08:19:52 Anon. Schenley: The links for hw1 writeup on the webpage seem down for me. Is it actually down or is my browser just messing up right now? 08:20:09 Anon. S. Highland: Do we get explanations for quiz answers? 08:20:12 Anon. Tech: We’ll check that and let you know. Sorry about that 08:20:15 Anon. Wightman: The link in piazza still works 08:21:19 Anon. Tech: @rtyr, Yes, correct answers and short explanations will be released. If you have more doubts about it, you can post them on Piazza or come to OH. 08:29:56 Anon. Star-Lord: If there is a random function (RNG), can an MLP model that too? 08:30:52 Anon. Gamora: someone has a beeping sound in the background, could they mute that? 08:31:19 Anon. Superman: should that threshold be 4 indtead of -4? 08:31:48 Anon. Capt. America: I have same doubt here 08:32:05 Anon. Star-Lord: ^Same. 08:33:09 Anon. Friendship: I think there's a mute all option for the host 08:33:54 Anon. Drax: @sdfsd I think the random function is a worst case to model for any MLP, but it can. 08:34:12 Anon. Star-Lord: Gotcha, thanks. 08:34:49 Anon. Hobart: integration 08:35:49 Anon. SpyKid1: is div() the abstract value of f(x) - g(x)? 08:37:05 Anon. Spiderman: ^I think you meant absolute value? Then looks like it. 08:37:14 Anon. Tech: @werwe Not literally, but yes. It is the extent of similarity between two functions 08:37:33 Anon. Tech: dissimilarity* 08:37:50 Anon. SpyKid1: Got it, thanks! 08:39:16 Anon. Myrtle: Do those samples need to satisfy Nyquist Sampling Theorem? 08:40:53 Anon. Bartlett: I don't think the target function will necessarily have a frequency for Nyquist to come in 08:40:56 Anon. Fury: I don't think it's necessary to sample that way 08:41:32 Anon. Drax: We don’t have to reconstruct the function exactly so nyquist sampling should not be reqd is what i am thinking 08:41:36 Anon. Friendship: Is it being sampled because we don't know every value of g(x)? In other words, is this supposed to be analogous to how the training set cannot represent the entire true set? 08:44:14 Anon. Friendship: No 08:45:06 Anon. Penn: What informs the weights and biases you start with or they can be chosen randomly? 08:45:07 Anon. Friendship: I see, thank you! 08:45:52 Anon. Fury: you can start randomly then with backprop you adjust ur weights 08:45:53 Anon. Murdoch: Weights and the biases are initialized using certain initialization functions 08:46:01 Anon. Tech: @asasd They can be random or they can be sampled out of a particular distribution. Depends upon the problem! 08:46:27 Anon. Murdoch: You will get to explore this in hw1p1 08:49:21 Anon. GreenArrow: Why do we want to make it pass the origin? 08:50:02 Anon. Wilkins: X is orthogonal to W 08:50:47 Anon. Penn: I think it’s the formula @uikuik. sum(WiXi + constant) is a hyperplane. If constant = 0 then it passes through origin 08:50:58 Anon. Wilkins: Cosine of the angle between them 08:51:40 Anon. Friendship: negative 08:51:42 Anon. Myrtle: negative 08:51:43 Anon. Liberty: neg 09:02:18 Anon. Drax: No 09:02:20 Anon. Bartlett: No 09:02:20 Anon. IronMan: no 09:02:22 Anon. Phillips: no 09:02:26 Anon. Centre: no 09:02:39 Anon. Centre: Redefine data 09:02:41 Anon. Wilkins: relabel 09:05:24 Anon. Vision: exponential 09:05:27 Anon. Penn: Combinatorial 09:05:28 Anon. Schenley: exponential? 09:05:31 Anon. Centre: exponential 09:05:34 Anon. BlackWidow: exp 09:05:37 Anon. Bellefield: exponential 09:14:10 Anon. Smithfield: still one but less error 09:15:20 Anon. Grandview: do we assume that the slope of the curve remains the same? 09:16:19 Anon. Tech: For the problem in the slides, yes 09:25:19 Anon. Drax: How do we get P(X) 09:26:06 Anon. Friendship: yes 09:26:49 Anon. Tech: @wqeqw P(x) will be the distribution function of x (training data) 09:27:41 Anon. Drax: @vaidehi but we only have g(x) right? The function which we have to approximate. 09:31:21 Anon. Drax: I think it was added to reach to the empirical divergence 09:32:01 Anon. Tech: g(x) is the function we are approximating over input space X 09:34:00 Anon. Drax: yeah, thanks vaidehi! Also, is the difference between true and empirical risk known as variance? 09:44:39 Anon. P.J. McArdle: @Bhiksha:You can open the hamburger menu on the top right and turn off ink to shape. Should prevent the annoying shapes from appearing. 09:49:09 Anon. Nebula: this particular method only works if the data is linearly separable right? 09:49:31 Anon. Drax: @cxxcv yes 09:49:33 Anon. Tech: Yes! 09:50:55 Anon. Drax: Is Bhiksha holding any office hours? 09:52:24 Anon. Tech: No, he is not! But you can ask your doubts on Piazza and he can answer:) Or come to the Office hours! 09:52:43 Anon. Drax: Sure, thanks @vaidehi 09:55:15 Anon. Batman: thank you 09:55:15 Anon. Beechwood: thank you professor! 09:55:16 Anon. P.J. McArdle: Thanks!