I have been considering
the whole issue of AI. You see I have been looking at this for about fifty plus
years now. In fact, when I arrived at Warner in 1980 or so, my boss, Gus
Hauser, sent me a note I believe the third day to ask to report to him on AI. I
knew Patrick Winston at MIT, he had published a book on AI from his
perspective, but I also know Minsky, Papert and others, so I had been at the AI
watering hole. To me AI was just the name of a watering hole, not a thing unto
itself. Thus Gus got my opinion. Skip ahead to 1986 and as the new Executive
Director of Research at NYNEX, now Verizon, I was being pressured to develop a
whole area in Neural Networks. I knew this area well, but I was also assured
that the then current computer systems were inadequate.
You see, back in 1971 I had a brief sabbatical at Bell Labs
trying to track Soviet subs. Massive data focused on pattern recognition. I
tried larges scale data and deep learning algorithms. Did not work well.
Fundamental problems existed in even gathering the data but that fact did nt
skink in, it may have hampered their contract. Thank God the Soviets never
attacked.
Now back in the mid-60s I spent time at the MIT
Instrumentation Lab, working on guidance and navigation systems for Apollo and other
projects. That is when I became enamored by the three body problem. The force
on a three bodies and the resultant sets of equation can be determined as
follows:
and likewise for the other two accelerations. From the
solutions of these three calculations we can determine the dynamics of a
spacecraft going between the earth and the moon. Now these equations are a
result of two of Newton's laws of gravity; (i) force mass and acceleration, and
(ii) force, mass and distance.
Now let us consider how Newton may have approached this
problem using massive amounts of data and neural networks. Namely let us assume
Newton could not think but he was a great coder. So Newton sees this apple fall
from the tree and he thinks that there may be something here he could use to
predict lots of other things, such as the trajectory of a cannon used in
battle. So Newton goes out and collects tons of data.
For example, since he has no underlying theory he must just
collect whatever he can. He is interested in such things as inputs and outputs
of this neural network so he must a priori define these elements. This is the
first step. But, and this is critical, he can only measure what he can define
and what he can measure with the tools available to him.
So Newton sits under the apple tree and gets hit on the head
with a falling apple. He then wants to know why the apple fell and how fast it
was going when it hits his head.
So what results would Newton like to get as the output of
his neural network? They may be:
Speed of the apple when it hits his head
Time it takes to fall
Distance it fell from the branch
We of course must assume he has the tools to measure these
things. But he does measure them and most likely with errors.
Also what inputs would Newton want to consider in his neural
network as drivers of his outputs for which his neural network will determine from
tons of data? They may be:
Temperature
Day of the week
Time of day
Color of his shoes
Species of tree
Amount of sunlight
Height of branch
Species of apple
Diameter of apple.
Weight of apple
Volume of the apple
Location of tree
Age of the tree
Latitude
Longitude
Angle of the sun
Color of the dress the Queen wore that day
and of course the list goes on. You see he has no idea what is
driving his result so he just gathers tons of stuff he can measure just in
case. Lacking a model he fills his network with "stuff" .
So now Newton goes out and spends days and weeks under apple
trees, he recruits many others, under order of the King, to also sit under
apple trees, and after a while half of England is sitting under apple trees
measuring the stuff Newton wants to get. Tons of data, massive amounts of data
arrive.
Alas Newton can enter this into his neural network and let
it grind away. So what is the result? Does he get the equation? No, not at all,
he now has a big machine that requires your to enter tons of data to determine
the speed of the apple when it hist your head under a specific tree, falling
from a specific height. Is there some equation? Nope! Just the machine. Did we
solve the three body problem, not even close.
Now back to the three body problem. In my Apollo days we had
a computer with 64K memory, yes computer geeks, 64K, not Meg, not Gig, not
Tera, K. That meant we had to think "smart" and not "hard".
We needed to viscerally understand that three body problem, when and where and
how much to fire the rockets for return.
Now let us move this to health care, say cancer diagnosis,
prognosis and treatment. We now move to the current date where we are trying to
diagnose say a thyroid tumor. They come in several varieties, papillary, follicular,
medullary, and others. Now each of those have some sub classes. Our output is
three stages; diagnosis (what type), prognosis (knowing what type what is an
outcome), and treatment (knowing the first two what should we do). Thus one can
consider a three output system, and some of the outputs having a multiplicity
of subtypes.
The input is now what we can measure, what we have tools to
measure. That is an important fact to remember since as we progress in
knowledge and in tools what we can measure today may be a small amount of what
can be done in a decade. There is no underlying physical laws to enforce, just
tons of data and hopefully an answer.
Now consider an alternative approach. Suppose as, first with
Newton, we had his laws. Then all we need is to find k, and solve the
complicated set of equations. That value could readily be found by a some what
dumb neural net. That is called a system identifies. Been there done that. But
we could use a well know system presentation of cancers, one where we identify
a measure, call it n(x,t), where n is an Nx1 vector where each element is the
local concentration of a cell of a specific genetic composition, say a
melanocyte with BRAF V600, or RAS, or N-cadherin and all possibilities thereof.
In fact here we have possibly hundreds of genetic profiles, starting with the
most benign and to the most malignant. n(x,t) may be a 1000x1 vector, and it is
a function of time and space.
Now we ask how does this state, genetic state if you will,
change in time, and on average. Well we have demonstrated the following. The
rate of change is equal to a diffusion state, a flow state, and a growth state.
This is a fundamental law of any organic system. This is Newton's law for
cells. We show this below and have presented it elsewhere.
Now if we were to collect massive amounts of data we could
determine a, b, c as above and then it could lead us to diagnosis, prognosis
and treatment. It becomes a system identification and in turn optimal control
problem, namely identifying the offending gene progressions and identifying
where and when to stop that process.
Thus AI should be more than blind data churning. For it may
have led Newton to a law dominated by the color of the Queen's hats and not a
function of the product of masses. AI, to properly work, must have a set of
underlying verifiable paradigms, models, which need further specificity. It
should not just be a black box which tells us nothing about reality.
Posts
