100 Replies to “TensorFlow and Deep Learning without a PhD, Part 1 (Google Cloud Next ’17)

  1. I only started trying to learn bout this stuff (A.I., Machine Learning, Algebra, Statistics, etc) two days ago and I am loving it.

  2. Martin, you have a superb teaching style. A terrific way of walking-through how neural networks work. I was so drawn in and engaged, that I hardly noticed that 55 minutes had passed. Thank you for really bolstering my understanding and knowledge of neural nets and CNNs. Hope to see more from you!

  3. I don't understand the training approach with multiple layers. With one layer one knows the correct answer, but what do you do with multiple layers? How do you know what the gradient is, i.e., which way down is? I couldn't find that in the video. Thanks.

  4. DON'T LET THE FALSE TENSOR MOVEMENT TO BE TRASHED BECAUSE THEY CAN BE MEANING FULL FOR THE ANOTHER TENSOR MOVEMENT…! OR ANOTHER FUNCTION PUT ALL THE MOVEMENT IN A SYMMETRY SO ALL THE TENSOR BIT CAN BE REMEMBERED AND THOUGHTFUL FOR THE HIGHER PROGRAMMING DIMENSION AND MORE INTELLIGENT PROGRAMMING…! LIKE POLYFUNCTIONS OF SINGLE UNIT ACTING AND DECIDING IN A SAME TIME IN A VERSATILITY IN A SINGLE UNIT FOR THE AUTOMATIC DIVERSION AND CONVERSION, DUAL OR MULTIPLE FUNCTIONALITY..! IN ANOTHER WORD NEW LIFE WHICH CAN THINK AND ACT INDEPENDENTLY…! BECAUSE ALL OF THAT WE ARE GIVING THE INPUT THEORY FOR THE PROGRAM FOR TENSOR MOVEMENT BUT MAY BE LETTING ITSELF LEARN FROM THE MISTAKE CAN LET IT DECIDES AND MAY BE THINK OR CREATE IT SELF…! I MEAN THE PROGRAM SHOULD ALSO CAPABLE OF CREATING FROM THE MISTAKE BUT NOT ONLY TO BE TRACKED IN A RIGHT PATH..!

  5. Hi Martin, Thanks it's really help ful. Can you please share the IDE that you are using and how to open tensorflow dashboard

  6. Thank you for posting these conference videos. This was incredibly helpful and I wish to attend these conferences next time when I get the opportunity.

  7. Why not make the picture big and the teacher small instead- its so hard to follow what he says, when he explains something without seeing the picture.

  8. You could chain a side neural network based on the first learning sequence to train the dropout for the second network. Having a random dropout seems absurd since you already have tensor information extracted from your data set.

  9. MNIST is a bit cliche, but I really love this lecture!
    It's concise and visually clear.

    Highly recommended.

    Thanks for putting this up, Google.

  10. Fantastic lecture Martin. Makes everything so clear. One of the best tutorials I've seen on any subject.

  11. but the RELU is not differentiable so how do you compute the derivates for computing the gradients?

  12. Interesting demonstration of the simplicity of the Tensor Flow. However, the real world data in not necessarily the correct one to ensure accuracy of fit. What if that real-world data is incorrect/false due to such factors as the human cognitive dissonances, uncontrolled variables and faulty classifications (human errors) to begin with? In this view a theoretical mathematical data set could be more appropriate to ensure purity and the right fit.

  13. Hi Martin,
    Could you explain the convolutional neural network again in your example?
    You choose one weight and stride it across the whole image. Am I right?
    What is the value of this weight? I read in the other materials and they said that we will use a small matrix and use dot product to find out the convolutional matrix -> use ReLU -> use Max pooling for the next layer,…
    Which one is correct here?
    Thank you so much

  14. I don't know what everyone is raving about here.
    The presentation is far from being clear. Way too jumpy. Some concepts are not properly introduced and have to be deduced.
    – That woodoo with made up issues of adding matrices with different dimensions is just that – made up issue! All because speaker decided to jump to matrix multiplication.
    – Also, what's the point of scaling if you already have a bias for each neuron which after exponentiation acts like scaling?
    – The words like "this" should not be allowed during the presentation as it is often not clear what "this" is.
    – At 9:12 "network will output some probability" – probability?! This concept wrt network was never introduced.
    – 10:21: what's the point of exponentiating something just to take log later?!
    – 10:21: with that definition the network that outputs all 1s is the one that minimize entropy.

