Anomaly detection with TensorFlow | Workshop 

Hands down my favorite ML teacher

The 45 mins workshop are really good. Please continue to make more of these workshops.

Fantastic video! You made neural networks very simple to understand.

This use of generative AI is just brilliant!

What a great learning experience! Thank you Laurence Moroney

Simply awesome!!! Thank you so much Laurence, it's always a pleasure to learn from you!!

He is my teacher in coursera. Awesome Teacher. He teach me with Andrew Ng

I love Generative modeling and your teaching style!

nice workshop love the explanation, and the materials given

Thanks. ECG is a time series..Can it be a simple Feed forward auto-encoder ? Should it not be a RNN Auto-encoder ?

Awesome content, very helpful for my research. More of this.

Great content and well designed example on using an autoencoder for anomaly detection.

Hi, first of all, thank you very much for the content. My question is if this would be possible with a simple classification model. If so, what are the pros and cons of each approach? Thanks in advance

Thank you for teaching this! I was confused for a long time...

Amazing, thanks for sharing your knowledge with us!

More of these .. would be great for decision tree as well

Awesome workshop! I really liked how you selected the threshold as mean + std_dev. But if we make the Gaussian assumption do you think that mean + 2*std_dev would be a safer value?

Excellent video and thanks a lot! Please allow me to make a question. Is the method shown in the video recommended to train and detect anomalies in financial transactions? Is the way of comparing the transaction against a previously calculated threshold recommended, or do you recommend any other I can use? I am new in ML, so if you know any example I can use would be incredibly appreciated.

Super useful content!

Thanks for the great content, Please if anyone could help in this, I have not understant why the python the dens layer started with 128 input and not 784?

I like your videos a lot! I have a question about the time series analysis. As I understood the shape of your network depends on the length of your data has. But what about the problem, when your single trials have different time duration, for example when one heartbeat is going longer than another one?

At 34mins 31s under the history code section, shouldn't the validation_data be using normal_test_data instead?

Thanks , this videos will be helped me!

Awesome! Why you do not have to add input layer?

Very valuable, many thanks...

تحياتي الخالصةthank you

Hi! Thanks for the great content! Is it necessary for the decoder to mirror the encoder? Why does it mirror the layers of the encoder?

Hello, could you please let me know what should be the value of dense layer like 32,16,8 if we have only 8 variables instead of 140 as ECG indicators??

@Laurence Moroney Very interesting, thank you!! But what if we are in un unsupervised scenario? In this case, i trained the autoencoder on the whole unlabeled dataset and made predictions on the same whole data, trying to spot anomalies. Given that the data are unlabeled and given that the percentage of anomalies should be less than about 1% of the entire dataset, would be correct to train the autencoder on the whole dataset, then predicting on the whole data and finally looking at the reconstruction errors in order to spot anomalies (points with the highest reconstruction error) ?

Very good content

Thank you Laurence. People would naturally worry about the 3 false negatives that happen to show low losses. Is there any remedy for that?

thanks coatch

perfect

Thank you very much

@Laurence Moroney where to see the ipython code here? could you please attach?

when i trained my model, I had a 96% accuracy but when i deployed it to an android application for what I needed, the prediction was totally different from what I got in Python, what could be the reason for this ? Any answer would be highly appreciated.

Which one works good for anomaly detection in time series whether autoencoder or Variential Autoencoder. Which one you would suggest.

Christopher Moriarty This is body copy and it's a little like this. This is body copy and it's a little like this.

so, data from high-dimensional solution space is reduced to its lower-dimensional solution space and then error is computed. Somehow, related to Proper Orthogonal Decomposition. Is it right?

There is an example for anomaly detection in images(like fashion MNIST dataset) ?

Is Multivariate Anomaly Detection possible with deep learning? For a dataset that includes different sensors readings?

Can anyone explain why we validate with both the normal and anomalous data?

Sir, would you like to suggest any book of tensorflow in which I could learn topics such as # Kalman Filter # BERT # LSTM and other advance topics

Thank you very much for your very useful video!

I have done the same thing on kaggle. using the ecg from physionet

Thanks ! Actually i am working on a project on "anomaly fault detection in power system"...and it really helps a lot...Can you suggest me something in line of my project(some links or articles will do)..?

Hello , Can you please explain how we are finding the number of layers and neurons in neural networks???

@TensorFlow could you please reply where we can look at the code?

Where we can get the code?

📺💬 He listening ... 📺💬 The decoding part is work very well on the extended type of coding. 🥺💬 You are correct it is a transformer and generator there are many references about them and it explains why it required encoding and decoding. 🥺💬 They are not required to have the same symmetric layers but they use functions from the generator with the objective, you can merge them into one model but translation will require for learning. By conclusion, objective and input are different and transformation information can be used in the target environment by the transformers part. 🥺💬The same as the cryptography class model and its concept transformer widely use in many objectives but itself can decode the input to the output they generated, the generators transformed input into messages. In your online class, there is word text generators that is the same. 📺💬 I use ECG 🥺💬 It is similar to the Fourier function transform we use it with repeating signals such as frequency but human brains or activities have an identity and there are many frequencies working at the same time as you test heart-working examines required to have opposite channels connected or all connected to investigate its working, compare to voices or natural signals. 📺💬It is count number of products, not Furrier 🥺💬 I explain about the transcoder ( transformer - encoder ) see the recursive function. I am not trying to talk about the name of the theory to make everything look too hard but when input as sequences pass though multiple layers of neurons it squeezes and expands when weaken connection is expanded and denses high connection. See the output graph or you can formula. 📺💬 ยุ้ยเชิญว่าอาจาร์ยขี้โม้ด้วยครับพอดีญาติกับซัน 🥺💬 ไม่มีใครว่าใครหลอกครับถ้าไม่เดือดร้อนดูเค้าทำกับผมก่อน แค่นั้นแหละผมก็ดูไปเรื่อยดีกว่าดูคนทะเลาะกันครับ ขอบคุณครับ สอนดีครับแต่เหมือนที่อาจาร์ยบอกบางข้อมันยากเกิน

This audio is bad.