The Filesharing Conspiracy

Up to 103 hours total processing time
(PLEASE NOTE: This is total CPU Clock tick time
for the entire grid system NOT linear time)

Processing of the Britney-to-Basso Profundo Voice is
STILL CONTINUING and is at 103 hours of CPU clock tick time
which will continue as per my APPROXIMATE estimate of 100 to
200 hours total time.

Sorry! There's nothing much I can do to speed it up
since during the day the network gets used by others.
So my apologies for this test taking as long as it has
but I can say that to please RESERVE YOUR JUDGEMENT
(Good or Bad) until processing is complete and then I'll
post it on youtube.

My software is as per what is displayed on the
www.midgrid.com website so very SOON you'll
be able to download it yourself for your own
amusement.

On a commercial basis, this software DOES REQUIRE
some considerable CPU horsepower to get things done
QUICKLY, but as home users, if you're willing to wait, your
results should be more than just half-decent.

And with 8-core and even 12-core CONSUMER CPU's coming
by first quarter 2011 from AMD and Intel, this grid-based
processing software will SUDDENLY BECOME VERY PRACTICAL
FOR HOME AND GENERAL BUSINESS OFFICE USE.

I will continue to post processsing times as this test completes
but so far I still say about 200 hours clock tick processing time.

No problem from me with the time, Stargate, I just want to see if this can be done..

Thanks for keeping us updated. I'm waiting to see what happens.

Mandatory second line purely for the sake of second lining.

Still processsing at 122 hours...So I do have to eat some crow
here in that timewise you'll need some extra horsepower
to use my software (8 gigs RAM, Hexacore AMD is a good start!)
if you want to get it done in a reasonable amount of time.

Our GRID network is mostly Single processsor and 2-core
AMD Athlon-based 2.3GHZ machines with 2 gigs or 4gigs
of RAM so they're positively ANCIENT by computer standards.

Our hexacore machines are unfortunately being hogtied by those
who have more immediate needs than I.....:-(( i.e. ME ANNOYED!!!

---

On a positive note: I have done something interesting
in terms of getting a pre-sample for me to hear using some
software and Operating System interrupt trickery to read
portions of a locked file so I can give a sneak peak at what
is coming out and I can state there are some half-decent sounds
I'm getting out of this test .... NO!!!!! DON'T ASK !!! --- You'll
get to hear it when I'm FULLY DONE!!!!

I think I can say with some clarity that SOME portions of this test
will sound quite good (if not EXCELLENT!)......and that's all
I'll say for now!

REMEMBER! The purpose of this test is to prove (or disprove)
that a qualitative voice change can be AUTONOMOUSLY
performed upon pre-recorded song material.
i.e. female singer changed to Basso Profundo voice.

The end result could be silly or it could be Insanely Great!

Who Knows?

So keep coming back for more updates.


PLEASE NOTE:

I will be in Northern British Columbia for 3 days
so I won't be having internet access because I'll be
in the middle of nowhere on a "paying job".
I'll be back Thursday evening or friday morning.
And based upon current timings I'm guessing
sometime during mid-to-late next week for a finish?!

I'll keep you all posted when I get back.

I have set the test to continue running during nights only
for the next 3 days but this won't affect processing much
because each CPU already has their CONSIDERABLE datasets
to number-crunch to their hearts delight.

Ok, so although I'm not really that interested in the timescale, just to see whether it can be done, I'd say it's failed the timescale aspect. I'd reckon in 122 hrs of "studiotime" (and probably well, well under that amount of time) Mike could easily find musicians, re-record the track with a bass singer and mix and produce it. This is unfortunately the difficulty new technology faces in its infancy. A very skilled, dedicated human with experience in their field can still be very useful and much cheaper to employ!

