To start off the week, I asked the students if they had any recommendations for learning about their projects. A couple of them mentioned a machine learning class on coursera that I should look at to gradually pick up the ideas. Another thing they had all mentioned last week was Deep Neural Networks, but I didn't really get the concept then, so I took some time to learn more this week.
So what are these neural networks? They are an algorithm to take data and classify it into groups based on multiple features. Well... that's the simple way to put it. What they actually do is really amazing, or rather magical, in the words of Prad (the student working on the cough project). It is a perfect example of machine learning. It is given a training data set and decide for itself how it should group the data into the correct groups. How does it do this? It uses a system of weighting and checking (not an actual term by the way). First it comes up with a set of weights, one for each input, which basically set the importance of the inputs. Then it makes a guess, and if that guess is wrong, it will go back and change the weights till it minimizes the error. Make sense? maybe not... let's try an example.
Say you want to sort a bunch of pictures of cats and dogs into two groups... cats and dogs. The neural network would first decide on inputs, in this case, features of the animal. It could look at size, shape, color, or something more specific like eye type, ear shape, or tail length. It would then pick an importance, or weight, for each feature. Based on those weights, it could decide that one feature determines that a cat is a cat, like its size. But this would not always hold, so it would then change the importance of the features until it can make the best possible guess. This is a little harder to explain than I expected; to make it better I'll add cute pictures of a cat and dog.
Besides learning by myself and with the students in the lab, I started auditing Dr. Berisha's class this week. The class is called Signals and Systems. It was a little difficult to try to pick up the class part way through the semester, but I am still able to learn a few things about signal processing, invaluable to my project.
Man... This might have turned into a pretty confusing blog post. Please ask any questions you have and I'll try to answer them as best as I can. Hopefully, next week I can learn more about machine learning and explain it all to you guys then.
So until then... Buh Bye.
Hi Luke! I'm so glad to see that you've been learning a lot in Dr. Berisha's lab and class. How is Deep Neural Networks used for looking at impaired speech?
ReplyDeleteBy applying deep neural networks to a signal processing/speech analyzing algorithm it makes it possible to do analysis on most types of speech instead of just speech similar to the test samples. The neural network will hopefully be able to learn enough to generalize to other speech. There is also the possibility of being able to do real time analysis because of the efficiency of a neural network.
DeleteHey Luke! I'm not gonna lie, but it was a bit confusing reading through your explanation of neural networks. However, I've made it a goal to research a bit on the subject so that the rest of your blog posts can make sense, because it seems very interesting! While auditing Dr. Berisha's class Signals and Systems, what was the concept you found most fascinating?
ReplyDeleteThe class right now is focused on just one concept, Fourier analysis. This is actually something I learned a little about earlier this year, but it is interesting to learn it in more detail. Hopefully I will be able to learn about the previous units that I missed, because they were more basic signals and signal processing units that seemed interesting.
DeleteHello Luke! I'm glad to see the improvement in the project and the use of machine learning in your project! I'm very fascinated in AI, so implementing machine learning in your project further peaks my interest. Just a question: How are you going to implement the neural networks to determining their speech patterns?
ReplyDeleteThe way I am looking at deep neural networks right now is in the context of the cough project. This means that it is used to determine what is and isn't a cough, very similar to the dog and cat example. the goal is to be able to take the neural network and train it enough to be generalized to all types of coughs and speech samples.
DeleteHey! What you are saying makes sense in my head, but I hope it's the same idea as what you posted. So is this basically an AI in a way? It's cool to see this progress, and I can't wait to hear more.
ReplyDeleteYeah, it is pretty much an AI, I will try to give a few details about it in the week 3 post, but i think the primary focus of that post will be the class I'm auditing. Just as a sneak peek, the unit in the class is fourier analysis, so you already know a little bit.
DeleteHey there Luke! So it is basically an AI? So if it is, does that mean it will the capacity to identify a patient to see exactly what disease they have and know a cure to the disease?
ReplyDeleteYeah, it is an AI. I believe that is the goal, to be able to diagnose a patient. I think that it is very possible for the algorithm to, once it has found the disease, find the cure.
DeleteHey there Luke! So it is basically an AI? So if it is, does that mean it will the capacity to identify a patient to see exactly what disease they have and know a cure to the disease?
ReplyDeleteHello Luke. I've actually been reading a little into deep learning recently, so its addition to your project makes makes me all the more more interested. You're doing some really awesome work, and I'm looking forward to next week's post!
ReplyDeleteOh nice, that's really cool how you were reading about it before hand. Next week's post will focus a little less on the deep learning, but I hope you still find it interesting.
DeleteHey Luke! I am still very much interested in what you are doing, but have a few questions. What practical uses does the Deep Neural Networks and how would it relate to helping find diseases?
ReplyDeleteHi Luke! That was a bit confusing to understand, I had to read it twice to comprehend what you were saying about the Deep Neural Networks by not getting distracted by the cute cat and dog gifs the second time around. Was that the same example they used to explain it to you? If not, what did they use to explain it?
ReplyDeleteYeah, when Dr. Berisha explained it to me the first time it was with that example. Another way that I heard it explained was with chairs. Prad said that the neural network should be able to learn what characterizes different types of chairs, and run through all the objects in a room and count the number of chairs. I'm not sure if that one makes more sense, but that is another way it was explained (I just thought a picture of chairs was less interesting than the dog and cat).
DeleteSeems interesting. I have a question about the algorithm you described. Is it essentially guessing and checking, or something more sophisticated?
ReplyDeleteIt is very similar to guessing and checking. For the most part, the meat of the algorithm is in the part after the checking though. It is supposed to be able to tune/change the guess to get closer and closer to the right answer.
DeleteHi Luke! It seems like you're learning lots with Dr. Berisha. How exactly does the systematic weighting and checking affect your study of impaired speech?
ReplyDeleteThe way I am looking at this system is mostly in the cough project, so a little less on impaired speech, but I can still apply it to that topic. I don't have as much knowledge on how it is applied, but this is my best guess. Very similar to how you look at the characteristics of a cat or dog, the network would look at the characteristics of speech, like tone, speaking rate, slurring, and classify the speech based on those.
DeleteHi Luke! It's awesome to see you progress in your Senior Project and see how fascinating your work is. Question for you, are you going to your own Neural Network?
ReplyDeleteI think there is a missing 'make' in there, so I am going to answer it that way. I am not sure how feasible it is for me to make my own, but I will be helping with the creation of the one for the project.
DeleteI am glad you are taking some of the classes you have mentioned to help you along in your project.
ReplyDeleteOn a side note, would it be possible to change the background picture on your blog. Personally speaking, its hard to read white font with so much going on in the background. Just a suggestion. What is the significance and connection of this background you have chosen for the blog?
Yeah, this blog background doesn't really have any significance, I just thought it applied well to the sound/speech part of my project. You can see, I did make a slight adjustment to the background, making the color behind the text solid instead of transparent. Hopefully, this makes it easier to read, if not I can change the background more.
DeleteHI Luke! After reading your example, I actually think I have some *minimal* understanding of deep neural networks! How is this being used in the different research projects?
ReplyDeleteI don't know the exact details of the implementation of it in other projects, but I do know that it helps them make better generalizations of the algorithm from training samples to real world test samples.
Delete