13:19:17 From Jonathan : the what curve? 13:19:21 From Andy Anker : Thx great answer! 13:22:59 From Haider Fadil Ali Hussein Al-Saadi : so its false positive on one axis and true positives on the other? 13:23:19 From Adriano : correct 13:23:40 From Jonathan : yes 13:27:54 From Sofus Kjærsgaard Stray : Why is it 50-dimensional? 13:28:08 From Jonathan : 50 weights i think ? 13:28:27 From Jonathan : 50 parameters to be tuned 13:28:32 From Sofus Kjærsgaard Stray : I see 13:28:51 From Simon Ulrik Hilbard : why is it not 20 + 5 dimensional ? 13:29:27 From Jonathan : yeah.. i don’t know 13:29:32 From Jonathan : good question 13:29:34 From Aske R. : because of the four different a for the s(x) I think 13:29:48 From Haider Fadil Ali Hussein Al-Saadi : yeah it should be 25 13:31:05 From Emy Alerskans : because there are two parameters for each activation function to be optimized 13:34:58 From Simon Ulrik Hilbard : which pararmeters i can only see a_i or w_i in the equations from the earlier slide ? 13:35:19 From Sofus Kjærsgaard Stray : That'd mean each connection has a_i and w_i so that'd be 50 total right? 13:35:41 From Haider Fadil Ali Hussein Al-Saadi : weights and biases? 13:37:12 From Emy Alerskans : so the parameters for each line is a and x_0 in your slides, or what? 13:41:07 From Haider Fadil Ali Hussein Al-Saadi : I dont understand how giving it both types is superior to both? 13:41:17 From Haider Fadil Ali Hussein Al-Saadi : either* 13:42:22 From Haider Fadil Ali Hussein Al-Saadi : the question is why is the 3rd bar highest basically 13:43:43 From Sofus Kjærsgaard Stray : But why is all variables better than JUST human-assisted? 13:43:58 From Haider Fadil Ali Hussein Al-Saadi : exactly 13:44:47 From Sofus Kjærsgaard Stray : yes 13:46:29 From Emil Schou Martiny : Also almost by definition if we don't get all the information in the human assisted variables. I we could calculate all relevant things, then why are we using a ML algorithm 13:47:07 From Sofus Kjærsgaard Stray : I guess the variables in question still need to be estimated 13:47:15 From Sofus Kjærsgaard Stray : So you know it's invariant mass but not what the invariant mass should be 13:49:30 From Moust : you said that you only can get linear combinations of the two input variables but isn't the idea of using a non linear activation function that the combination isn't linear 13:50:33 From Haider Fadil Ali Hussein Al-Saadi : not necessarily sofus. If we see it as the example he gave with ratios, the raw variables were for example X1 and X2. You can then define the ratio of these two as X3, and now you can then feed X3 to a neural network. However, feeding the network only X3 does lose quite a bit of information from X1 and X2, since the same ratio can be achieved with many different values from X1 and X2, so the optimal result would come from feeding it X1,X2 and X3. 13:51:30 From Sofus Kjærsgaard Stray : Okay, that makes sense 13:56:43 From Emil Schou Martiny : There is also somethin going around with optimizing structures with fluid dynamics 13:59:10 From Jonathan : will we get more into how to tune hyper parameters later? 13:59:34 From Jonathan : :D 14:00:09 From Simon Ulrik Hilbard : will we get example code for NN today ? 14:00:45 From Simone Vejlgaard : Will the disturbances introduce more exercises? 14:01:01 From Yane García : I have heard that neuronal networks are slow some cases? does it depend of sth else? Or the nature of the dataset? 14:01:04 From Emil Schou Martiny : I made a neural network from scratcg in another course, it took 2 weeks and was only roughly 2-3 orders of magnitude slower than sklearn :h 14:02:45 From Haider Fadil Ali Hussein Al-Saadi : I made one as well, god it was a pain in the ass :D 14:05:10 From Yane García : thanks 14:05:45 From Simone Vejlgaard : Will the disturbances introduce more exercises? 14:05:49 From Andy Anker : I cannot find the sklearn transformation package, could you link to it here? 14:06:50 From Carl-Johannes Johnsen To Troels Christian Petersen(privately) : Det er kun host der kan hente listen af deltagere, men du kan dog også hente det for gamle sessions: https://support.zoom.us/hc/en-us/articles/216378603-Generating-Meeting-Reports-for-Registration-and-Polling 14:06:57 From Troels Christian Petersen : https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.quantile_transform.html 14:07:57 From Simone Vejlgaard : Will there be any additions to the exercise today? 