Originally posted by J.Clear
Ok, so although I'm not really that interested in the timescale, just to see whether it can be done, I'd say it's failed the timescale aspect. I'd reckon in 122 hrs of "studiotime" (and probably well, well under that amount of time) Mike could easily find musicians, re-record the track with a bass singer and mix and produce it. This is unfortunately the difficulty new technology faces in its infancy. A very skilled, dedicated human with experience in their field can still be very useful and much cheaper to employ!

---

Failure is Good!!!!

You can't know what to fix until it goes wrong!

And in this case the issue is more to do with
limited computer horsepower (i.e. cheap & old 25 AMD Athlon
32 bit and 64 bit CPU's - 1 & 2 cores).

It's only this year that the first true 4, 6 & 8 core CPU's have come
out which will make your complaint pretty much a MOOT POINT!!!!!!

And even a FEW AMD Hexacore 3.2 GHZ systems will BLOW AWAY my
network and since my software WILL take advantage of both Multicores
AND networked CPU systems processing times will only decrease!

AND I should point out that our Midgrid software can LEARN on it's own
and after this experiment is done, it will have an extended experience
base that will allow the system to perform MUCH FASTER the second time
around (and the Nth time around!).

I expect that THIS TEST WILL SHOW UP issues and problems BUT I still
expect that the final result will LIKELY be better than you might think.

YES! it won't be perfect but I say that when YOU were a kid,
were you ABLE to ride a bike or skate or swim perfectly the
first time around??? I THINK NOT!

In this case though since Midgrid IS a neural net, it can learn
from it's mistakes and CHANGE its behaviour over time as the
boolean logic trees become more refined.

IN my case I expect that the phonemes of my Basso Profundo
while not quite fully overlap the female voice of the Britney song
because the software is NOT YET smart enough to tell the
female singing voice from some of the background instruments
because SOME of the frequencies will overlap and thus NOT ALL
of the female singing voice will be cut out of the background.

However as I run the tests on DIFFERENT songs over time,
the software WILL get smarter and smarter making better
choices as to what frequencies to cut out and replace with
new foreground material.

As I am now back I will continue to update the number of
processing hours which is now at 154 hours total CPU time.

From a studio point of view they can AFFORD to PAY for a $150k
array processor to finish this type of test in a few minutes.

Hope this helps.

Hehehe.. well, I think everything you've said will be a "moot point" til we hear the results. Of course, these developments will probably make a massive difference, but I wouldn't say it's a totally cut and dried situation.

I understand how your neural net will work, very cool in theory, but could you point me in the direction of a neural net that has done what its inventors claimed it would be able to do after gaining enough experience? I'm just curious, there was one senior lecturer at my university who was mad into em but to be honest I don't think they got near any end product with them as far as I know. It got them very excited, for sure, but it has yet to live up to their expectations.

Originally posted by J.Clear
Hehehe.. well, I think everything you've said will be a "moot point" til we hear the results.

Of course, these developments will probably make a massive difference, but I wouldn't say it's a totally cut and dried situation.

I understand how your neural net will work, very cool in theory, but could you point me in the direction of a neural net that has done what its inventors claimed it would be able to do after gaining enough experience? I'm just curious, there was one senior lecturer at my university who was mad into em but to be honest I don't think they got near any end product with them as far as I know. It got them very excited, for sure, but it has yet to live up to their expectations.

---

Your comment illustrates a common problem with ALL neural nets
in that the initial rules must come from SOMEWHERE, which is usually
assigned to the programmer themselves who are generally the WORST
people to be told to create such initial rules.

The initial logic tree rules of ANY nerual net SHOULD be set
by subject matter experts and NOT programmers. In MY CASE,
I need to break my own rule in that I'm doing the initial logic tree
database BUT I am also setting up a formal set of procedures and
software-based neural net logic tree database input filters that
will ALLOW ANYONE to START an initial neural net database with
some confidence that specific subject matter will at least be
FILTERED and refined to an acceptable level right from the outset.

