Thank you very much ..😊 A very good explanation. Breaking it down into simpler concept. Liked the way you explained how 1x1 conv is networks in networks.
In this data, there should be seasonality because a person generally sleeps at the same time every day. After this, we can use seasonal decomposition and then apply the SARIMA model.
@@deeplearningexplained Yes, I mean that seasonality refers to the regular, predictable patterns in the data that repeat over a specific period. In this case, it could be daily sleep patterns, where a person tends to sleep at the same time every night. By identifying these seasonal patterns, we can better understand the data and apply seasonal decomposition techniques before using the SARIMA model for more accurate forecasting.
@@siddharthvj1ah right yes that would be awesome to have that circadian pattern information per participant! This study was conducted in a sleep lab though with heavy wiring of equipements so it might skew a bit the normal sleep pattern. There are two recorded night per participant I’ll check if the sleep stage pattern per participant match in another video!
Yes, linear or no activation functions makes the network only learn linearly separable planes. Non-linear activation functions are crucial to make the deep network actually deep.
@@deeplearningexplained Great! I was also watching your video about EDF files, I worked a lot with eye tracking on my PhD. Do you have some resources about eye tracking and Machine learning?
@@gustavojuantorena I'll check if I have something from my Master. I did some work on eye tracking + physiological signals feature fusion for a study on autism a while back.
Hello there! 👋 64 is the number of filters in a given layer. So in the (3x3, 64) case, it means you have 64 learnable 3x3 convolutional filters in that given layer. Let me know if that helps!
Amazing video, thanks so much for this! I had a question regarding what to do if you are conflicted on which roadmap to pursue. I eventually want to either work for or start my own AI startup, but I am also really interested in the research side of things. I happen to be a full-stack SWE currently, and was wondering do you think it is worth it to pivot to a ML engineer or Research Engineer role, maybe do some networking and and try to start my own company?
Ah great question! What I would suggest is to not go into research then, especially since you already have the full-stack SWE skills. If I were you I would go through the ML Engineer route since you are already pretty much there in terms of core skill set. What you have left to learn is in general how model are trained and deployed, then you are good to go to inject AI into whatever application you want to build. That's a very powerful combo. Afterward, when you get comfortable having AI into your SWE skill-set, then I would dive a bit deeper into certain research area to make whatever project or product you are working on better. That will be much more rewarding for you than going the opposite route with research first.
It is indeed very confusing for newcomer, but I’ve realized that a lot of the early deep learning open source developers came from a Physic background interestingly!
Too underrated video! Thank you so much for this video, now my vision is clear and I have set my goal - ML Engineer Job. Will soon comment here back, when I get a job/internship!
I still don't get why it's supposed to help translational invariance. As you say, the convolution is already capable of that. If I move the image contents 50 pixels to the right, the features should also move 50 pixels, given stride 1 and padding. Exactly the same is true for traditional sobel edge detectors. The edges don't change if I convolve the image with the edge detection filters translated or not
No the convolution help with translational equivariance, but not with translational invariance. The idea of adding pooling is that you are forcing exact nearby pixel/feature information to be “lost”. This means that the network is forced to learn more generalizeable components of the picture you are showing it.
Very useful indeed, I just think the 3D representation could be a bit better (I'm used to seeing the filter rectangle behind the first layer and not on the side, but that's probably just me)
Hey thanks for the kind words! Will do, I have a few more videos ready for next week :) Do let me know if there is a specific topic or question you would like covered!
Glad you liked it. It’s the second time this week I had this request for where to start in deep learning, I’m setting up a video on that topic will publish soon!
Ah then no, VGG shouldn’t be your pick here. It was expressively designed for image classification. Take a look at the various model on PyTorch made specifically for audio signal: 📌 pytorch.org/audio/stable/models.html
Yes you can, but the internal of the model is tailor built for image. If you are able to express your 1D signal input as an image I would say it might be worth it to try. However, there are other models made specifically for audio.
Hey, fyi I had to reshoot some of the section on this video because I couldn’t stop saying Drop Path (from Fractal Net) instead of Stochastic Depth. There is still 1 wrong mention of drop path in there that I wasn’t able to fix haha That's what you get from reading two paper simultaneously!😅
Yes, but this shouldn't have happened. The first file contains 2M characters, while the second only 1M thus |file_2| = 2 * |file_1|. The author of the video wanted to have a first file with 500'000 "ab" (thus 1M chars) but that's not what he did.