DeepLearning.AI Learner Community Event ft. Ghaith Sankari

Good morning everyone and with consideration to the time zone, good afternoon, good evening and good night. Welcome to the first DeepLearning AI mini event related to AI for medicine specialization. I'm honored and excited to be your first host and speaker for this event. And I'm very happy to be here with you. For sure, this event could be weekly event and each week you will meet one of the men and I chosen to be the first one because I the laziest mentor between the old. Everybody was busy with answering your questions, but I was busy with answering the email from others about this event. My name I will start with brief about myself and my career. My name is Raif Sankari. I'm from Syria, Software engineer and AI engineer. I graduated as a bachelor from University of Aleppo in Syria in 2008. My specialization was AI artificial intelligence. And at that time it was very excited to go into AI. In fact it was just kind. There is no resources to understand exactly what AI. What made me choose this specialization? Because I was excited about the genetic algorithm at that time Deep Learning 10 wasn't. So we know that there is something called artificial neural network. But the genetic algorithm was very good idea to start with the AI because genetic algorithm is just. You are encoding the problem that you have into a solution using something we name it as chromosome and then you are going to make motivation and crossover to generate a new solution. What an idea. It's something great to do. That's why I choose this specialization. And I studied two years in University of Ariform. When I graduated and I go to the market I found very hard to work in. It wasn't in my plan to go on the research and to make master and PhD and work for that. I want to add real value to the real world according to my point of view at that time. So in 2010 I moved to UAE and I was working for one hospital, one of the biggest hospital in the Middle East MIDI Clinic. And each time we're facing some problems or facing some challenge that needs solution, I was suggesting to use artificial intelligence to solve this solution. And at that time, I was the only one who was singing on this bathroom only means that he wants to use the architect. Everybody was looking for me and just telling me oh, what you are doing, what you are trying to do. So until 2013 at that time, when the concept of AI of deep learning appeared, everybody start to check, okay, who was talking about this? Who has a certificate? Yes, I was able to find some appreciation at that time in 2013, 20, 2014. And I start let's say check my way on this one. The most challenging thing here is to keep learning and how to keep learning. The good thing that in the last two years I just discovered course era and discovered Dr. Andrew and also I get the deep learning specialization at that time for deep learning specialization I like it very much was very simplified all the concepts that I have during my study but it was very simplified and very clear idea. So I finished the 5 courses specialization within one month only. Thank you. This is my son, he wants to join but I will tell him just please don't. Okay, According to that after that I wasn't able. I wasn't able to just tell him please don't open the door. Sorry for that. So according to that I was trying to keep learning by project. I was trying to work on the project and to keep working on this project and find any challenge over the Internet just to keep working and to have more education about it related to the AI for medicine. In fact I have early trial on this and it was my graduation project on the university just to create his system that has diagnosis, uses AI and according that I used at that time forestry and it was very nice disaster project because for sure I have a lot of knowledge that I don't have. I have right now a lot of knowledge I don't have at that time. But at that time it was very good start for me When I joined the hospital it was good also to have some information about the health informative and according to that also this is help me on understanding how to label data. Since there are some countries they have SIS system or EHR system over the country then the label data is already there only you have to use your data, cleansing your data, balancing your data and then you can implement the solution that you have. So briefly, this is about my career right now. Finally starting from two months back, I started my own startup here in Dubai with the partnership with Golden Bear in India. Right now I'm working on something related to the AI for medical diagnosis related to the COVID 19 detection using the X ray images. In fact, the research itself, there is a lot of papers it's already mentioned and already I shared some of the labels with you. And also you also shared with me some data sets shared with everybody on the Slack channel. The good thing about what I tried to do, I tried to mix both CT scans and X rays and also I have good findings related to this mix and which one is giving more accuracy on the detection. But also my focus is not to create the or to reinvent? No, my focus was to create a product something that able to be integrated with any ESR system or any hospital information system. According to that, everybody will get this service as able to be used. And right now we reached that point right now and we are trying to launching this on many hospitals all over the world, not only in Dubai. So according that I find that there is a lot of challenges during this. The most challenging part here is the client for our work, for our clients. Our client, they are doctors. And the doctors they are finding any kind of building a machine learning model for diagnosis as a challenge for them why we are unable to diagnose the situation. No, they are able to do that and in fact they are our baseline, they our information resource. But it's not easy to tell them that this is what we are going to. We need you for. Especially here in the Middle East. They find it direct challenge