These neural net "Starter Kits" will be built right into the Midgrid software
with an accompanying "Connection Engine" so that other people who
DO HAVE subject matter expertise can contribute on a GLOBAL BASIS to
the start of a globally available "Subject Matter Expert App Store" which
will allow users to give, lend, lease, rent, sell or GIVE away their
subject matter expertise in a manner similar to the iTunes Store.

So if you need a Finite Element Analysis neural net that evaluates
the properties of various construction elements in an aerospace application,
you just click on the subject matter database logic tree you want to
download and then click on the "Checkout" button paying by credit card,
paypal, debit card or via some sort of barter or giveaway or other credit
schema AND BINGO you'll be able to START your own neural net
logic tree which YOU can add to yourself or you can add to from
the expertise of MANY others of similar training and expertise.

There will be a FORMAL expertise certification, peer review and
rule-set verification program to ensure that certain neural nets
will conform to accepted scientific standards and/or legal qualifications.
You don't just want ANYONE's neural net designing aircraft engine parts
or giving legal advice....!!!!!!

So in answer to your question as to what is an acceptable level of neural
net performance, we actually CAN MEASURE that by duplicating what we
do to children or students...We train them (i.e. general Neural Nets) on
proven problems that have verifiable solutions and/or accepted
methodologies of problem solving. And like training children,
we keep making the neural nets do many homework assignments
until they reach a specific expertise-grade of problem
solving ability.

Then you can graduate your neural net into college level problem solving
by allowing it to tackle problems that are difficult but not insurmountable
or totally unknown by the scientific establishment. If the neural net
performs decision making that brings about an answer that is within
specific minimum and maximum performance metrics, ONLY THEN
SHOULD YOU USE IT for "Grand Challenge Problems" which need
decision making abilities that out-perform humans.

In many cases, neural nets will be asked to perform tasks that require
general subject matter expertise but NOT super-intelligence.
These neural nets will be used to replace lower-level and/or less skilled
workers at attractive price points for small to medium sized businesses.

Higher-performance neural nets will have the expertise of MANY
experienced and WELL-REGARDED subject matter experts built
into their logic tree databases. These neural nets will be VERY
EXPENSIVE but many business people will be HAPPY to pay dearly
for them if it will improve their bottom line quickly!

Others will WANT TO HAVE the power of many well regarded experts
within a little black box which will be used to tackle those
grand challenge problems which currently require many
expensive scientists or professsional-level experts.

And since neural nets can ALWAYS be expanded, as subject matter
experts are brought on board to see the beauty and power of having
MANY-EXPERTS-INPUT types of neural nets available, I can see
a burgeoning market of scientists, legal personnel, MBA's, technical
experts and yes even artists, being paid BIG BUCKS to allow
THEIR KNOWLEDGE to be input into a wide variety of saleable
neural nets. I can even see having a ROYALTY system in place
where for every neural net problem that uses their expertise,
a set of percentage-based or fixed-rate payments will be made
to them...!!!!

On the downside, this is NOT GOOD for the lower-level or less skilled
worker as THEIR VALUE in terms of academic or expertise input
will be SEVERELY limited and eventually discounted as MANY
newly borne neural nets WAY OUTPERFORM AND OUT-PACE
their inherent abilities. Unfortunately, it is already TOO LATE,
as people like ME begin to offer the software and payments
infrastructure for creating, selling, renting, leasing, licencing
and managing general and task-specific neural nets.

---

My test is beginning to start its re-layering and thus
will start the final edit process REAL SOON NOW!!!!!!!

Today I'm not going to say anything in terms of final finish time,
because it is OBVIOUS that my earlier time estimates have been WRONG!
I still believe the final result will showcase that neural nets CAN be used
in multimedia production, but that they do need some work to get to a
TRULY professional level of quality...HOWEVER...I can countermand that
statement by saying that it is ONLY A MATTER OF TIME and enough
computer horsepower to be able to put a Band-in-a-Box or
Picasso-on-a-Disc for EVERYONE to play around with!!!!

I'm in Seattle right now and will be back Thursday!

