@@nathanwycoff4627 matrices and linear algebra are really useful for math and engineering less so for general programming. Different languages focusing on different usability concerns quite interesting.
I don't like it. I wish people stopped being overly-lazy with Numpy and just wrote their own libraries so they'd understand what they are actually doing. No, scratch that, if they can't accomplish the same thing using only Assembly, they're a total noob, should put down their keyboard, and get an MBA instead...
I watched this video when I was studying in grade 11. Had no clue what he was talking about but I tried to understand as much as possible. Now I watch it again as a university student, it is so satisfying to understand everything now.
It happens to me several time. Sometime you just stumble on a knowledge and can't understand a single thing about it then suddenly 1 or 2 years later you completely understand it without any try.
Took a Machine Learning course in university and this is what we did the whole semester in Matlab. Tensorflow was introduced right at the end for the final project.
If you make more deep learning videos with numpy and math(without any framework) just like in this video, it would be great for begginers to learn basics!!! Do you think to keep continue??
Hey guys, a reply would be highly appreciated. I want to plot the cost vs the number of iterations but I am not able to figure which parameter to plot ? I am a beginner and I would really appreciate the help. Thank you
Here's a course you'll need. Face Mask Detection Using Deep Learning and Neural Networks. It's paid but it's worth it. khadymschool.thinkific.com/courses/data-science-hands-on-covid-19-face-mask-detection-cnn-open-cv
@@anishojha1020 you're probably not a beginner anymore so I hope you found your answer! Unfortunately, youtube comment section isn't a forum and a lot of people disable notifications(including me) so an actual forum although people are sometimes really rude and condescending, is your best bet for future questions.
Hey guys, a reply would be highly appreciated. I want to plot the cost vs the number of iterations but I am not able to figure which parameter to plot ? I am a beginner and I would really appreciate the help. Thank you
Just your intro alone in your motivations was so capturing. You laid out everything so clearly, including creating those row and column matrices in the early steps. Thank you.
I'm so glad you actually went in depth with the math explanation. So often people will just explain surface layer and then "alright lets jump into the code".
I've never heard any of this explained before. After watching this once, I understand the mathematics behind neural networks and why the functions are used. Great job with the explanation here. Many thanks.
Most of the videos are titled “how to create a blabla” when they’re actually teaching how to use… so I really appreciate your video! This really contributes to knowledge 🥰
This was interesting, it certainly made neural networks far more approachable to me as someone who's never needed to/been inclined to try making one, but encounters them frequently by being involved in STEM. Your explanations coupled with my familiarity with numpy as opposed to dedicated libraries for neural networks really helped - thanks!
This was really neat. The math explanation was frustrating the first time around but really made sense after working through the code. Thanks for sharing.
Most tutorials I watch online about ML, you can just tell that the instructor doens't know whats happening. They've just memorized libraries and tensorflow syntax, and I don't want that to be me! This is exactly what i've been looking for! THANK YOU!!!
This was a really good video. I’ve never build a neural network but it was interesting seeing how the fundamentals add up to build something a little more complexed.
What an awesome video! Thank you for sharing this insightful walkthrough, it was really helpful in getting a better understanding of how neural nets works. Thank you!
Here's a course you'll need. Face Mask Detection Using Deep Learning and Neural Networks. It's paid but it's worth it. khadymschool.thinkific.com/courses/data-science-hands-on-covid-19-face-mask-detection-cnn-open-cv
Maaan, I am so happy you made this video. I was looking for somebody to train the Neural Network from scratch. I will go through it several times to get into the subject. Your English is excellent! Many, many thanks!
Just learned basics around the neural networks and saw this video. So satisfied to all the math formulas are laid out clearly in numpy and real-world coding and training neural network with back propagation. It really helps beginners like me. Thank you so much!
this type of learning is honestly the best, i implemented k means clustering by myself in c (pretty easy stuff but still) , and i can never forget it now, makes me happy that i can do stuff too
When I was in high-school algebra I programmed an algebra calculator to do my homework for me, and for some reason I never actually needed it. Programming something really is a great way of learning it, even if it does take significantly longer than just some p-sets or flashcards.
I remember when I tried to implement a decision tree on paper !! With a very small data dimensions (maybe 5x6 dim? Can't remember). I spent all the night doing the math but after 5-6 hours I realized I made a mistake in an iteration 😂😂 that's when I realized that we're lucky to have computers to help do it because a human mind can't build completely without doing mistakes in the process (can't stay focus for long time)... I also remember when I implemented a PCA from scratch on excel ( still have the Excel 😂)...😮
This is pure gold, MSc in Data Science and Artificial Intelligence, no professor ever gave me the answer to "what is the code inside the libraries we use", until I found you. Thank you
I don't want to sound too catchy and annoying but the NN's in Tensorflow and PyTorch are not actually implemented like this. They don't store functions to compute gradients for every single option rather they use AutoGradient which does all backpropogation job. I would highly recommend to watch Andrej Karpathy's tutorial on micrograd (mini AutoGradient which you will implement)
I got a master in physics and statistics but I do know how to code a lot of "machine learning" techniques from scratch. Yet human resources look at my degree and think I am incapable, so they rather hire master in AI. I can also code CFD, SPH and FEA from scratch but HR say I am dumber than engineer who just uses third party software (ansys).
@@michaelpieters1844 welcome to recruitment in 2024... you need to feed the recruiters what they want to hear, so that you can then get to the guy who you actually want to talk to about your stuff.
Really excellent breakdown of a Neural Network, especially the math explanation in the beginning. I also want to say how much I appreciate you leaving in your first attempt at coding it and the mistakes you made. Coding is hard, and spending an hour debugging your code just because of one little number is so real. Great video
Another thing that would be helpful for those of us that want to copy what you did and experiment with it is to have all the code together instead of separated as it is using Kaggle - this way you can put in some comments with the code explaining the different features. Again, very good video.
This is great. Built a backprop in C thirty years ago to solve the same problem. Just for a goof. It worked well before I finished debugging. These things are awesome and now I want to take another look. Thanks for posting this.
Super cool! Would also recommend the series from The Coding Train about creating a neural network from scratch, going a little more into the details of math and what is a perceptron and so.
You should continue making video similar to this maybe something a training course for machine learning and reinforcement learning AI. You have a real talent for explaining it in the best way possible then from what most videos I’d watched. 👍
Great video! I did the same thing in python about a year ago, but I didn’t like relying on numpy so much. Your video gave me the motivation to write both a matrix manipulator and neural network from scratch in Java
What an impressive speed run! Just nitpicking: 15:45 `rand` is for a uniform dist U(0,1) and `randn` is for the standard normal distribution N(0,1), therefore unbounded, not U(-0.5, 0.5)
I had many Machine Learning seminars in University and saw a lot of videos online on this topic. This is definitelly one of the best i ever saw. All relevant information in such short time, explained with such a high didactic quality. Wish i've had such docents at University. You should go teaching in MIT.
In case any beginners to ML came here wondering why they are really confused, this video isn't really for beginners and he doesn't really explain that. Its "from scratch" in the sense of not using any prebuilt models in the code. Its a good explanation for people who are already familiar with neural networks, prebuilt layers, loss functions, etc. not for people starting their understanding "from scratch."
actually im new to ML, (2-3 months in) and this helped me understand a lot, i am implementing it on my own now, without even using numpy so i can code out stuff like transpose on my own and learn more. Random is tricky tho lol
Hi Samson! I'm a developer and trying to learn the basics of ML. Much of the beginner stuff I see is using pre-trained models and frameworks which might be convenient to get things going. However, for me this is something completely new and I really what to understand what happens behind the scenes. Thank you for posting this! /Kevin from Sweden
Samson, this was such a great walk through. Just wanted to say that if you ever made other videos recreating machine learning models from scratch, I'd 100% watch them. In any case, hope all is good and thanks for this great content :)
Hi! I did a recreation of your code with more hidden layers and noticed what I think is a bug in the db calculation. Changing it to db = 1 / m * np.sum(dZ, axis=1).reshape(-1, 1) was able to get me better results. I think the old db = 1 / m * np.sum(dZ) sums the entire dZ to one float. Very good video though!
noticed the same thing. The way it was implemented here returns db to a float and thus b will always be "similar" to the random initialization, only shifted up/down by a constant.
Hey, I know you posted this a while ago, but I noticed the same thing and saw your comment. I am still not sure how to solve this, dZ is still a 1D array (1 by 10) so in your solution, what does axis=1 do? won't .sum*() just turn the 1D array into a scalar regardless, and then you are back with the same problem of updating all your biases the same way?
Numpy requires some strange things when you have only 1 dimension: Verfied that without this change the final biases weights aren't being updated. With it, training works better. Didn't verify the details of David's solution, just that it was needed, and that it seemed to work. def backward_prop(Z1, A1, Z2, A2, W1, W2, X, Y): one_hot_Y = one_hot(Y) dZ2 = A2 - one_hot_Y dW2 = 1 / m * dZ2.dot(A1.T) db2 = 1 / m * np.sum(dZ2, axis=1).reshape(-1, 1) dZ1 = W2.T.dot(dZ2) * ReLU_deriv(Z1) dW1 = 1 / m * dZ1.dot(X.T) db1 = 1 / m * np.sum(dZ1, axis=1).reshape(-1, 1) return dW1, db1, dW2, db2
I see the same. Also, either this is old enough that something has changed in Python or numpy, or he hasn’t included other things as well. Using his code line for line and the same data set, I get a divide by zero error on the softmax function.
I’m always too intimidated to try some of these things. But seeing your process makes it really seem feasible. Need to brush up on my linear algebra again tho 😆
Amazing video for beginners to gain an insight in how neural networks work. You just have to have programmed a simple neural net from scratch once to have a good basic understanding.
An excellent nice video with abundant mathematical insight. It may be worth to note that instead of partial derivatives one can work with derivatives as the linear transformations they really are, and also looking at the networks in a more structured manner thus making clear how the basic ideas of BPP apply to much more general cases. Several steps are involved. 1.- More general processing units. Any continuously differentiable function of inputs and weights will do; these inputs and weights can belong, beyond Euclidean spaces, to any Hilbert space. Derivatives are linear transformations and the derivative of a neural processing unit is the direct sum of its partial derivatives with respect to the inputs and with respect to the weights; this is a linear transformation expressed as the sum of its restrictions to a pair of complementary subspaces. 2.- More general layers (any number of units). Single unit layers can create a bottleneck that renders the whole network useless. Putting together several units in a unique layer is equivalent to taking their product (as functions, in the sense of set theory). The layers are functions of the of inputs and of the weights of the totality of the units. The derivative of a layer is then the product of the derivatives of the units; this is a product of linear transformations. 3.- Networks with any number of layers. A network is the composition (as functions, and in the set theoretical sense) of its layers. By the chain rule the derivative of the network is the composition of the derivatives of the layers; this is a composition of linear transformations. 4.- Quadratic error of a function. ... --- Since this comment is becoming too long I will stop here. The point is that a very general viewpoint clarifies many aspects of BPP. If you are interested in the full story and have some familiarity with Hilbert spaces please google for papers dealing with backpropagation in Hilbert spaces. A related article with matrix formulas for backpropagation on semilinear networks is also available. For a glimpse into a completely new deep learning algorithm which is orders of magnitude more efficient, controllable and faster than BPP search in this platform for a video about deep learning without backpropagation; in its description there are links to a demo software. The new algorithm is based on the following very general and powerful result (google it): Polyhedrons and perceptrons are functionally equivalent. For the elementary conceptual basis of NNs see the article Neural Network Formalism. Daniel Crespin
Just 1 minute in the video and I can easily tell that you're gonna own a multi-billion company within a few years. You've got the IQ, the voice, the clarity, the confidence, and the right personality. Best of luck Mr. Zhang
The yt algorithm only recommends me this now, 1 year after i've encountered a similar discontent with neural network tutorials. Still very interresting to see how someone else does it. I did give myself a bit of help by using a library called Eigen for the matrixes calculations. Very well done nice video
Haven’t finished video yet, but this looks like the missing piece of my experience learning about neural networks at a high level…I probably lacked the linear algebra skills I have now though. Whoa! This could be incredibly exciting! I can’t wait!
bro, I am watching NN course from online platform for 1 months, but still difficult to get grasp on it. But you made me understand it in just 30 mins. many thanks
i have no idea what your were really saying but at the same time i do because you explained how the math is used and implemented for the code. thank you !
There is one thing I do not understand. Because the derivation and chain rule stuff, shouldn't the derivative of the softmax activation function also be included somewhere?
This is a great way to teach ANN - congrats. However, I would like to suggest you to not worry too much about the time to finish the implementation. Double-checking all steps will avoid coding errors.
This is great. Librarys are good for using, but not so good for learning and really understanding. So I really appreciate this tutorial! You made my day!
Here's a course you'll need. Face Mask Detection Using Deep Learning and Neural Networks. It's paid but it's worth it. khadymschool.thinkific.com/courses/data-science-hands-on-covid-19-face-mask-detection-cnn-open-cv
A great introduction to neural networks is Parallel Distributed Programming by Rumelhart and McLelland from about 1986. They do something similar and give a lot of additional background.
Hi, i found this video very helpful for beginners. Could you please tell how you came up the equations of dz,dw and db? That would be really helpful as well
Perhaps I overcomplicated matters compared to your approach when I did this a couple of years ago, but like you, I wanted to program it "from scratch". My language of choice: java. I actually simulated "neurons" which were a class that stored its activation data value, and its connections to the next layer, so that it "looked" like a K_m,n graph, and the connection was an array which stored the biases along each "synapse" so to speak. Then when the hidden layers activated, I had each neuron simply sum the outputs from each synapse connecting to it from the previous layer, which was just the product of its activation value and its bias, then sigmoided this to get its own activation value. Note that while each neuron's activation was only in (-1,1), I let the biases be free parameters. When I programmed the backprop algo, I did the gradient descent the same as you, but effectively set that alpha parameter to one. It didn't occur to me to mess with that. Starting the network out with random parameters, then training it on randomly chosen sets of 10,000 images five or six times seemed to work pretty well. I saw 93% accuracy on the test data. And just for fun, I put the network on a discord bot so my friends could feed it images of the same size and see its guess. Two interesting results came out. The network fails on inverted colors: i.e., drawing white on black using MS paint or something wouldn't get reliable predictions. Secondly, using MS paint to give it new data did work, but at a much lower rate. Our best guess for why this happened was due to the sharpness of the lines between the number and backgrounds.
You can actually use momentum for gradient descent. The result is slightly better (I tried on your nn and it gets 91% accuracy) // I'm a beginner at ML so your video taught me a lot. Keep up your great work you're doing man. It's really cool.
Amazing stuff! Just wondering what value does the coding timer add to the video? I mean instead of correcting your mistakes with overlapping text you could have taken a little bit of time to review your code instead of rushing it through. But again, amazing content!
Very good video and explanation! Thanks 😊. I just would have liked it if you had explained the backprop a little more in depth. Like how the derivatives are calculated on each layer (chain rule etc.) But other than that one of the best nn videos
Here's a course you'll need. Face Mask Detection Using Deep Learning and Neural Networks. It's paid but it's worth it. khadymschool.thinkific.com/courses/data-science-hands-on-covid-19-face-mask-detection-cnn-open-cv
Here's a course you'll need. Face Mask Detection Using Deep Learning and Neural Networks. It's paid but it's worth it. khadymschool.thinkific.com/courses/data-science-hands-on-covid-19-face-mask-detection-cnn-open-cv
@23:30 you also see two errors, there is no axis argument for the np.sum(), the lines should be db2 = 1 / m * np.sum(dZ2) ... and ... db1 = 1 / m * np.sum(dZ1)
It's a MLP, you easily computed the backpropagation step in closed form, but I wonder how those famous frameworks can compute any network's partial-derivatives tensors automatically
usually the partial derivatives in backpropagation are of functions specifically chosen to be convex and have nothing to do with the problem you are working on, but are just ones that work nicely for ML algos
Hey, I found a flaw in your code and would be great if you answer it......The updation that you are doing for the bias' is not all needed as per your code because all the bias are changed by same factor hence it's still random( you have used a scalar to update the bias instead of a column vector)......I found the correct solution to it but getting an error. you should add the axis=1 in the sum function.