15:10:37 From Haider Fadil Ali Hussein Al-Saadi : can you answer the question i posed a while ago? 15:13:36 From Haider Fadil Ali Hussein Al-Saadi : what do you mean by hits? 15:14:59 From Haider Fadil Ali Hussein Al-Saadi : yes 15:15:02 From Haider Fadil Ali Hussein Al-Saadi : thanks 15:17:30 From Carl-Johannes Johnsen To Troels Christian Petersen(privately) : Mht polls, der var det fordi at du sagde du sad og ventede på at 3 svarede, og så var det jeg tænkte om det var mig, Adriano og Zoe! Jeg har ikke lige set hvordan host enden af en poll ser ud :) 15:30:19 From Haider Fadil Ali Hussein Al-Saadi : so the week 1 schedule says the exercise is just to get a neural net working and compare with tree algorithm. Is there any more to be done after that is over with? 15:31:19 From Troels Christian Petersen : No - that is it. If you feel that you’ve played around with NNs enough to have a basic idea, and can compare the NN to the tree-based results, then that is it. 15:32:11 From Haider Fadil Ali Hussein Al-Saadi : Alright. You wrote to me earlier that you wanted my code for the tree algorithm to put up on absalon, do you want it on mail? 15:32:45 From Troels Christian Petersen : Yes, that would be fine. I already got one other example, and my idea is simply to put them on the Week1 webpage for all to see and get inspiration from. 15:55:12 From Jonathan : does anyone have an example of how to solve this exercise with pytorch? in that case I would love to see it :) 15:56:35 From Runi : Anyone got below 10% using sklearn and MLPClassifier? Trying to change hyperparameters, but no luck so far to get it better 15:57:13 From Emil Schou Martiny : slightly below. like 9.8 i think 15:58:02 From Jonathan : I got 9.9% with sklearn, technically thats below but… not much better :P 15:58:44 From Troels Christian Petersen : @Jonathan: No, but I take it that you’ve tried to take that path. If you think that you’re close to getting something, don’t hesitate to send it to us for a look 15:59:00 From Runi : What did you change to go from 10% to 9.8/9.9% ? :P 15:59:21 From Troels Christian Petersen : General question: How many of you are transforming the input variables before feeding it to the training? 16:00:34 From Yane García : I am not transforming 16:00:47 From Yane García : ok 16:01:38 From Runi : My error rate goes to 11% if I don't transform, so transforming seems better here 16:06:51 From Aske R. : https://www.datahubbs.com/deep-learning-101-first-neural-network-with-pytorch/ 16:06:58 From Jonathan : thanks :D 16:08:06 From Aske R. : for now i am not moving into the improving accuracy yet 16:08:16 From Aske R. : since that seems complicated 16:12:13 From Yane García : have a question regarding the model fitting. In SKlearn there is a function "score" for the model (X_test , y_test) I had 0.90...and the r2 score (y_test, y_pred)...0.48... This later one does not look good. 16:15:03 From Yane García : and wrong fracc 9.97% 16:15:56 From zoeansari : As far as I know about the r2_score in sklearn , the best value would be 0 16:16:24 From Yane García : as far as I know should be close to 1 r2 score 16:16:44 From Yane García : oh 16:17:09 From Troels Christian Petersen : https://scikit-learn.org/stable/modules/model_evaluation.html 16:17:30 From Yane García : Thanks Troel and Zoe 16:18:29 From Runi : Troels if I'm using the Classification report in sklearn, the weighted f1-score is the best result to report right? 16:18:49 From Runi : weighted avg* 16:20:49 From Troels Christian Petersen : Well, it is not that easy, as different scores focus on different things, which may or may not be what you really want… 16:21:43 From Runi : Well it corresponds to error rate right? 16:28:20 From Yane García : hhaah 16:29:21 From Simon Ulrik Hilbard : it was fun 16:38:08 From Carl-Johannes Johnsen : @Runi No, the error rate (1.0 - accuracy) and the weighted avg f1-score from the classification report is not the same: >>> test = [0, 1, 0, 1] >>> pred = [0, 1, 1, 1] >>> accuracy_score(test, pred) 0.75 >>> print(classification_report(test, pred)) precision recall f1-score support 0 1.00 0.50 0.67 2 1 0.67 1.00 0.80 2 accuracy 0.75 4 macro avg 0.83 0.75 0.73 4 weighted avg 0.83 0.75 0.73 4 16:38:52 From Yane García : if a data sets consists on samples with replicates (e.g. 3 replicates per sample) and they are all but identified wihin the raw data. Should I consider the raw data set as it is? or should I consider an standard deviation for each sample? does it have any relevance at this point when doing ML? 16:40:28 From Runi : @Carl-Johannes, yeah "accuracy" seems to be the inverse error-rate, so I'm using that now 16:41:19 From Runi : I'm fluctuating between 9.4-9.8% error rate. 16:43:58 From zoeansari : I am not sure about what you mean on using the standard deviation here Yane, could you explain it a little bit more? But in the case when we have same data in the training , we defiantly need to get rid of the copies, cause it affects an overtraining which might not even be observable in our checks after the training. 16:46:09 From Runi : Btw for those using sklearn I recommend setting verbose=True , as you can follow the iterations as they are calculated in the console 16:53:45 From Runi : got 9.16% :D