Lecture - 28 Back Propagation Learning

[trisky1234]

Test coming tomorrow. This helps, thanks!

[rflood89]

im only 9 mins in so far but already THANK YOU! :)

[hxmstu]

Love the moustache my friend. LOVE IT.

[ZERL1NG]

Excellent explanation, thank you very much!!!!!

National University Federico Villarreal

Computer engineering student

Lima, Peru

[daviddalbylive]

I have a back propagation neural network, its my favorite NN.

[hmjkg]

Slightly off-topic, but I love that intro. Anyone knows the artist?

[fsl4faisal]

@hmjkg me too

[manuGori7]

i realy like this videos

i am fan of prof.p.dasgupta

[genghiskii]

thank you for the lecture, if you train with the first sample, the calculated answer will be closer to the real answer of the first data, after that you train with the second sample, how can you guarantee that the when you input the first sample again, the error won't be larger?

[AndreasBacher85]

@genghiskii This is because you calculate the change of the whole error E by changing the weight Wi,j - not just the vector component error Err.

[25380421859]

wow nice handwriting

[selotmani1]

@25380421859

I think the same

[1888junkteam]

excellent work!

[sabriath]

I personally use my own algorithm:

=max(1-(T(P-x))^2,0)

T is the tightness (0.1 to start and increment)

P is the point of interest (I use 0.5, works fine)

This creates 1.0 if it matches the point and grows outward from the point in BOTH directions reaching 0 at the same distance from the point, the speed of the drop is determined by the tightness. Easier to calculate, and the error function is easily streamlined along with the feedfowarding.

[1shiOne]

Another question:

Should I correct the value of the bias as well ?

[MsBernis]

what is the difference betwee activation function and transfert function.

[thisgonabeegreat]

I youst making kandi this tuf

[1shiOne]

Nice lecture !

But I have a question. At 34:00 he is using g'(x) instead of g(x). What is g' ? g was the activation function (sigmoid).

[KirusRaysor]

g'(x) is the derivative of g(x)

If g(x) is choosen to be the sigmoid, then g'(x) would be:

g'(x)=g(x)*(1-g(x))

[OriginalAtomicSheep]

Gradient of the sigmoid?

[feebdaed]

yes, to find the direction to reach a minimum

[sajinkoroth]

Great Lecture.

[berjberj1]

interesting

[viviom]

Thank you , this lecture it was helpfull!!

[vangtid]

great !

[rg071970]

Very good, thank you, you talk slow and clear. thank you...

[rohit85727]

it will be better if figures ar pre-drawn and more stress should be given on explanaton.

[rakimo]

Fantastic lecture. You really have the gift of teaching in an understandable way.

[tweety504]

so nice...

[aan1923]

lovely...