policy, the IMPACT ON THE GLOBAL ECONOMY is a HUGE broader impact. (This year, every time I use the popular word "huge", I think of the equally popular word "unpresidented".) That's where my mind WANTED to go yesterday.
But: the bottom line economic growth ov er the next ten years is a SUBSET OF THE AGI THREAD.
Whichever way it goes, short of total collapse to the Stone Age, the financial system ten years hence is a SUBSET of the internet, NOT the other way around. That is a simple basic principle we need to get straight, to avoid the many opportunities for fiscal and physical suicide.
HERE AND NOW, major nations like India may not be able to afford the degree of lockdown we see in New York and DC. The "broader impacts" term in THEIR policy decisions may cut quite differently from how they cut in these areas.
But beyond that... global economic plans which assume EITHER that we get to print unlimited paper money OR that money is a hard physical constraint like gold are BOTH too extreme to survive. I have been reminded a lot of how inflation in Germany in the 1920s, after Spanish flu, might have even been decisive in causing Nazi rise and World War II. Getting money straight is VERY important now, more important than the flu itself. And MONEY IS AN APP DEPENDENT ON ITS PLATFORM!!!!!
Confused as climate policies seem to be in most of the world, that kind of platform policy is far, far worse. Will we be able to survive the growing strength and focus (myopia) of "Godzilla, King Kong and the Borg" (my names for the three main serious platform development realities I see emerging right now)? TBD. As they say, "God help us. Please."
Best of luck,
Paul
apologies... I fear i was not as helpful as my heart otherwise seeks help my mind, to be.
If there was sufficient distributed compute to solve some of these problems, how would you define the ACLs...
I guesses, that's my underlying point. The work on (royalty free) standards, etc. is in most ways, all about how to answer that problem. If that puzzle is solved well, then we have a chance to do some good - whilst seeking to solve others.
but horse, before cart. ATM, there's some who have automobiles, others trying to feed their horses petrol hoping they go faster, when in fact there's a decision to be made - build new tech, or get back onto the grass...
| ||||||||||||||||||||
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment