Neural Networks from Scratch - P.4 Batches, Layers, and Objects 

lol you are weird xD

I literally wait for RU-vid’s notification telling me Sentdex dropped a new tutorial 😂

@@usamanadeem7974 me too brother ahahahah

same here

cringe

My week starts when he drops video!!

@@akiratoriyama1320 countdown starts today for the next video

Agreed, I'm waiting patiently also :)

i think you might have to make new days cuz it is taking 2 weeks for his next vid

I was going to say the same thing

yes, absolutely. i come from a time where programming meant you buy a book that acually was outdated the moment you bought it. thank you 🙂

I agree with you, although you can check out Deeplearning.ai on Coursera. It's pretty good.

@@lucygaming9726 No thanks . Im too poor for that.

@@aleksszukovskis2074 its free and by andrew ng, the legend

@@janzugic6798 thanks

@@janzugic6798 you need a coursera subscription ($49/mo) after 7 day trial period regardless of the course being free

Please clarify the concept of the Gaussian Distribution that you introduced when talking about np.randn

if i draw a neural network of 12 inputs imaging into 3 output and connect each neurons to the output, there will be 36 lines in total, that means there has to be about 36 weights but the weight you took had only 12 weights in array, how is that possible ?

@@anjali7778 He has only 4 inputs to the output layer... therefore number of weights = 4*3 = 12 If instead , you have 12 inputs ,you will get 12*3 = 36 weights

Am I wrong, or is there something missing around the 9:08 point?

Question: why did we need to transpose weights 2 since they are both 3x3 matrices, index1 of the would equal index 0 right?

ASMR for eyes, thanks Daniel!

@@Saletroo ikr 😂😂

No plans to stop any time soon!

Glad to hear it!

better than most ivy league schools

I banged my head on numerous videos too. They assume a level of knowledge that was hard to peice together. This series is filling lots of gaps for me. The concepts are starting to jell, this whole field is fascinating!! Kind of empowering.

They are there sometimes :) but yes fairly rare.

-comment

I was looking for this comment. Thanks for pointing that out!

Sister: "If that's informative, then what's educational" Me: "Glad you asked!" *starts to explain neural networks and basic QP* Sister: "NO! Make it stop!" *Never asks again*

Story of my life 😁

Youssef I really hope @sentdex reads this ;)

I think it was explained in a previous video, how biases help making predictions. Check out the last video guys

He explained that in previous videos, but no all your questions

@@carloslopez7204 I agree it was explained a bit but I really didn't feel the explanation gave me a deep understanding of the why unfortunately, just a very rough surface level and vague hint of what might be going on.

I think that now some things, like the biases dont make sense now. But when you get into training(the lerning process) is all start to make sense.

Glad we could help!

x = we're arriving at the sexy parts... print(x)

Yeah, that really confuse me more than it should have

yes. It is called broadcasting.

Because matrix multiplication is not commutative

I think it is the nature of what we are doing - We are taking inputs - applying weights and biases - and delivering outputs. or entering exiting a decision. so we can't use an entrance to a neuron to exit another neuron. I think the demonstrations by Harrison are to cement the concept and awareness of ValueError shape()... and he also showed how multiplication works between array and vector (multiplication) I went back to lesson 3 for 2 things. 1. I like inputs being the first entry so 'My doors' are labelled correctly 2. Use the npArray().T in that example If he had not shown us 3 before 4 - I would have found it harder to appreciate transpose() - I don't think I will ever just reverse the args when I am coding this stuff. import this

Something that I find interesting that I think might have to do with this is that specifically in the case of 1 dimensional arrays, the shape is different. I am used to thinking of a matrix as rows by columns. For example, [[2,2],[3,3]] , would be a 2 by 2 matrix. 2 rows and 2 columns; however, lets take the example [1,2,3,4] , I would have expected the shape of this to be 1 by 4 ( 1 row and 4 columns) but it is not. The shape of [1,2,3,4] is (4, 1) . So, the way to think about it is by elements in a list of lists. The first entry x in shape (x , y) is how many lists in a list and the second entry y is how many entries within each element or list. In his example, his first inputs [1,2,3,4] the shape is (4,1) and when he put [[1,2,3,4],[1,2,3,4],[1,2,3,4]] the shape became (3,4) and if you are thinking about this in rows and columns that wouldn't be the case. I hope that made some modicum of sense lol

You could perform the same operation by transposing the inputs however keep in mind the matrix rule (A.B)' == B'.A' , e.g. inputs.(weights.T) == (weights.(inputs.T)).T aka. the output of inputs.weights_transposed will be equal to the transposed output of weights.inputs_transposed, the issue with the values probably comes from adding the biases without first either transposing them or transposing this output matrix back, as they will be added in a completely different order.

daniel!

So op

U R AWESOME

I'll keep that in mind

Glad you liked it!!

Happy to do it!

Woo! Hope you enjoy!