Neural Networks Lesson 2: Probabilisitic Neural Networks

[dougfinn2]

Very nice - but you could lose the mournful and distracting background music.

[bigfong]

at 1:59, does the area under the 3 curves all add up to 1? Thank!

[PhilosopherEight]

Can anybody recommend any Neural Network freeware for PC that will implement PNN, as well as other methods? Something that doesn't require me to write C++ code, but will allow parameter flexibility and easy file I/O would be great. Does such freeware exist anywhere?

[darkness9484]

@PhilosopherEight

try neuroSlab. it's a nice freeware software. just google it.

[PhilosopherEight]

@darkness9484 Thanks for the help. I looked at it and it doesn't seem to process time series though, only images. Anything available for time series?

[darkness9484]

@PhilosopherEight

take a look at Zaitun Time Series. it should be good enough for your needs...

[PhilosopherEight]

@darkness9484 Zaitun looks cool, but it doesn't implement PNN like I need. Backprop is useless for me.

[wsubsee]

Please add more videos about neural networks.

[mohitrathore15]

Great stuff. Please add more videos about other applications of neural networks

[ph0chiz0]

*facepalm*

[Houshalter]

I don't get it.

[hyperthreaded]

This is great. I can only imagine how much work went into putting these slides together. It's very instructional and incredibly useful. Thank you very much.

[jadlerhalofan]

You really do an amazing job of explaining these, please do more and avoid assuming knowledge on the part of the viewer. I'm a CS major, but I'm not up to some of that math quite yet. :)

[TheNoodlyAppendage]

OK, so you gave us a high level description of gaussian methods, but how about actually teaching something like showing step by step how you would generate the weights on paper, not using handwaving and pseudocode.

[fatalmystic]

@TheNoodlyAppendage true, it would be very interesting to see how to put that all together step by step - on paper. but one has to say thanks for this great lecture! that's how education should be: high quality and free! ;)

[N3CR1S]

@TheNoodlyAppendage One way, which i use, is by using a genetic algorithm. make many "chromosomes" (structures) containing the values for the weights, different amount of neurons, the thresholds for each neuron. Run the data to every "chromosome" and order them according to which hade the best results. make the winners have offspring, merging their values, allow mutationrate and overcrosses.

after some (10 -1000) generations, you will have good weights and a good neural net.

[rubbercraft]

But for your example problem, with relatively evenly spaced class points, and a uniform gaussian PNN functions with a size that was fudged-compatible with the sample class data points, it works well as a stripped down version to visualize the PNN application idea.

However, saying no further training is required, is only true of YOUR example, and not general PNN, which can adapt the gaussian scale and rotate the kernels as the training data requires.

[rubbercraft]

I understand the examples, but I see great oversimplification of the actual problem. All gaussians shown were of the same X and Y scale and all the same size. In real PNN you would often require gaussians that can be scaled and rotated, not just placed on a training point. Why? To minimize the misclassification errors due to heirarchical scaled clustering and spread of differing class points where the classes are still unique, but may come very close to each other.

[Seantron13]

Please do more. These are amazingly helpful.

[CorporationsSuck]

Shouldn't the value on the right be .62?

[SewTroy]

no, because the maximal response is .61 :)

[theantichristchannel]

But at 2:32 it says that the pattern units are summed thus the blue ones are .61 and .01 which equals .62

[lautarocarmona]

Is this related to GMM (Gaussian Mixture Models)? Is like an hybrid of NN and GMM?

[nickrohn93]

this is a bit beyond my algebra 1 math experience.