00:22:43 Reshmi Ghosh (TA): Please lower your hands when you are done. Pretty please !:P 00:28:56 Anon. Batchnorm: Hi Detective! 00:29:06 Anon. Batchnorm: Ignore this sorry 00:29:10 Reshmi Ghosh (TA): Huehuehue 00:29:17 Reshmi Ghosh (TA): Shoot:P 00:29:40 Reshmi Ghosh (TA): Sorry that was for a private message 00:37:47 Anon. 70mV: Why would it be the entire input rather than the input up to some point? 00:38:16 Anon. Kaggle: maybe in the case where its bidirectional im thinking 00:41:32 Anon. Gabor transforms: if at each point we receive all input sequence, what makes the outputs change? 00:41:48 Anon. Mac: a or an 00:42:11 Anon. Markov Chain: So is this bidirectional? 00:42:40 Anon. Sequence: not necessarily 00:42:50 Anon. Gabor transforms: I guess the inputs of outputs change 00:42:58 Anon. Fast RCNN: I think the inputs are fed into the network as a sequence and each timestep has an output. To get the most probable sequence you have to wait till the end 00:43:14 Anon. Andy: double letters? 00:45:25 Reshmi Ghosh (TA): There’s an inherent dependence of outputs generated from previous inputs. Imagine a sentence, you are trying to come up with the 4th word, it would be dependent on the what words came before that 00:46:22 Anon. Gabor transforms: that makes a lot of sense, thank you 00:46:52 Anon. Mac: so instead of the future directly influencing the output, we just interpret the result afterwards based on the entire input sequence? 00:48:46 Reshmi Ghosh (TA): If you see the diagram on the slide which was just shown we have a probability of what may come out for each time step, Does that make sense? 00:49:30 Anon. Kaggle: so if i'm understanding correctly for each time step t each node we'll calculate p(x) of all possible symbols then get the highest p(x) and combine with each time step until X n-1? I'm abit confused bcz based on us having our input sequentially how will u get the symbols of vocabulary of the words or symbols u haven't encountered yet in ur time step? 00:49:42 Anon. Directed Edge: Hmm what is the input for x0 for example? Is it the whole sequence still? 00:49:43 Reshmi Ghosh (TA): It is okay if you don’t; we have to get it right otherwise beam/greedy search won’t make any sense 00:53:22 Anon. Mac: so what we DON’T want is to just output the highest probability label sequentially at each time, because that’s like repeated hitting the keyboard suggestion on an iPhone and generating nonsense 00:54:32 Anon. Loss Function: when calculating prob distribution of a position, can we see the hidden states or prob distributions of subsequent positions? 00:56:37 Anon. Mac: I thought the probability distribution of a position is just the output of the network at that time step 00:57:19 Reshmi Ghosh (TA): @sammy et al.:we have a prob dist as output at each time when we are feeding the net an input. Imagine this you are trying to predict “I am a dog”, at each point when you feed an input there will be a prob dist of ‘am’ ‘a’ ‘dog’ or something else; and then you do a search on the distributions generated at all points 00:57:26 Reshmi Ghosh (TA): Does that make sense? 00:57:35 Anon. Directed Edge: Thank you 00:58:54 Anon. Kaggle: yeah it's making sense now 01:03:16 Anon. 70mV: So essentially we want Is the most probable path under the joint distributions rather than a naive greedy search? 01:06:06 Anon. Kaggle: would it make sense to add transition probabilities to this to make it output a good sequence? 01:06:26 Reshmi Ghosh (TA): What do you mean by transition prob? 01:07:37 Anon. Kaggle: like in an markov chain where u'd have odds of going from 'I' to "went' and from 'I' to 'is' would have less odds of happening 01:07:55 Anon. Kaggle: so for each transition btn words there is usually a P(X) that is associated with it 01:08:39 Anon. Kaggle: so if ur model say ' I went is market' and ' I went to the market' the second one would be more likely to be right 01:09:43 Anon. Mac: but for “I have a..” vs “I have an…” , is “a” or “an” more likely 01:13:41 Reshmi Ghosh (TA): We are already using prob distribution to find out the final sequence through the search process; I m not sure how you associate the transitional prob to each distribution for each output at a particular time; I am not asking this to Bhiksha rn as we are already behind; but feel free to post on Piazza 01:14:10 Anon. Kaggle: alright, thx 01:21:32 Anon. Residual: Train the model on all possible alignments? 01:21:38 Anon. Algorithm: train over average of all possible alignments? 01:22:59 Anon. Algorithm: do we need to know the order, or can we expand to an un ordered sequence? 01:28:06 Anon. Mac: just the last occurrence of each continuous segment? 01:30:34 Reshmi Ghosh (TA): @Jeff I lost it 01:30:41 Reshmi Ghosh (TA): Might post on Piazza? 01:31:09 Anon. Directed Edge: ^ I think he was just answering an question the prof was asking 01:31:19 Anon. Mac: @ Reshmi I was just answering a question professor posed, thanks 01:31:22 Reshmi Ghosh (TA): lol 01:31:31 Reshmi Ghosh (TA): Yay poll is coming! 01:31:51 Reshmi Ghosh (TA): damn 01:31:54 Anon. Mac: rip poll 01:31:55 Reshmi Ghosh (TA): My hopes were high 01:31:56 Anon. ASGD: lol 01:31:57 Anon. Directed Edge: lol 01:32:00 Anon. Gaussian: ahhh 01:32:02 Anon. Quadratic: xD 01:32:07 Anon. Algorithm: lol 01:32:09 Reshmi Ghosh (TA): I had it ready since the last 30 mins 01:32:10 Reshmi Ghosh (TA): :(( 01:39:22 Anon. Mac: because each output is independent 01:39:38 Anon. Mac: so we can multiply them right 01:40:18 Reshmi Ghosh (TA): Eh multiply what:O? 01:40:33 Anon. Mac: the probabilities at each node 01:42:19 Anon. Mac: it’s ok Im ok now lol 01:43:07 Reshmi Ghosh (TA): XD 01:44:22 Reshmi Ghosh (TA): If you are not okay; feel free to ask again 01:45:28 Reshmi Ghosh (TA): Strong hands everyone has got; keeping it raised for soo long:O 01:46:32 Reshmi Ghosh (TA): LOL; I am not even gonna ask to lower hands now 01:54:28 Anon. hello_world.py: rip poll 01:54:29 Reshmi Ghosh (TA): There is a poll again; and now I am gonna try! 01:54:59 Reshmi Ghosh (TA): My heart is breaking 01:55:10 Anon. Directed Edge: Lol is poll appearing today 01:55:34 Reshmi Ghosh (TA): I am gonna try one last time 01:55:39 Reshmi Ghosh (TA): After bhiksha finishes 01:57:41 Anon. Kirchhoff: I have another class, can I leave please? 01:57:47 Reshmi Ghosh (TA): sure