My Purdue Experience - Dimitrios

 This week I will be at Cornell, giving a presentation on my work, "An Accuracy Comparison of Support Vector Machines and Decision Trees in the Morphological Classification of Galaxies based on Their Cosmic Color Indices." This work has nothing to do with the research I was performing at Purdue but has been influenced a bit, as I was parallelly working on both. Since I am presenting only on Wednesday, and the rest of the program has nothing to do with actual work and is mostly PR and meeting people, I will talk a little about the connection between the two studies today. I talked to Dr. Nakazawa, and he said that would be the best option, as I cannot really write the blog in the same way I was up until now, for this week. The most significant Data Release in Astronomy history was released around two and a half years ago. The Pan-STARRS survey was finished in 4 years and produced 1.6 Petabytes of astronomical data. Massive volumes of data are produced annually by the different...

 There were 3 presentations total taking place. The first presentation was about greedy reclustering from a professor at Cornell, the second one was about Self adaptive systems in general, and then I was next. Before I went up to the stage, I thought of the reflections I had to write, and I took a look at the reflection, where I talked about the conversation I had with the professor. Before going up to the stage, I knew there was a specific word I had to have in mind. The crucial word here is necessary. You should always tell a tale and just provide the absolutely necessary information. You need to take into account two primary factors. One is your target market. The degree of intricacy should align with the amount of expertise your target audience is anticipated to possess. You must explain everything if the audience is high school pupils. You don't need to explain much if it's a group of experts. If you can't provide enough technical details about the study, make your dat...

Tomorrow, I will be explaining the techniques I used to test the two algorithms, the hypothesis, and what I have concluded. Thanks to my little yet important experience in researching, I was able to make a hypothesis based on prior studies and do what many people call and Newton first famously said, "Standing on the Shoulders of Giants". Each genius, inventor, and brilliant thinker has had their own personal "giants" that they have admired. It turns out that Aaron Bernstein and Felix Eberty's science fiction works served as an inspiration for several of Albert Einstein's thoughts. It was claimed that he "breathlessly attention" as he ate up their work. According to legend, Bernstein's theory of special relativity was sparked by one of his tales. I understand the phrase to imply that the accomplishments of today's scholars build upon those of earlier ones. I kind of always had a simple idea of what the phrase meant, but truly it is a brillia...

  I am grateful the presentation happened, and I think it was a great and insightful experience. Not so much so because of the feedback I got, but because of the questions. The question we got made our project more bulletproof and really gave me more creative and debugging thoughts. I would not start by saying this if I did not have proof of how much I appreciate the experience. After I got all these questions about how PCA and UMAP clustering preserve the local and global density of the cluster points, I thought, "wait a minute," what is the dimension of the vector input in the clustering algorithms. After checking the dimension of the code with simple input, it was not what I expected. For each vector representing a picture, I was expecting to see the 20 dimensions extracted by the Autoencoder multiplied by the number of pixels of each image, which I thought would be something like 10000 pixels or so. Turns out the vectors were 784x20. Why was the number so small? Took me m...

 Today was the day we eventually presented the work I was talking about in my previous post. I am sorry it was a bit long, but I knew I wouldn't have the chance to talk about it in this or the entry tomorrow because we wouldn't be doing any actual work apart from presenting and discussing. Therefore, I would not have any new details to share.  It makes me feel really good that I know that someone is following all the progress in my work, even the details because this is not a common thing in research. After we finished presenting our work to the professor, the professor first said that it is important to know how to sell your work. He said that from all the Ph.D. students he asked, over so many years, the thing they learned best in their Ph.D. work is how to sell their work. To be honest, I totally agree with him. It doesn't matter if you have two, five, or 50 minutes to present your work; there is always enough time. The only difference lies in the number of details. That ...

 Tomorrow is the day when Deepak and I will be presenting our research up until now on the quality assessment of ultrasound images to fully automate the ultrasound diagnosis process. I collected all results and research on past papers we have performed and compiled a 55-slide google slides PowerPoint presenting what is already in the market, what has not been explored, what we tested, what we got, and our assessment of it. Since we were asked to give more details of our work, I will provide some, as it will also be helpful for me, since I have to present them tomorrow. To be totally honest, from the conversations I have had with the professor and the reactions of other Ph.D. students that presented their research over the course of the internship, I am kind of scared. The professor is a brilliant and sharp man. He has decades of experience with research and is a highly respected man in the field of robotics, with Purdue University being among the best universities in engineering in...

After walking in today, I had a small chat with Deepak and then worked on testing the same model of the algorithm but applied on a different Neural network structure, which is very common in image segmentation, a resnet. I collected results, and it seems like it might be an improvement in the accuracy of the clustered and classified ultrasound pics.  Towards the end of the day, I asked Deepak for some feedback on how I have been doing so far and what he would like to see changed. I was relieved to hear that he really likes my research skills and that I work really fast on research and application. He said that he rather give me tips than feedback because he doesn't really want to give me feedback on research that I see for the first time. He said that I have improved a lot from the first day I walked in and that my ideas are better constructed but are not yet engineering ideas. I did not ask for a clarification here, but what I think he meant is that, when you have an idea, you can...