[edit on 2010/7/20 by StargateSG7]

[edit on 2010/7/20 by StargateSG7]

[edit on 2010/7/20 by StargateSG7

[edit on 2010/7/20 by StargateSG7]

Sounds great, I'm very interested in neural nets. It sounds like you'll be trying to create a whole outlet or even career path for professionals. But again I'll ask, are there any neural nets you know of that look like they're coming close to their objective? I'd love to check it out..

Originally posted by J.Clear
Sounds great, I'm very interested in neural nets. It sounds like you'll be trying to create a whole outlet or even career path for professionals. But again I'll ask, are there any neural nets you know of that look like they're coming close to their objective? I'd love to check it out..

---

Today there are TWO SETS of well-known rule-based
"Expert Systems" aka Task Specific Neural Nets which
have been VALIDATED for many years now!

These are the Visa - Mastercard - American Express
purchase/payment validation systems which check to see
if there are illegitimate purchases and/or payments made
using your credit cards.

While not true general purpose neural nets...they are in fact
task-specific rule-based expert systems....Visa/Master Card
/AMEX have used these system to both prevent credit card fraud
and examine purchase and payment patterns for advertisers.

Another set is the FICO score (Fair Isaac Company)
which uses 350+ rules to set your credit score which
can range from simple criteria such your as debt to equity ratio
all the way to such estoteric criteria as how many hospital
insurance claims you've had within an X period of time (i.e. there
is a valid statistical correlation between the types and number
of medical insurance claims and people who have problems
paying general bills such as credit cards and mortgages)

These expert systems ARE USED DAILY and their expertise
levels are being raised higher and higher as more and newer
rules are verified and added to their logic tree databases.

I have given some links below of systems which I PERSONALLY
HAVE RESEARCHED and read up upon...and at last count I have
found over 35,000 neural nets being used for everything from
fraud prevention to building design to surf board design and
manufacturing to giving legal advice in specific areas of expertise.

I can confidently predict that neural networks WILL BE ABLE to
match and soon OVERTAKE human abilities in MANY AREAS
of scientific, legal and artistic inquiry.

I personally know of 50+ neural nets in use TODAY that solve common problems such as:

---

Composing Music with LSTM Recurrent Networks
Blues Improvisation:

www.idsia.ch...

---

Music Classification Made Easy with Neural Networks

news.softpedia.com...

----

Sentient Autonomous Vehicle using Advanced
Neural net Technology

infolab.stanford.edu...

---

Intelligent video editing system using a
neural network coding scheme:

spie.org...

---

Intelligent Flight Control System for Automous Piloting of Aircraft:

www.nasa.gov...

---

Ship Identification Using Probabilistic Neural
Networks (PNN):

www.iaeng.org...

---

integrating neural networks and rule-based reasoning
in the legal domain:

portal.acm.org...

---

Detecting Patterns of Fraudulent Behavior in Forensic Accounting:

www.springerlink.com...

---

Managing Patent Legal Value via Fuzzy Neural Network
Incorporated with Factor Analysis Based on Patent
Infringement Lawsuits :

ieeexplore.ieee.org... cision=-203
---

Optimum design of ship design system using neural network method in initial design of hull plate:

www.springerlink.com...

---

Optimization using finite element analysis, neural network, and
experiment in tube hydroforming of aluminium T joints:

pep.metapress.com...

---

Artificial Neural Networks for Finite Element
Modeling of Giant Magnetostrictive Devices:

membres.multimania.fr...

---

As you can see the list barely covers a small subset of the
full range of what many think are tasks reserved for
human abilities.

So I can say to you who disbelieve that your job is SAFE
from nearly sentient software-based neural networks,
look again! Hasta La Vista Baby... they're coming SOONER
THAN YOU THINK.....!

In my case, I am using some brute force DSP and neural net-based
filtering methods to cut out human voices from background audio
and replacing it with another voice transposed to another gender
but keeping the original timing and inflections as much as possible.

