Re: Bayesian Inference Engine

On Mon, 11 Sep 2006 09:16:16 GMT, Michael Olea <oleaj@xxxxxxxxxxxxx>
wrote:

HMSBeagle wrote:

On Sun, 10 Sep 2006 05:55:56 GMT, Michael Olea <oleaj@xxxxxxxxxxxxx>
wrote:

HMSBeagle wrote:

On 9 Jun 2006 04:38:24 -0700, "JGCASEY" <jgkjcasey@xxxxxxxxxxxx>
wrote:

The winner of the DARPA grand Challenge is a Bayesian
Inference Engine.

I dont know who wrote this first. I will go ahead and attribute it to
JGC.

Anyways, its false.

...

What is it that draws idiots to C.A.P. like moths to a flame?

Google: Thrun Bayesian Darpa

The truth is out there.

We will see who the "idiot" is now won't we?

Well, no, not "now" -- it was already crystal clear when you wrote "Anyways,
its (sic) false" about a question of fact that is easily checked. Maybe
"easily" is a little too strong; checking up on the facts of the matter
does require 1) that you understand how to use words like "winner", "first
place", "second place", and apply that understanding to the results of the
DARPA Grand Challenge (a task you failed the first time around), 2) it also
requires that you have a clue what is and what is not Bayesian inference.
You don't. But I will grant you that you have made an already rock solid
case even stronger. You quote from a paper describing a Bayesian Inference
Engine, and, apparently because it does not use that exact phrase, cite
[snip]

I see the name-calling is not going to end anytime soon. Anyways,
calling the car a "Bayesian Inference Engine" is a completely
different matter from the fact that the vehicle uses bayesian methods
inside of parts of its software. I emphasize *PARTS*.

this as evidence that a "probabilistic reasoning" system does not use
Bayesian methods. Do you have any [meaningless intensifier elided] clue
what Bayesian methods amount to? You are willing to pontificate on matters
you know nothing about. That makes you, in a word, an idiot.


I know exactly what they are. And the vehicle Stanley is not a
"Bayesian Inference Engine". I will continue to support this
conclusion more and more in the rest of this reply.



... That was wrong. It was the winning
vehicle that used them (Stanley).

So you have at least learned a little something. Now, do you think it might
be germain to the question of fact you presumed, in all ignorance, to
dispute, to probe a little deeper into just what, technicaly,
algorithmicaly, constitutes a "cloud of guesses"? Here's a clue, moron:
google "particle filtering". You might also try "importance sampling". And
while you are at it, check out "approximate Bayesian inference".


That's not a clue at all. The cloud of laser points is not a particle
filter. You claim to have been doing this for 15 years? I suggest
you not quit your day job. What's worse, "Importance Sampling" has
absolutely nothing to do with the Stanley vehicle whatsoever, not even
in some bizarre twist of math that you think constitutes "proof".
The vehicle takes an equal number of readings per cycle regardless of
the nature of the last readings. Importance sampling would only help
if the car was sitting still in a room with no weather changes. These
vehicles never engage such environments. They are moving at a fast
rate of speed most of the time. The Z (height) value of a reading is
updated based on when the last reading was taken.





... To demonstrate, I will quote the pdf
found here:
http://robots.stanford.edu/papers/thrun.stanley05.pdf

Here's the thing, pilgrim; it's not enough to cut and paste text to make a
point. You actually have to have some clue what the text means. You don't.

This pdf was written by Sebastian Thrun himself. ...

The list of authors is in fact:

Sebastian Thrun, Mike Montemerlo, Hendrik Dahlkamp, David Stavens, Andrei
Aron, James Diebel, Philip Fong, John Gale, Morgan Halpenny, Gabriel
Hoffmann, Kenny Lau, Celia Oakley, Mark Palatucci, Vaughan Pratt, Pascal
Stang, Sven Strohband, Cedric Dupont, Lars-Erik Jendrossek, Christian
Koelen, Charles Markey, Carlo Rummel, Joe van Niekerk, Eric Jensen,
Philippe Alessandrini, Gary Bradski, Bob Davies, Scott Ettinger, Adrian
Kaehler, Ara Nefian, and Pamela Mahoney

... First and foremost,
the word "Bayesian" appears NOWHERE in the pdf. Absolutely nowhere.
The word "bayes" does not appear anywhere in the pdf. The phrase
"inference engine" appears nowhere in the pdf - not even once.

You see? This is what idiots do. They ignore any conceptual analysis of
content, which would require some actual understanding of the material, and
instead look for keywords to tell them what it was all about. Make no
mistake, this paper describes a Bayesian inference engine, just as surely
as Hawkins' "On Intelligence" described a Bayesian Network model of
neocortex. This is immediately obvious to anyone conversant in Bayesian
methods. You do not have to search the index, or search the text for
telling phrases. All you have to do is understand the algorithm described,
and understand Bayesian inference. You, dog breath, do not.


That's nonsense. There are probabilistic updates of readings. This
methodology is bayes-like in its intention, but that is only one iota
of the entire software system in this vehicle.

A stack of books on bayesian inference would never tell you to couple
the RGB camera data with the data from the laser sensor. This idea
is truly unique and the PDF cites those who first coined the idea in
print.


