Activation Functions - Softmax 

so for anyone wondering. there is an issue in the derivative of this one! It is not as simple as I stated. I was a bit younger when I did this so please have mercy with me :) The rest should still be correct.

Thanks for the most concise and forthright explanation of the softmax activation function I've ever seen!

The best explanation I've heard of.

Amazing explanation, I loved the fact you took some example numbers, did the calculations, and showed how to values are modified by the function. Really got the point home. Keep it up!

Good explanation 10/10

Damn, why couldn't everyone explain it like this? I am dumb and I need an explanation like I am a 5-year-old but most of the explanations on the internet assume that we are all smart as fuck. Thank you!

Thank you so much for the explanation!

Great, found the difference between softmax and sigmoid. Thanks

GREAT VIDEO ,Thank you !

At 5:30 the light bulb went on. THANK YOU! :)

Thanks, easy to follow👏🏾👏🏾

I Love it :) thanks

Excellent explanation.

Perfect ✌🏻✌🏻

I think what you mean in the last part about the difference between using a Sigmoid or a Softmax for classification is that for a binary classification problem you only need the probabilities of the two outcomes and a threshold, let's say if I predict A with over 50% probability then my prediction is A, otherwise my prediction is B. For a multi-classification task however you want to normalize over all possible outcomes to obtain a prediction probability for each class

at @2:06, there seems to be a typo. x should be [0,1,2,3,4,5] instead of [1,2,3,4,5,6]. f(0) = 1/1+e(0) = 0.5; f(1) !=0.5

I wonder in which cases it's advantageous to use softmax over using percentages of the total sum? Numerically it seems softmax is good for separating big values from smaller ones: EDIT: **googles** Apparently is exactly that. To make high values more evident.

You just got a new like and subscriber

I love your videos!!!!! It helped me create my very first AI ever! Ur tutorials are so concise! I was wondering, if you know how to do it, could you make a tutorial on q learning in Java, then deep q learning in Java, the deep q learning being something I have been struggling with implementation

This (3:20) little thing tells a lot about you and is the way to reach more subs. Thanks!

There’s no problem with sigmoid , all activation functions have their uses

thanks a lot

Excellent explanation

Super excellent

respect

Thanks, Finn- helpful! I don't think you mention why you need the exponential functions in the Softmax definition. If you showed some negative example values as components of your a-vector (totally legitimate outputs of e.g. a layer with a tanh activation function) it would be easier to see that without them, you wouldn't be guaranteed probabilities bounded by zero and one.

the problem of softmax is about the derivative form here math.stackexchange.com/questions/945871/derivative-of-softmax-loss-function?rq=1 u must consider two case 1. i not equal to k and 2. i equal to k I have calculate it by myself but I m not sure if it is right can u go through it once?

thanks for the explanation Finn!! I have a question. whenever ı google "derivative ofsoftmax function ı always find something like this " eli.thegreenplace.net/2016/the-softmax-function-and-its-derivative/ ". I am working on a project that is about pure java implementation of multi Layered NN can you help me how can ı use derivative of softmaxfunction.

why do we use e?

Congrats on making this so simple to understand, you actualy know what the function does. I sometimes wonder if people actually even understand the content they reproduce or are just too lazy to try to put things in a way others can understand. Einstein did famously say: "If you can't explain it simply you don't understand it well enough"

why do we use "e"?

thanks for the explanation.

what is "e" ? how i can get the "e" value ?..