to them to be to create machine learning models for diagnosis, for example, which is my focus. In fact, I'm focusing on the legacy also in the health insurance part. And also some of you maybe they checked on the deep dive my research on how to predict the approval of the claim. So this is the audience. They are the most challenging part in our world. The second it's dealing with the data and trusting the data source Also very important because we are dealing here for medical researches. We are dealing with. We are not dealing with anything else. It's not about economical crisis, it's about death or life. So trusting the data source and get trusted data source is very challenging part. And this is what we trying to fix on our research for COVID 19 detection using X ray. And by the way, according to my research, X ray is better than CT scan for that. Maybe some people will tell me no, it's not the situation, but this is the time. So this is the most challenging part in our work. And according to that I have one advice to all the learners. Just to choose your track from the beginning. Either you want to be a researcher, the one that who is involved in research for universities or they are creating great things. But at the end it's not product. There is somebody has to change, make it as product. So this is one track. The other track is what I choose to be AI engineer because mainly I'm software engineer. So to know to get this research and to create the suitable structure to give it to the people and to make it usable and not keep it only on the research. This is that one track and that one is another track. So to do to Know exactly how to learn and where to learn. For shoes you have to choose those. For me I'm for sure I'm one of the deep learning AI community members so I get all the benefit of that. The batch each week from Digital Andrew Worker also is great place for learning new things and also for test shooting skills. Also the Kaggle and the project mentioned there. And also I prefer to learn using the project. Using the project you will face the real life challenges, not only the research challenges. No, you will be on real life. This is the most important part is according to Dr. Andrew make your hands dirty and start working by yourself. And really this is let's say gold advice. Not only an advice, this is the key strong advice for anyone who want to go deep on this thing. So according to that, this is my introduction for this event. I hope it wasn't long for you. I will start to get your questions and to answer you for sure. I am encouraging you to post your question on Slack channel right now. I have a couple of questions already with me. I will start with Anthony Morada question he asked me how do you draft plan for new projects, do you diagram, write out that process or just jump to the code, etc. I guess I'm curious about what does the step before writing code look like? Okay, right now we have to define what's the meaning of the project. Because the project has many meanings. Because if you are in software development field, you know that the project itself is a product. That means it might be a client requirement and then you start as project. It might be a product that you find it's important and you want to make it as a project. And for AI it might be a result. That means you want to create a new CNN artifact, for example to detect some disease or something. And according to that it's project. But it's at the end you are not looking for final product to be in the market first, you are looking for that. All of these cases there is one keystone here. It's problem statement. You have to create your problem statement and to make it. This is the key to more projects. After that you have if you are talking about product or project or client, you have to determine the risk. All of this something not related to the AI. It's something related to any project management determine the high level risk. What might stop this project from going on, what might the feedback. All of this has to be studied. But if you are looking for for example, if you are talking about research, the risk is the data set. For example that source that you need. Also there is another thing you have to do when we are working on research is to check other research, to have access to other research and what other teams are around the world try to do and where they are reached. Because it's very important to start from the end point of everybody, not to reinventively we are able to build on our work together. So this is also we are not starting with a code until now. We are studying that problem itself. The risk around the problem. The research is the other research. Also it's part of the study that we have third step, it's very important to have some statistical analytics about the problem that you have. This is also explainable. In two cases you are talking about product or a recent. So what happened there? What's the data science take? What's the information you have to get full from this information. Until now we didn't speak about the coding. In fact, when we are working on products or projects for the client, the project starter is playing the main rule here. Because you have to develop the project charter, including the risk analysis plan, the risk facing plan, the change request plan, all of this, even if it's AI project. But also at the end it's for a client. And since it's called a client, you have to be ready with the budget chart. After that, when we are ready for that, we have to determine our final what we are going to do and according to that. Yes, it's very important since you have a team with you. Nobody can work alone. Since you have a team with you. It's very important to make diagrams, flowcharts to determine for each block of your cause or project what is the input and if so the output. Now here we also we have to differentiate between the research or the AI research or any other product software brother that you have. You. You are able to for the AI you have to work on the data and is the data is ready with you or not yet you