/*
This article describes the robot Stanley, which won the 2005 DARPA Grand
Challenge. Stanley was developed for high-speed desert driving without
human intervention. The robot's software system relied predominately on
state-of-the-art AI technologies, such as machine learning and
probabilistic reasoning. This article describes the major components of
this architecture, and discusses the results of the Grand Challenge race.
*/

Just what the hey do you think they mean by "probabilistic reasoning"? Just
what the hey do YOU mean by Bayesian methods?

I don't even agree with that abstract. The vehicle does not use
"reasoning" at all. It only uses probabilistic updating of sensory
readings to refine them. I also do not agree that the vehicle is
engaging in "machine learning" because that would imply that it is
acquiring a history of its interactions with the world. I'm certain
the vehicle does not use LEARNING methods (in the sense of storing
memories and acting on them in the future.). The vehicle is
mostly programmed to take certain speeds in certain conditions of the
road in front of it. It does not LEARN an association between what
it can get away with on certain roads and what it cannot get away with
in terms of driving speed.


I will now quote copiously from the pdf itself.

And you will do so with no comprehension at all of the text you quote. You
will utterly miss the fact that you are quoting more or less random
snippets describing a Bayesian Inference Engine.


A stack of books on bayesian inference would never tell you to couple
the RGB camera data with the data from the laser sensor. This idea
is truly unique and the PDF cites those who first coined the idea in
print. The fact is, the vehicle as a whole (while using bayesian
methods in its a few parts of its perception loop) is far and above a
mere "Bayesian Inference Engine.". Your claim was that it was that,
AND NOTHING ELSE. That claim is wrong.

Just because Stanley uses a bayes-inspired updating of laser sensor
readings, it does not mean that the WHOLE VEHICLE IS A BAYESIAN
INFERENCE ENGINE.


[pdf]
"Obstacle detection on laser point clouds can be formulated as a
classification problem, assigning to each 2-D location in a surface
grid one of three possible values: occupied, free, and unknown.
A location is occupied by an obstacle if we can findnd two nearby
points whose vertical distance exceeds a critical vertical distance
DELTA."
[end pdf]

These "laser point clouds" amount to a distribution over a biased (i.e.
prior) hypothesis space. Can you say "hypothesis space", beagle boy? Do you
have any clue how the DELTA is calibrated? Do you have the slightest
inkling of the bidirectional flow of evidence between diagnostic and causal
support?

People reading your little tantrum here might think the PDF says the
DELTA is "calibrated" which it is not. Your tantrum will be easily
demonstrated by me linking readers to the source itself which can be
found here:
http://robots.stanford.edu/papers/thrun.stanley05.pdf

DELTA is simply a critical "tallness" value on the road. The laser
sensors are trying to find bumps. That's about as complicated as it
gets. The the height difference between adjacent points (that have
been measured) is distributed according to a normal distribution
whose variance scales linearly with the time difference between the
laser readings. That's not "bayesian inference", that's just a
statement of fact. The authors of the PDF point out this fact on a
graph of readings on page 13.

If this were "bayesian inference", there would be no a priori
assumption about how the readings change over time. Stanley assumes
a LINEAR VARIANCE OVER TIME, FOR EVERY READING. That's not bayes.

And it gets worse, because of the linear variance, each reading gives
you a upper and lower BOUND on future readings which can be culled as
"inconsistent" and ignored by the vehicle. Bayesian inference does
no such thing. This is just a good algorithm.






Gee, probabilistic update of the belief distribution in the light of
diagnostic support and causal inference -- what could that be about? What
probability calculus is involved? Do you know one damn thing about what you
presume to dispute? No, you do not. Geez, what an idiot. You really have no
clue, do you, that the "updated probabilistically" methodology described in
the paper is FREAKING BAYESIAN INFERENCE!


Yes. That one tiny part of the whole vehicle loop is bayes-inspired.
That does not make the vehicle a giant "Bayesian Inference Engine".
You are wildly jumping to conclusions here.

Some related papers on Thrun's home page:

J. Pineau, G. Gordon, and S. Thrun. Anytime point-based approximations for
large POMDPs. Journal of Artificial Intelligence Research, 2006.
forthcoming.

D. Lookingbill, J. Rogers, J. Curry, D. Lieb, and S. Thrun. Reverse optical
flow for self-supervised adaptive autonomous robot navigation.
International Journal on Computer Vision (IJCV), 2006. Forthcoming.

J. Diebel, S. Thrun, and M. Brüning. A bayesian method for probable surface
reconstruction and decimation. ACM Transactions on Graphics, 25(1), 2006.

ALL these papers are about Baysian methods in robotics.


A stack of books on bayesian inference would never tell you to couple
the RGB camera data with the data from the laser sensor. This idea
is truly unique and the PDF cites those who first coined the idea in
print.



Yep. Calculate a new probability distribution over the hypothesis space,
given the old distribution and new evidence. It is called "belief update",
and it relies on efficient "belief propagation" algorithms. This is Baysian
inference in action. Puppy, Beagle boy, Wiener Dog, has not one iota of a
clue about that which he presumes to dispute.