  15. Martin is an excellent teacher, but this is the 3rd or 4th time I'm seeing the same presentation given by him. We want the next level Martin

  16. It's actually not that well explained, unless you already have experience with neural networks. Here are some things that should be improved:
    – explain each variable and matrix variables with more details, especially on the second screen of python
    – take a bit of time to explain various runs. for example what's the difference in parameters between sess.run(train_step, feed_dict=train_data) and sess.run([accuracy, cross_entropy], feed=train_data), even if it's just for the display purposes. why are the different parameter names used (feed_dict vs feed), etc.
    – naming of the variables should be better. X, Y, Y_ is not very intuitive.

  17. Nice intro to Tensorflow! I found the run through of a single problem helpful.
    A bit of a nit to pick, though — shouldn't that 99% accuracy have been tested on a final final test set that had never been seen — how many informal model tweaking iterations occurred after peeking at the accuracy on the test set? Perhaps the ending model would not do so well on a truly novel test set. Not really important here, except that we should never forget that the whole point is to generalize to unseen data, which may be drawn from a different distribution. And it is a pain in the butt to not peek.

  18. More "democratization" of ANNs for the masses (thank you, NVIDIA, thank you google). Next thing you know, we'll liken them to commodities like toilet paper.

  19. Fantastic presentation Martin! Just one question: where can I find the training and test images that you used in your tutorial?

  20. Thank you Martin. One of the best tutorials on TensorFlow! Btw did you use Tensorboard for the realtime visualization?

  21. This video was a great way for me to get up to date with my newfound machine learning skills after taking Andrew Ng's online course. It just amazes me how radical some of this stuff is since 2011!

  22. 22:10– In reality if you want to reach the bottom of the mountains very quickly you should take long steps. ๐Ÿ˜€

  23. The next video in the series in online: https://youtu.be/vaL1I2BD_xY "Tensorflow, deep learning and modern convolutional neural nets". We build from scratch a neural network that can spot airplanes in aerial imagery and also cover recent (TF 1.4) Tensorflow high-level APIs like tf.layers, tf.estimator and the Dataset API. Developers that already know some basics (relu, softmax, dropout, …) I recommend you start there to see how a real deep model is built using the latest best practices for convnet design.

  24. Recommended Top Data Analytics and Deep learning courses:
    this awesome post from website Kupons Hub helps you to learn Deep Learning quickly
    Machine Learning, Data Science, Deep Learning, Artificial Intelligence A-Z Courses
    http://kuponshub.com/machine-learning-data-science-deep-learning-artificial-intelligence-a-z-courses/

  25. This video alone was soooo useful. Here is what I was able to do https://github.com/djaney/ml-studies/blob/master/06_conv.py

  26. When I changed batch size from 100 to 50 the program does not work. But the program worked fine for batch size > 100. Weird behavior.

  27. Best explanation , really easy of understand !!! Big thanks !!! can someone tell me the tool name he is using? those graphs of the error and the converge I dont see it those in the tensorflow I've installed

  28. metrixย  666x666x666์˜ 3D space๋กœ ์ด์šฐ์–ด์ง„ ๊ฒƒ์ด ํ๋น…..์ด๋ผ๊ณ  ๋‚œ ์ƒ๊ฐ์„ ํ•ฉ๋‹ˆ๋‹ค.

  29. ์—ฌ๊ธฐ์— ์‚ฌ๋žŒ๊ณผ ์ƒ๋ช…์€ ์ธ์ž๋กœ์„œ ๋ณ€์ˆ˜๋กœ ์ž‘์šฉ์„ ํ•˜๊ณ  ์ด๊ฒƒ์ด ์‹ ๊ฒฝํšŒ๋กœ๋ง์„ ๊ตฌ์ถ•ํ•˜๋Š” ํ•˜๋‚˜์˜ ๋ณธ๋ณด๊ธฐ๊ฐ€ ๋˜๋„๋ก ๋…ธ๋ ฅ์„ ํ•ฉ๋‹ˆ๋‹ค.

