Slashdot Mirror
New Google Project Lets You Collaborate On Doodles With A Neural Network (tensorflow.org)
Long-time Slashdot reader Giant Robot writes:
Google Brain's latest experiment is a neural network that allows you to collaboratively draw with it inside of your web browser in real-time. The neural network is trained using the drawings collected from an earlier web game called Quick, Draw! released a few months earlier.
"Once you stop doodling, the neural network takes over and attempts to guess the rest of your doodle," explains Google's page about the project, adding "You can take over drawing again and continue where you left off."
"Once you stop doodling, the neural network takes over and attempts to guess the rest of your doodle," explains Google's page about the project, adding "You can take over drawing again and continue where you left off."
This project will make many headlines, and some people will try to build a business around the api. Google will then close it in three years because only a few hundred thousand people use it regularly.
Any guess?
OK I tried it and I was not exactly whelmed. Forty years ago I was playing with predictive pattern matching models in an attempt to create a Go player. Although the basic idea worked it took too much expensive computer storage (CDC 7600) to actually play a game on a 9x9 board in real time. (The processor was actually fast enough.)
This exercise is pretty much the same but this software has the benefit of lots of cheap processor cycles and storage space. (I'll spare you the numeric equivalent calculations from the 7600 to modern hardware.) And instead of winning combinations it just guesses stuff with a wide range of possible positive feedback responses.
We will see if Alphabet's infinite money supply can make something out of this. Personally I just don't see much but I have been wrong on the "vision" thing before.
Trump doodles on the bills heâ(TM)s supposed to read while imagining having sex with Ivanka!
What could possibly go wrong?
I suggest we all collaboratively draw erect cocks to screw up this IA
Neoron network is which?
Several users, whether for privacy or anti-malware reasons, have decided to abstain from running JavaScript at all, including many who replied to this story. If blocking all scripts becomes commonplace, how will things such as "Google Brain's latest experiment" be built? Will such experiments instead need to be wrapped in Electron for Windows, Electron for macOS, Electron for X11/Linux, and whatever is used to package web apps on mobile? Or would people who do not tolerate JavaScript instead tolerate a clunky workaround using server side image maps?
And Google added the spokes! Man, that thing is really smart. Next up: boners!
I assumed Quick Draw was used to build a training data set for the GBoard "draw an emoji" feature.
... boring.
It little behooves the best of us to comment on the rest of us.
lol :)
A good engineer can reverse engineer an algorithm with a number of test inputs. This POS showed its 'advanced neural method' in a few tries.
Draw the first obvious geometric outline (squashed oval for 'pineapple', some sort of quad for 'garden', a rounded cone for 'windmill' etc) and the 'algorithm' finds weights for this outline (pos, size, orientation) and places obvious 'features' (legs of 'spider', blades of 'windmill) to correct scale, position. Big whoop.
Any more sophisticated input (ie., the actual complete form of the 'object') and the 'algorithm' effectively draws random junk on top. It's nothing but a con, like all so-called new-age 'AI' (which is actually a simple pattern matching algoritm exploiting the growth in data storage and data searching/processing).
Did you know that the 'advances' in voice recognition only happened when 'smart' methods were dropped, and massive indexed databases of pre-existing voice samples were adopted instead, allowing recognition to be done by a 'closest fit' method? Earlier the same idea lead to a breakthru in written language translation, when UN records (where everything a member sez must be translated and recorded in a myriad of other languages accurately) were compiled into massive indexed computer databases. Then pattern matching became the translation method, dropping all 'smart' methods that tried to find semantic language rules.
Computers aren't smarter than Humans (the big Slashdot lie to boost 'AI' investment). They simply process far more data at a vastly greater speed. They literally work hard, not smart. And now CPUs are so fast and cheap, and storage so fast and cheap, the 'best' algorithms are the dumb ones using vast amounts of storage and data processing.