A little GA Tech humor/puzzle.. - Subaru Enthusiasts Car Club of the Sierras

JC · 2005-10-12, 07:14 PM

I got this in an e-mail today, thought I'd share.

Quote:

Your advisor has left you in charge. A pesky grad student comes up to you and states that they have developed a new Neural Net. Each perceptron in the Net has an output that is a summation of its weighted inputs (no thresholds).

You yawn and state that the Net is uninteresting. Why?

(No, its not because you haven't been getting enough sleep)

lol

dknv · 2005-10-12, 08:42 PM

Lord, JC, what are you taking? Quantum physics?

Is the answer, because the net is a male model? ( lol.)

Good stuff, keep it coming, our (my!) brains can always use some weighted inputs.

ScottyS · 2005-10-13, 09:31 AM

Quote:

You yawn and state that the Net is uninteresting. Why?

Because you know that a surplus of grad students creates a scarcity of interesting projects --- as they all vie to promote their own project as critical to the advancement of science, when in reality it's just another useless niche with great buzzwords.

sperry · 2005-10-13, 09:34 AM

Sounds like that Neural Net can't learn anything. Even if the weights are being modified during the learning procedure, each perceptron in the layer will be outputting the same value to the next layer, thereby removing any learned behavior. Essentially, the network will act as a simple adding machine where the result is scaled by the learned weights.

...I think, I never really learned much about neural networks.

And JC, stop getting other people to do your homework.

Nick Koan · 2005-10-13, 09:38 AM

Now, correct me if I'm wrong, but won't each node only recieve and transmit the value 1?

MPREZIV · 2005-10-13, 10:59 AM

wow. my head hurts.

ShawnS · 2005-10-13, 05:58 PM

Quote:

Originally Posted by MPREZIV

wow. my head hurts.

Alcohol will fix that.

Dean · 2005-10-13, 06:09 PM

Sounds like a cascading pyramid of nodes resulting in a flat "true" at the edge of the net. Boring....

sperry · 2005-10-14, 10:03 AM

Quote:

Originally Posted by Dean

Sounds like a cascading pyramid of nodes resulting in a flat "true" at the edge of the net. Boring....

If all inputs are 0, then the output will be zero as well. It's essentially a massive, complicated OR gate.

Dean · 2005-10-14, 01:12 PM

Quote:

Originally Posted by sperry

If all inputs are 0, then the output will be zero as well. It's essentially a massive, complicated OR gate.

Sorry, I assumed there would be some actual true bits somewhere in the data. But yeah, a giant OR net...Whopie

cody · 2005-10-14, 02:37 PM

Um, 9?

SlickNick112 · 2005-10-17, 06:52 AM

Quote:

Originally Posted by Dean

Sounds like a cascading pyramid of nodes resulting in a flat "true" at the edge of the net. Boring....

Ha ha ha, he said nodes!

JC · 2005-10-21, 09:50 AM

Answer:

The Neural Net is uninteresting because, *no matter* how large and complicated it is, it can be replaced by a single perceptron (node) for each output.

Training a single node is far faster than training 500!

A single perceptron here represents a linear combination of the weighted inputs. Without thresholds, the output will always just be a linear combination of the weighted inputs.

The perceptron can represent a vector in an N dimensional space, where N is the number of inputs. If you add two/three/a billion vectors together you just get another vector.

So the entire Net only has the power to represent a vector for each of its outputs.

This means that it can be replaced by a single node for each output with no loss of representational ability.

sperry · 2005-10-21, 09:57 AM

Cool, looks like we were on the right track for the answer, but I like the way it's phrased better in JC's solution.

cody · 2005-10-21, 09:59 AM

Damn, so close...

sonicsuby · 2005-10-21, 11:22 AM

Quote:

Originally Posted by ScottyS

Because you know that a surplus of grad students creates a scarcity of interesting projects --- as they all vie to promote their own project as critical to the advancement of science, when in reality it's just another useless niche with great buzzwords.

I think this is the real answer.

2005-10-13, 09:34 AM	#4
sperry The Doink Real Name: Scott Join Date: Nov 2002 Location: Portland, OR Posts: 20,335 Car: '09 OBXT, '02 WRX, '96 Miata Class: PDX/TT-6 The way out is through	Sounds like that Neural Net can't learn anything. Even if the weights are being modified during the learning procedure, each perceptron in the layer will be outputting the same value to the next layer, thereby removing any learned behavior. Essentially, the network will act as a simple adding machine where the result is scaled by the learned weights. ...I think, I never really learned much about neural networks. And JC, stop getting other people to do your homework. __________________ Is you is, or is you ain't, my con-stit-u-ints?