I still cannot say with certainty as to the qualitative results
of this test....I'm hoping for the best ... but will probably end up
with something hilarious!

I will keep you updated....!!!!!

Up to 187 processing hours (i.e. Aggregate Network CPU clock time
NOT linear time)...the re-layering process is happening
and will continue for PROBABLY 25 to 35 CPU hours more.

Sorry! I can't make it go any faster....!!!!

Sweden - Pirate ISP launched

like a vast library of endless books
endless and endless and endless looks

teh futar ees hear mang. :B

Originally posted by Skyfloating
Sweden - Pirate ISP launched

Pretty big stuff! I think this thread would be a good reference also for that thread..

I gotta say, this thread has made me think about dling and torrent. I was thinking about a certain new movie and going to see it or some other internet means and the phrase popped into my head, "no more [company making wonderful animated films] if everyone's filesharing". The extreme case, of course, but a definite future scenario put forward here by someone early in the thread, big companies and expensive productions simply not being able to exist anymore. I think there's an awful lot of dross but it would be a huge loss to lose some of the more creative companies making worthwhile contributions.

reply to post by J.Clear


Thanks for referencing this one in the other one...Im hesitant to reference threads I OP'd .

Im following both threads because I think this is one of the major issues of our time (despite many not having woken up to the fact of it being an issue yet).

[edit on 26-7-2010 by Skyfloating]

The discussion on the other thread is already starting to veer into what was covered here. No problem

Just an update of we are at 216 hours of total
aggregate network processing time for the
Voice Replace test.

---

Not to worry about the initial length of time this test is taking.
As the Neural net does this more and more often, it will get faster
over time because it will learn what to cut out of it's processing.
Right now I'm doing a brute force DSP and audio raytrace which
is thorough but not necessary in future iterations.

It's time consuming but effective in "teaching" a neural net
what to avoid or skip over next time. My final result will be
decent and hilarious and put up on youtube.

3 weeks linear time but that is ONLY because our internal
network is being used for paying jobs and thus this task
is set at a lower priority. MANY home users will be able to
set this at a high-level priority during the day and night
and be able to take advantage of newer multicore systems.

Remember that this is a test of the ability of software to create
multimedia content AUTONOMOUSLY. It will INITIALLY take
longer than a skilled human but the results will eventually
get to a point where it could be considered truly professional.

I will update as this gets down to the final
nitty gritty processing tasks which right now
I look with a bit of egg on my face in terms
of time estimation but I'm not discouraged.
We WILL get through this!

By the way... I hope those web links on what Neural Nets
are doing TODAY will help you understand that software
IS GETTING BETTER and BETTER at doing what we humans do!

This is to be expected!

WE HUMANS CREATED THE SOFTWARE !!!!

Therefore, I see no reason and no obstacle as to why
software CAN'T equal to better our ability to do everyday
things and provide even technically or artistically
sophisticated services or creations!

And like humans who go to college and perpetuate
a culture of higher learning in THEIR children, I see
no reason why many learned scientists and artists
would not group together to create digital versions
of themselves which will become highly proficient
and SKILLED at their designated tasks as the software
begins to LEARN BY ITSELF and begins to CREATE NEW RULES
of both learning and doing!

Originally posted by StargateSG7


Not to worry about the initial length of time this test is taking.
As the Neural net does this more and more often, it will get faster
over time because it will learn what to cut out of it's processing.
Right now I'm doing a brute force DSP and audio raytrace which
is thorough but not necessary in future iterations.

One tiny point on this.. how many times will it get the opportunity to do this though? In a studio, surely the day-to-day tasks will vary quite a lot, depending on the style of music being recorded. Also, the artistic element surely also demands some off-the-cuff creativity to make it unique. Maybe the neural net would never get enough experience at a task to be as effective as a human expert because it would be useless to repeat the task. How will neural nets develop a human level of this kind of flexibility?