  30. ๋‰ด๋Ÿฐ์€ NEW RUN๋‹ฌ๋ฆฝ๋‹ˆ๋‹ค.ย  ๊ทธ๋ฆฌ๊ณ  ์ธ๊ฐ„์˜ ์—ญ์‚ฌ๋Š” ๋‹ฌ๋ฆฌ๊ธฐ ์—์„œ ์‹œ์ž‘์„ ํ•˜์˜€๊ณ  ์ด๊ฒƒ์€ ์ถ”์ ์˜ ์—ญ์‚ฌ์„ ๊ฐ€์ ธ์˜จ ๊ฒƒ์ด ํ˜„์žฌ์˜ ์‚ฌํšŒ๋ผ ๋‚˜๋Š” ๋ด…๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด ๋‰ด๋Ÿฐ์˜ ์ƒ์„ฑ๋ชฉ์ ์ด๊ณ  ์ด๊ฒƒ์ด ์ด๋ฃจ์–ด์ง€๋„๋ก ์ƒ๋ช…์ฒด์˜ ์‹ ๊ฒฝํšŒ๋กœ๋ง์ด ๊ตฌ์ถ•์ด ๋œ ๊ฒƒ์ด๋ผ ๋ด…๋‹ˆ๋‹ค. ์–‘์œก๊ฐ•์‹์ด๋‚˜ ์‚ฌํšŒ์ƒํ™œ์„ ํ•˜๋Š” ๊ฐœ๋ฏธ์˜ ์Šต์„ฑ ์ด๋‚˜ ํ˜น์€ ๊ฑฐ๋ถ์ด ์˜ ์žฅ์ˆ˜ ๋น„๊ฒฐ๋“ฑ ์ž…๋‹ˆ๋‹ค… ์‹ ๊ฒฝ์ด ๋‘” ํ•˜๋ฉด ์˜ค๋ž˜ ์‚ด๊ณ  ์‹ ๊ฒฝ์ด๋‚˜ ๋ฐ˜์‚ฌ์‹ ๊ฒฝ์ด ๋น ๋ฅด๋ฉด ๊ทธ๋งŒํผ ๋” ์šฑ๋” ๋นจ๋ฆฌ ๋…ธํ™”ํ•ด ๊ฐ€๋Š” ์Šต์„ฑ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค๊ณ  ๋ด…๋‹ˆ๋‹ค. ์ด๊ฒƒ์ด NEW RUN์˜ LIFE CYCLE๋ผ๊ณ  ๋ด…๋‹ˆ๋‹ค.

  31. About the image in 45:33, shouldn't the first convolutional layer have dimensions of 24 x 24 x 4? If each patch is 5 x 5, you can scan this patch across 23 possible places in the x direction, and 23 places in the y direction, correct? Or does the padding='same' make it so that the first patch's position is with 4×5 pixels on off to the left of the image ('looking at padding'), and with 1×5 pixels on the beginning of the actual image?

  32. Very amazing teaching! my grand-mother also could becomme Data Scientist. Great. Thanks a lot and I hope to hear you more and more. Woulb be the same thing if I want to do a simple regression wit mixed data?

  33. This is the most helpful tutorial that I have ever seen. It combines the theory and the practice together. The explanation is also very clear.

  34. For some reason Martin's scripts run a lot faster than my replicated jupyter notebook code. Can anyone offer some insight?

  35. I do not understand why now this is being taught when it's been know for 2 decades. This tutorial has no current application and nowhere to go because were actually very much past this.

  36. I've been casually watching machine learning tutorials for over a year and this is by FAR the clearest explanation of how a Convolutional Neural Network works, out of around two dozen that I've seen.

  37. What I would never have expected it's that he got >98% accuracy without making any "shape" correlation (that's almost magic to my eyes). CNN definitely important but maybe plays a bigger role with more difficult datasets.

  38. you might not need a phd but a high school certificate in maths and 3 years working in programming is certainly not enough

  39. 45:05 – Could someone pls explain W1[5,5,1,4] ? … i dont understand whats a patch of 5,5 and applying 4 of those to the images.

  40. We don't need to have PhD just for utilizing Tensorflow/DL, however that doesn't mean we can learn them within 1 hours without having any pre-knowledge or reading a text before. I'm also an novice not having any real tensorflow coding myself, but have read a related text, still am confusing at many parts. So I visited yt and watched this. I'm not sure it's perfect or not. But this video was very very helpful to me. Thanks Mr. scarf. ๐Ÿ™‚

  41. Very nice talk conveying the big picture and the general intuition. I do not think anyone can convey and address everything from the big picture to the nitty gritty in a 1 hr talk, given that I think this was a excellent introduction. Nice work Martin !

  42. This is a great presentation. Thanks for sharing. How can I use this with my own color images instead of using minst data set ? Can I create my own color mnist data set (how) ?

Leave a Reply

Your email address will not be published. Required fields are marked *