2005-10-13, 09:38 AM	#5
Nick Koan JDM Cowboy Real Name: Nick Join Date: Oct 2003 Location: Somewhere Posts: 8,642 Car: 2015 Mazda 3	Now, correct me if I'm wrong, but won't each node only recieve and transmit the value 1? __________________ While a standard engine is powered by a belt connected to the crankshaft, a turbo engine runs on its own exhaust steam, making it more energy efficient. -- CNN

2005-10-13, 10:59 AM	#6
MPREZIV Token Real Name: Le Stig Afrique? Join Date: Jun 2005 Location: sitting next to a big yellow box Posts: 3,589 Car: 2001 Impreza 2.5 RS Class: 05 TDSP No, I won't work on your car. F your car*	wow. my head hurts. __________________ "...these condoms have a topical anesthetic to reduce sensitivity, so you can last longer. What a paradox. You can't feel a thing, but you can f*ck for HOURS..."

2005-10-13, 06:09 PM	#8
Dean Señor Cheap Bastarde Real Name: Dean Join Date: May 2003 Location: $99 Tire Store Posts: 9,294 Car: $.04 STI Class: Fast,Cheap & Reliable=STI Deal, did somebody say Deal? Oh, Dean, yeah that's me.	Sounds like a cascading pyramid of nodes resulting in a flat "true" at the edge of the net. Boring.... __________________ I am a Commodore PET --- Now get off my lawn you kids...

2005-10-14, 02:37 PM	#11
cody Candy Mountain Real Name: Cody Join Date: May 2005 Location: Californication Posts: 7,751 Car: 03 Pussy Wagon, now with more pink! Class: TESP OMG Internet!	Um, 9? __________________ Slow and low, that is the tempo.

2005-10-12, 08:42 PM	#2
dknv EJ207 Join Date: Dec 2002 Location: 39n53, 119w90 Posts: 2,698 Car: RX-8 Class: CS maybe	Lord, JC, what are you taking? Quantum physics? Is the answer, because the net is a male model? ( lol.) Good stuff, keep it coming, our (my!) brains can always use some weighted inputs.

2005-10-21, 09:50 AM	#13
JC Ask me about dubs! Real Name: JC Join Date: Dec 2002 Location: Orange, CA Posts: 2,895 Car: 2013 Triumph Speed Triple R	Answer: The Neural Net is uninteresting because, no matter how large and complicated it is, it can be replaced by a single perceptron (node) for each output. Training a single node is far faster than training 500! A single perceptron here represents a linear combination of the weighted inputs. Without thresholds, the output will always just be a linear combination of the weighted inputs. The perceptron can represent a vector in an N dimensional space, where N is the number of inputs. If you add two/three/a billion vectors together you just get another vector. So the entire Net only has the power to represent a vector for each of its outputs. This means that it can be replaced by a single node for each output with no loss of representational ability. __________________ Actually, I am a rocket scientist.

2005-10-21, 09:57 AM	#14
sperry The Doink Real Name: Scott Join Date: Nov 2002 Location: Portland, OR Posts: 20,335 Car: '09 OBXT, '02 WRX, '96 Miata Class: PDX/TT-6 The way out is through	Cool, looks like we were on the right track for the answer, but I like the way it's phrased better in JC's solution. __________________ Is you is, or is you ain't, my con-stit-u-ints?

2005-10-21, 09:59 AM	#15
cody Candy Mountain Real Name: Cody Join Date: May 2005 Location: Californication Posts: 7,751 Car: 03 Pussy Wagon, now with more pink! Class: TESP OMG Internet!	Damn, so close... __________________ Slow and low, that is the tempo.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
OEM skid plate for MY 04+ WRX $20	cody	User Classifieds	18	2007-01-22 07:05 PM
Solo Annual Tech Inspection	Kevin M	Motorsports Chat	12	2006-03-30 04:34 PM
Tech inspection for track day	doubleurx	Motorsports Chat	14	2004-08-30 11:47 AM
Update***Track Day Pre-registration and Tech Inspections!	MattR	Motorsports Chat	9	2004-08-25 03:23 PM
Clutch replaced	dknv	General Subaru Discussion & Club Chat	19	2003-03-13 09:49 AM