don't have the data. Then you have to find a way to get the data. Otherwise you have the data. Then you have to play with this data to make it ready with you. To make it ready to use to be input for any kind of solution that you have. And also part of the project planning for AI project is to set let's say of a process that you are going to use. For example, one of the projects I determined three steps. First, I want to try the random forest tree. Second, I want to try the CNN one of the CNN common object and the third one is the RM all of this, it's happening before writing the code. All of this is real job and it's very critical job. And it will be before writing, to be honest. Sometimes you will jump to the code directly. Why you want to test something you want to try for a small Data set of 100 images, for example, what is the result that they are going to have? And according to that, you will find if this project is promising or not. So sometimes you will jump into the code, but mainly jumping to the code. It will never, will never give you a clear and successful. This is for Anthony. I have another question in fact from Dr. Adam. Dr. Adam, he's talking about his medical doctor and he has. He is looking for creating ML machine learning collaborative projects. So for him he is asking how to break into AI. Dr. Adam, you are exactly to say you are the center for us medical doctor. He's encouraged to go on this field on this and he is not finding it challenging for his skills. You are. You are the gold beast for it. Because this is the right way. You are our exact. You are the source of information. For example, if we are talking about right now, what the last project that I work on, the COVID detector. If there is no medical doctor and researchers tell us that really this Covid has real impact on the X ray image of the patient, we are unable to do anything. We will never think that maybe machine learning model will solve the learning. So you are already in the right track. You are already in the right place this AI for medicine specialization. And exactly you have the right background for sure. Maybe some technical points will be hard for you, but we are here and also our colleagues, for sure they are as I told you, they are better than me as a mentor. So they are here to help you. And also we are waiting for your input because your input is the most appreciated input that we are looking for right now. Personally, I'm looking for someone like. Okay, I will set another question. Right now I have. What are the potential deep learning application in medical service which is seeking work. Sorry, maybe I didn't get that question. The question with vendor. I didn't get the. What you mean with that question. Exactly. If you are able just to post a small clarification about it. The potential deep learning application in medical science which is. Okay, maybe I didn't get what you mean. Forgive me English speaker, but my mother language is Arabic. Sorry, so maybe I missed them. Makama, would you encourage us to have a study group where we will practice together to solve problems during and after the course for sure I encourage you for this is very important and since you are on in this community for deep learning AI maybe you have find this group with us for sure it's not my decision maybe you will find that you are able to be alpha tester and once you are alpha tester you will be able to test new courses and also to check whatever happens. Now if you want to study group it's very very very good idea for you because this will help you out. Nilesh on career path research this AI engineering what about AI product management? Do you see people that are skilled at bridging customer requirements problem domain through AI engineering team? I will not go deep with your question maybe I will choose this part to answer. I will tell you there is common let's say yes to bridge the gap between the requirement and the research the business analyst is main rule for folk business analyst at the end he is the one who has to be there in each and every technical project either it's normal classical project for software development or AI According to Johns Hopkins University I got course from them related to the executive data science. They are mentioning four rules which is important for each team that's working on any data science project including a machine learning. The first is the data engineer, the one who is able to present the engineer to the model that suggested the data scientist brochure and then project manager or product manager together. And the most important one is the business the business analyst has to and this person, he is able to bridge the gap between the requirements of the science and the engineering. I hope this will answer your question. Ritika great in science could you speak to what you think are some of the biggest bottlenecks in reliably applying AI for medicine? Biggest bottleneck in fact is the trap most of the time when we are dealing especially something related to the diagnosis, for example brokennesses for medical prognosis it's also important but at the end it's kind of statistical. But for magnesis nobody will trust that okay, this diagnosis is coming from machine. At the end if we visit a doctor and he tell us that you have this disease, maybe we go to get another consultation from another doctor. And this is the let's say he is a human, he is very clever, he's very smart and he has the necessary experience. So for us to go there, we need to convince the community, the medical society that okay, this model is trusted enough, the science is there and maybe when we are dealing with the market, this is the, let's say the most critical Dealing with the market. That means you are not dealing with the research center in one university or something. You are dealing with people who are just facing the patients and doing the daily protein for them. And you are going to them and tell them, okay, I have this model, it will finish and according to their understanding, it will finish the work of the regulators that you