Originally posted by J.Clear

Originally posted by StargateSG7


Not to worry about the initial length of time this test is taking.
As the Neural net does this more and more often, it will get faster
over time because it will learn what to cut out of it's processing.
Right now I'm doing a brute force DSP and audio raytrace which
is thorough but not necessary in future iterations.

One tiny point on this.. how many times will it get the opportunity to do this though? In a studio, surely the day-to-day tasks will vary quite a lot, depending on the style of music being recorded. Also, the artistic element surely also demands some off-the-cuff creativity to make it unique. Maybe the neural net would never get enough experience at a task to be as effective as a human expert because it would be useless to repeat the task. How will neural nets develop a human level of this kind of flexibility?

---

In my case I have over 10,000+ HOURS of audio & video ranging
from court case depositions to training videos from corporate
clients ranging from lawyers to aerospace companies to military
to food processing and whole lot more. Most of this material
is STILL ON Betacam SP or DV format videotape and thus needs
to be transferred ANYWAYS to hard disk.

I can always get our transfer software to send it through
our neural net for examination processes to use as a
"Teaching/Learning Mechanism".

One of my goals is to use all that video to TEACH our general
Midgrid Neural Net software how to RECOGNIZE faces and voices
from the various Technical & Conversational languages used
within all 10,000+ hours of that old video.

We can teach it what traits are considered to be Human, Male,
Female, Child, Teenager or Adult faces, voices and mannerisms
all the way down to differentiating between ethnic or educational
levels and backgrounds.

Ergo, I can teach Midgrid to RECOGNIZE and UNDERSTAND
what characteristics contribute to a definition of humans
that are of a specific gender, age, educational level or
physical ability...including "Physical Tells" as to whether
a human is Lying or Telling the truth.

I could also teach it to scan through multiple videos
(at Terabytes per minute speeds!) to find out what is
a Red Corvette or what is a Book or Piece of paper
or what is a chair and what BRAND or TYPE it is. I can even
teach Midgrid what is considered Good Quality Video and Sound
and what is Bad Quality Audio & Video.

That tens of thousands of hours of ULTRA HIGH QUALITY video
is a SUPERB base to begin Neural Net Training and we estimate
its technical/teaching value to be in excess of tens of millions
of dollars worth of high-level TRAINING and LEARNING material
for top-notch Neural Net initialization.

This means Midgrid will START OUT with Built-in object recognition
and language & text processing abilities that would be EQUIVALENT
to a high-level expert.

For a few bucks, you get a SUPERB Print to Digital Text,
Voice to Text, Text to Speech and Audio/Video Search tool
that would be a LOT FASTER and a LOT MORE ACCURATE
than MOST HUMANS. Midgrid would be able to work
24/7/365 and get BETTER & FASTER over time!

And since Midgrid is a GENERAL NEURAL NET you can teach it
ANY task from Welding to Music Composition to Video Editing
to Designing CPU chips...It'll be UP TO YOU as to what the limits
(if ANY!) are as to what MIDGRID will do for your business
or personal life.

Our key is to SPEND SERIOUS MONEY on quality subject matter
experts to infuse their expertise into a machine readable form
in order to create Hundreds of Thousands of "Rules of Thumb"
which will allow the Neural Net to perform and/or simulate almost
ANY task we give it at a High Level of accuracy AND expertise.

It is the Initial Quality of these Rules of Thumb which will give
Midgrid a SERIOUS ADVANTAGE over OTHER commercial neural
nets which were initialized by PROGRAMMERS rather than
proper subject matter experts!

And because I've had to learn via the "School of Hard Knocks",
we've found out the hard way that Subject Matter Expert
Experience and Ability really do COUNT when it comes to
initializing a Neural Net with a properly configured
Multi-state Boolean Logic Tree Database!

Like with ANY child, INITIAL growing conditions and/or
environment REALLY DO MATTER as to how WELL the final
skill set and inherent abilities of a Neural Net will turn out!

Garbage In = Garbage Out applies here!