 I want to start by saying that I really appreciate the comments I get on Moodle and that they can be strongholds. Today was a testing day. I stress-tested the quality assessment algorithm on our data, evaluating its performance on very noisy and low-contrast quality pictures. The results showed that an unsupervised neural network could perform image segmentation for feature extraction and classification. If we say we have 5 levels of qualities for a picture, then a random guess would get it right, about 20% of the time or 1 in 5, right? The algorithm gets about 70% of the time right and is classified with a very high confidence of 90%. I am sure there are ways to improve it by flickering numbers here and there, but the main point is that it is working.  But how can we say that the algorithm got it right or not? Is there a subjective way to access that, and if not, then what is the whole point? As I said previously, I always like to look at the greater picture.  I dont ha...

 I know I have been saying I need to work on the NLP project, but I always seem to find more work. When you have an older sibling, they usually get to do most of the chores, and they get blamed often, even when the parents know who's a fault it is. Although I have a big brother, I now know how it feels to be the bigger brother, thanks to the internship experience.  I consider myself a more existential thinker, someone who always looks at the bigger picture, but that sometimes means I miss the small picture, maybe because I dont want to see it. I am really excited about the end goal, and I was so happy when we finally made a workable quality assessment algorithm. I immediately started thinking of the applications and the presentations, but I did not really think of the little things we needed to do before all of these even began. In research, we must be very thorough and test every single bit. That is what I have been doing for the past 3 days. Testing, testing, testing. I know...

 Today was a very hard day, but I gave it all since the weekend is coming and I can relax then. I performed a 3d clustering with 3 features extracted from the ultrasound images, using an Autoencoder Neural network, and the results improved the performance seen from 2 feature 2d clustering, but still, we need to do better. This is because we are talking about actual humans here. 90% sounds like a great number, compared to nothing, because there is no other similar product in the market, but imagine the machine diagnoses you with a random disease the rest of 10% of the time. How would you feel? That, for me, is really important. Because, at the end of the day, we will release a very good research paper. But, I am of no use if I cannot see that this product will not go well in an actual situation with human feelings involved. I feel like in research, there is a very delicate balance between looking at the small tasks and the milestones and also considering the big picture. Considering...

 I did not write a reflection yesterday because I was very exhausted to do so.  We have an autoencoder Convolutional network, where we extract 2 features from the ultrasound pictures and then cluster the data to classify the pictures. You can see the clustering in the photo below: Generally, I know that working in research is challenging mentally and that it is normal to get responses like, "oh, this is not going to work," and " we already tried this." But, because I get these responses, I always feel like I am not very worth having in the office or that I am not very useful because I always move slowly as I am learning. But. I realized this is a terrible way to see it because of two things.  1. Yesterday, the computer we were getting the ultrasound images off was not working due to a windows corruption error. Deepak and the other guy working on that were getting a sad :( bluescreen issue, saying "System Thread exception." Although that might sound very ba...

     Although I did not show it, I was extremely happy today when I left the office. The reason is that we finally managed to have a prototype of the feature extraction code. Although we are not even halfway through, I really believe that setting small goals, or milestones as we called them in data structures, is essential. That did not work in data structures because I think the projects were too small to make a difference. Nevertheless, it indeed does work in real-life research. They are essential, of course, in terms of keeping up with deadlines and work to be done, which is why I was using them from GSTR essay to Taichi practice. Though, there is another reason they are essential, which I had not thought of before. This is the excitement and motivation it gives you to move on.     Starting the theoretical research now on finding an area of interest in the ultrasounds, I learned from my mistakes and worked with more patience and diligence. The way I...

  Today, as i said yesterday on my reflection we worked on finding the best algorithm for the quality assessment. We implemented in auto and colder network, a very simple, one in order to extract features from their ultrasounds or generally pictures.  The hardest part was determining the features we wanted to extract from the auto code of network using simple Mathematical equations. In your network it’s very important to take care of your data set because the biases that exist in the data sheet as well as the noise will reflect on the final results. I learned that from trial and error.  The reason We chose the auto encoder in the first place,    is that after critically reflecting on all the papers I read about quality assessment using deep near network it seems like all the methods were based on very good train data sets. It’s important that we know and understand that on our research there are no high-quality ultrasounds we can get as well as label data, there...

 As I said in my previous entry, my goal for today was to work on finding 2-3 algorithms and then run them to see which one is better for quality assessment. Or at least that is what I thought when I walked into the office today. After less than an hour of working, Deepak walked in, and we discussed my progress. This discussion took more than half an hour; maybe he should start reading my reflections! A crucial thing came out from my summary of what I had learned and what I was looking at. He advised me to stop looking at complicated algorithms and focus on seeing some simple ones work. This, he said, would help me understand the flaws of different solutions and identify what I really needed. I respect Deepak deeply because he is always ready to support and help me while also being a mentor for me, as he will not tell me the solution even if he knows it, he wants me to find it. All our lives, from Middle school to high school and even college, we were always given problems, which t...

I think that today was the first time I got the actual experience of being a researcher. When I first walked into the office early in the morning, I sat down and continued from where I left it, reading about Quality Assessment CNNs and seeing which one would work for the Ultrasounds. Before today, I would just look at some papers for about 1 or 2 hours and then implement these algorithms on our dataset to decide. So far, I have not been successful in finding an algorithm that would be able to assess our dataset, trained without supervised data, I think due to the nature of Ultrasounds being very noisy and with low modality. Since both Deepak, who is the person in charge of the project, and the lab director were in the lab, I decided to ask them to help me reflect on why I was failing.  I asked what I should look for and how can I start looking at why my algorithms were failing. We had a discussion about debugging, and the director said something that sounded dismissive yet was a si...