have. Okay? This is disaster for them, okay? They will not be happy to do especially for the market. In the medical market, you have to find that people they are trying to deal with the patient as clients. They are not patients, they are clients. So in fact the trust is the biggest you have. I will check another question. Sorry if I didn't answer your own question. I will make some notes on Slack channel after finishing the event because according to plan we have only 10 minutes. Maybe we are able to for another 10 minutes. But please forgive me if I didn't answer all of the questions. I will make some notes on the channel after the Ivan, in your experience, what are some of the challenges you encountered when working with the data you receive to train test your algorithm? For example, difference in how medical professionals process data or even Aquarius could lead. Difference in quality. Yes, it might happen. First, I start from the end according to my experience, as I told you, I have good information about the health informatics and medical codings. So when I start with the first two machine learning core medical for medical diagnostic model, I start from labeled data and labeled and standard. So I used the claim information at that time and this data is labeled already we already know the disease and what is the procedure performed by the physician and all of this why I use this because this is bad. But later when we are. When I'm dealing with this for example Covid detector or diagnostic model, the most important thing is the resource. Because for example, I have a data from Canada, from Singapore, from China. But I'm trying to implement this model in Dubai and in India for example, and in Nigeria Right now this is the bottle that I'm importing. So the quality of the machine, the quality of the technician who's working on this X ray machines, it's different. First and the quality of the reporting and the language of the reporting is different. So how to get labels better? This is the first tenant. The second tenant is when you are dealing with data set like this, you have a lot of, let's say, concern about the legal part. Because if the patient is approving to use his information for this result, some of us are not using personal information. I'm not using his name and father name and all of this. But the gender, date of birth, X ray for his situation, his admission date, all of this information will be included on the data. So also the legal part is important. Maybe somebody will tell us. No, I'm not happy with that. I don't want to be part of this research. So. And right now, according to this isolation, we have to manage all of this approval from home. So you can imagine what kind of challenges we have. So all of our research according. For example, as I told you, I have data sets from all over the world. But I'm telling everybody I'm creating a prototype. I'm not creating the final product right now. The final product will be when I get the data from the country that who needs the product. For example, if we are talking about hospital here in uid, I need the data from uae. At that time I can retrain the model and gather that. Also one of the challenges, it's the data side. Until now, for Covid research we have small. We don't have enough information, enough CT scans, enough X rays to get this for sure. We tried by transfer learning to face this and we get until now good result, Very good result. And we are talking about, for my model at least, I'm just preparing the white paper for that. We are talking about 98% accuracy. But until now this is. I'm unable to trust this 100% that it will be a product ready to be used for all over the world. I'm taking also another question we have blasphemy. What? What are the possible ways of applying machine learning in neurology? In fact, I am unable to answer this question clearly because who will answer this question is a mix two people, let's say AI engineer or data scientist with a neurologist. He has to tell us some of the information related to how they diagnose the neurology. According to that, we might start researching. There is something. Sometimes some data scientists come with, let's say crazy ideas and it works. For example, they want to maybe from the face image to detect some weirdnesses related to neurology. Maybe this will work. Maybe there is some symptoms. I'm not sure about it. He finds it and then he. He's able to build a model according to it. But who is that? Let's say again to the Dr. Adam question. Who is the one who will lead us to go there? He's the doctor, the neurologist. For example, here, Mohammed. Hi guys. Hi. In working with machine learning, how do you deal with the lack of Data. When you cannot find large enough data set we have to find it the easy answer. But for example, when you are talking about something that image processing it's very useful and I find it very useful really to use pre trained. Let's say I use pre trained resnet and pre trained inception network just to retrain or fine tune it to be able to differentiate between Covid or classify the image in Covid and non Covid. So by using pre trained it was pre trained for example for animal classification that this image for cat and this image for dog. There's 1000 class on that network and I just add the final layer just to train it to work for Covid analysis. So for example, this is one way transferred learning. And also for sure you. You maybe you notice that Pranav on the. On the courses he mentioned something related to the data augmenting or resampling. Resampling the user clock. Because for example I face 7000x ray image available for pneumonia but it's not related to Covid. For me it was non Covid images and I want to find a suitable amount of COVID images for that. So resampling its workload for me in this situation. Joshua, thank you for great insight. Thank you. Thanks for you for being here. First, how do you handle privacy issues on the deployment of medical AI given in fact I'm as one person he's unable to, let's say to solve this issue. This issue it will be able to be solved