Yes. This is bayesian inference in action. But it is one part of the
vehicle. One tiny, small, insignificant part. This is the part
where it finds out if that part of the road is flat. That's a pretty
minor detail in the whole flowchart of the vehicle.


Notice this describes the REFINEMENT of points in the cloud over time
as *NEW MEASUREMENTS" come in, just exactly as I had described in my
post.

You actually did not describe anything "exactly". But, by gumbo, just what
do you, in all your manifest and profound ignorance, imagine Bayesian
inference is all about, in its exacting calculation of new distributions of
probability density functions from old distributions of probability density
functions, given new observations? Just what the [meaningless intensifier
elided] do you think "REFINEMENT of points in the cloud over time as *NEW
MEASUREMENTS" come in" amounts to, you stupid little [fill-in-the-blank].


I said in my original post that the updating of points is
bayes-inspired. I can do without the drama.


The part of my post that was completely correct was that Stanley
couples the COLOR of flat road with laser sensor data that is
confirming that it is flat. The car then drives towards this color,
assuming that it will not change dramatically over time. Dr. Thrun
himself admits that he did not invent this methodology. I quote Thrun
in his own words:

Words you have not the slightest qualification to understand. The part of
your post that is completely erroneous is the contention "Anyways, its
false".


It's still false. Referring to the Stanley vehicle as nothing but a
large "Bayesian Inference Engine" is so innaccurate as to be false.
What would be more true, would be to say it uses a bayes-inspired
method in parts of its software.


[pdf]
"To find the road, the vision module classifies images into drivable
and non-drivable regions."

Do you know anything at all about classifiers? What is the relationship
between the VC dimension of a classifier's hypothesis space and the number
of training samples needed to achieve a given PAC (probably approximately
correct) level of classification accuracy? Huh? Any clue? How does that
answer depend on the PDF of the stochastic source of drivable and
non-drivable images? How does this relate to Bayesian Inference? Got clue?


I could do without the drama. Anyways, the vehicle is going to simply
project the laser readings onto the camera image. Yes. It is simply
going to PROJECT THEM ONTO THE camera image. People reading your
little tantrum here may be led to beleive that some esoteric PAC
scheme is going to be used. That's false. I refer the wayward
reader to look at the PDF in its own words:
http://robots.stanford.edu/papers/thrun.stanley05.pdf


...we already have such drivability information from the laser in the
near range. All that is required from the vision routine is to extend
the reach of the laser analysis.

And that extension is, as described in the paper, done probabilisticaly,
using Bayesian Inference.


Now you are just lying.

.
Stanley finds drivable surfaces by projecting drivable area from the
laser analysis into the camera image.

Done probabilisticaly, using Bayesian Inference.


Another lie.


.
The learning algorithm maintains a mixture of Gaussians that model the
color of drivable terrain."
[end pdf]

Oh geez, a "mixture of Gaussians" prior over a parametric hypothesis space
Baysian inference 101. You just cannot get more Baysian than to go mixture
of Gaussians, as a matter of computational convenience. I have done the
same many times and in many contexts, during my 15 years of professional
experience applying Bayesian methods to pattern recognition problems.


No. The PDF does not say that. This is like the 17th time in your
post that you have said something that the PDF does not actually say.
In the end, who do you think the readers are going to trust? You?

[pdf]
"The learning adapts to the image in two possible ways; by adjusting
the previously found internal Gaussian to the actual image pixels, and
by introducing new Gaussians and discarding older ones. Both
adaptation steps are essential."
[end pdf]

The algorithm uses an adaptive update which discards Gaussians already
storied. That's not bayesian inference, that's just an adaptive
learning method. Anyone can go read the PDF and see who the liar here
is.

I repeat for emphasis *THE COLOR OF DRIVABLE TERRIAIN*.

What you utterly ignore is that "COLOR OF DRIVABLE TERRIAIN" is, in this
paper, diagnostic support for a Bayesian Inference Engine.


Oh no... you are stacking your lies two layers deep now. The color
of drivable terrain is created by mapping the color of flat terrain to
the color of distant terrain that the laser sensors cannot reach.
Since colors in the real world are often LUMPY, Gaussians are used to
smooth out these irregularities. Anyone can read the article itself
and see I am practically quoting it word for word.

You are picking out little details of the PDF, and ignoring the bulk
of it. The authors themselves referred to their adaptive method as
"essential". (See above).

I will not continue to
respond to posts in the future that are written by people calling me
names.

Oh, right. You hypocritcal little chiken ***. You post some highly
insulting ignorant crap attacking my integrity "I dont know who wrote this
first. bla bla bla. ... Anyways, its false", follow it up with a post
illustrating beyond any shadow of a doubt you have no clue at all about
anything Bayesian, and then have the goddamn gall to talk about people
calling you richly deserved names. Hie thee home, fragment. And when you
get there do be sure to commit an anatomically impossible act. Idiot.


Given what I know about AI, you must have had a parent who rewarded
you at a young age when you threw your tantrums. (There must be a
behaviorist or two reading this who will know exactly where I'm going
with this.)
.

Quantcast