Daniel Bourke walks through his study session on an AI capstone project when motivation is low. He uses 15-minute momentum bursts followed by 25-minute focused blocks, takes walks and power naps between cycles, and builds an LSTM model to classify speech using mel-frequency spectral coefficients.
Chapters
What is going on, y'all? Welcome back to another episode of Learning Intelligence. And I just woke up from a nap. I was sleeping on this couch here for about 20 or so minutes. And I'm about to get back into the voice user interface capstone project for the Udacity AI Nanodegree. I'll be honest with you guys, because that's what this channel is about. I want to share the good times and also the not so good times, the hard times. I was facing a little bit of resistance today. What does that mean? Well, it's when you know you should be doing something else, but you're sort of filling in the time with other tasks. So, this morning I've done a workout, I've cooked some food, I've listened to a podcast, amazing podcast on sleep on the Joe Rogan Show actually with Matthew Walker. I'll link that below so you can check it out as well. But I've been doing a bunch of other things instead of studying, finishing off the Artificial Intelligence Now degree because I'm up to a sticking point, I'm up to a hard point. And that's, I don't know, that's where you always hit a roadblock is when you get to that point where it's not it's not as free-flowing as what it is at the start. That excitement sort of wears off a little bit and it's just about getting the work done, getting into it. That's where I'm at. Essentially, it's just me being like, "I don't want to do the hard thing, right? So, I'm going to not listen to that resistance force anymore, the part of me going, 'Daniel, you could do all these other things today.'" I'm going to get into it right now. So, that's the project for today. I'm going to spend at least 3 hours on it. I'm rested. It's about lunchtime now. I've got the whole afternoon cleared to get back into some study. Alexa, 25-minute timer. 25 minutes starting now. See you in 90 minutes. Alexa, stop the timer. That's the first 90 minutes done. Oh, I can, I can hear myself talk now. It's amazing what just 15 minutes of dedicated work. I mean, why 15 minutes? Well, that's, that's what I find about the— once you get about 15 minutes into something, you're on a roll, right? So, that time just flew by. I've made some progress on the project and I'm really, really happy with it so far. So, what I got here is a, is a LSTM or RNN model training to recognise speech input from this sort of stuff. So, this is a— what is it called? A Mel-frequency spectral coefficient. So, that's a mouthful there, but essentially what it is, is a sound wave coming through the air, getting translated or transformed into an image by different heat maps. So this is intensity of sound wave and whatnot. And then we're using RNNs to, to model the sequences that it comes in off that. And that's probably not the greatest explanation of, of how this model is working. But once I finish the entire project, I'll be sure to upload it to GitHub so you can check it out. But otherwise, I've read two blog posts here that have really, really helped me out, understand a few things. This one is from Machine Learning Mastery. Which is understanding the difference between return sequences and return states for LSTM and Keras. I use that— where did I use that? Over here, maybe? Return sequences equals true. So I'll let you read that blog post so you can get up to date on that as well. And then one more over here by Cola, or C-Ola, I believe. C-Ola or Cola, I'm not entirely sure. Understanding LSTM networks, which is a relatively old blog post in terms of this field, how fast it moves, but it's a great explanation on LSTM networks with a bunch of different diagrams. I'm gonna have a 30-minute break, then do another 90-minute cycle, and hopefully my network's finished training by then. I try not to go on my phone because that thing is, is just a distraction device. I mean, it's okay for little bursts here and there around downtime and whatnot, but when I'm in study mode, I want to be in study mode. I don't want to get distracted by anything too heavy. Usually, I'll go for a little walk, as you just saw, in the backyard. I'll have maybe a snack or something, maybe, maybe not. I just had an apple because, I don't know, I felt like having an apple. Now, I'm going to have the rest of my break. I'm gonna go back to this couch again and sleep and let what I've just learned sink into my brain, because I just listened to that podcast on sleep. Seriously, listen to it, you'll love it. And then we'll get back into another 90-minute cycle. Hey Siri, stop the timer. Hey Siri, stop the timer. Now back into coding. Alexa, set the timer. And there we are. 3 hours done. And why was I procrastinating so much? Who knows? But my advice is, is when you're in one of those funks, just, just get in, set yourself a timer for 15 minutes or something like that. And once you get that initial momentum done, the hardest step is usually— oh, sorry, yeah, the hardest step is usually the first part, and then you get the momentum done. And all of a sudden, I'm 75% through this project. I've got my Second last model training here. It's going to take a while, but it's only through one epoch. I'm going to stop it early though, because I've got to go somewhere and look after a house. But I've got a few errors here. I'm going to fix them up. I've left myself some notes of what to do. I do this every time I'm with a project, so I know what to do when I come back next. And we can knock this off the to-do list today. So see here, I had artificial intelligence Nanogri times 6 Pomodoros. Is done. 1, 2, 3. And Pomodoros are 25 minutes each, so an hour and a half equals 3, and then another hour and a half. So 3 hours done of concentrated work on this. For the rest of the week, I've got work at Max Kelsen for Wednesday to Friday, so I'm really, really excited there. I'm not sure what we'll be working on. Still concentrating on trying to finish the artificial intelligence nanodegree. We're going to wrap this video up, and as always, Keep learning. But a side note to that, don't forget that even people like myself, we get into funks as well. We're all human. We're all on this, this learning journey together. When the times get tough, just remember, what's the alternative? I just not do it and just, just continually have that feeling. But it's like, no, I'm gonna get in, I'm gonna get this project done, and I'm not gonna listen to, to that voice in my head that's like, ah, but there's so many other things you could be doing. Not this time, I win. Leave a comment below. What do you do to get over those points where everything's difficult and it's like, "I don't really want to do it." Not this time, I win.