when you create a proof of concept according to the open data that available for you and then you go to the, let's say some organization or big hospital research center and ask them you need support at that time they will be able to deal with this if they believe me. The good thing that I found this. I found this feeling and this is. I really appreciate that from all the bosses that I'm working with because they are able to support our ideas on this. Maurice. Maurice. He starts with Marhabalai. So I can understand that maybe he is from Lebanon or he's from Syria. For my country. I would like to ask what do you think is the best way to learn how to manage an AI tool? Best thing to learn, best way to learn how to manage AI. I don't know if you are. If you are involved on deep learning specialization from deep learning AI. The presenter of the Inspector he was the. There is one course for that specialization. It was really how to manage AI. That course I really. That's the best. I think it was the third course on the on the series of five courses it is the best. Really it's very, very, very good place to start and to learn about Vector list. How would would you advise striking the balance of no understanding enough about a field coming from healthcare background thinking and not fear. Okay, how do we know we know enough to be able to bridge the gap between the technical aspects. Of course, for in fact this is very nice question and I'm not from the. Let's say from medical background, but because I worked for Mediclean for five years and I was involved a lot on insurance processes and in health informative and I worked with the latter. So I was able to understand exactly to get enough information, let's say really not enough. But right now, let's say for our example, enough information to tell somebody or something. But how to raise this gap? At the end, the AI is on the high level of the AI. We are talking about machine. So when you are following some courses related to that or to get into some projects to the scientific ideas that we are looking for for researchers, not for products, you will be able to learn to get the information needed to build AI models. But at the end you will not be a software developer, for example, or technical software engineer or software architect. But at the end you are able to manage your project with your team to forward it to get the desired result. And then there's another part of it is to create suitable assets. So we are talking about mathematics here. So at the end there is nothing, nothing. There is a lot of things you have to learn with the software aspect, but you can start from the AI itself as much. Again, I can forward you to deep learning specialization from DeepLearning AI because it's very great and there is no need to have technical background for it. You can go ahead and start it. You have to have strong background. Ashish, I hope. Ashish Kumar. I hope I mentioned your name incorrect way, sorry for can AI able to help us to predict the pandemic and inventing medicine of such diseases like COVID 19. There is a special field in AI for medicine related to discovering and developing the drugs. So yes, already there's some research related to that. You can find something already published. There's some companies and startups already working a lot on that. And in fact I was suggesting to deeplearning AI also to have 4 course maybe or another course related to AI in drug discovering and develop. Yes, there is something and it's helpful now to predict the pandemics. To be honest, according to my notice, all the let's say the processes to face these pandemics around the world right now it was a work from data science. I can feel that there is data science mill behind how to just make the isolation, the physical isolation. Nobody has to go to anywhere or something. This is not political decision, this is data science. So it's already we are facing the randomness. According to my, let's say expectation using AI Hatin the ruling, do you think that AI is able to make the medical diagnosis also, although the result cannot always be accurate, how can we trust its results, especially in such critical fields? In fact, this is the term that I mentioned during this. Yes, we cannot. But when we understand the AI process and since we are trying to the concept, we are learning from the errors, okay? And we let that network or the architect that we have to learn from the error that we have. And according to that we are minimizing the error to the lowest part, to the lowest portion that we are able to. So we are sure because we are trust mathematics. Okay? We are sure it will work. And the the error level it will be limited. Now, for example, I'll give you another example for what bosses I suggest for copy. I suggest to make the sensitivity of the model as high as possible. And then if it's sensitive enough, that means the model that I create it will give me positive, even for some negative cases. So according to that I was trust enough that if the model tell me this patient don't have cov, doesn't have CO B, then I'm sure he doesn't have. But if he has a Covid, then we have to go and make PCR screening for him. According to that we will be sure if he has a Covid. So we solve 50% of the problem by raising the sensitivity. Since we are unable to reach 100% accuracy and without overfitting. Right now, according to the data that we have and to the situation. So by doing this, yes, we are. For example, our model will tell us that if the model return negative value for the X ray image, we will be sure that he don't have code. But if he. If it's returned a positive value, then it will be. Maybe he is negative, but we make a mistake. So this is some of the processes how to make decision before implementing the model itself. The model it will be able to help. Again we have Arwen. Hi. So I'm radiologist doing my own little side project in AI for radiology. You are some I've deployed to the club. But I always wonder what track I Should take. I have seen, I have seen some job description here like data science, AI engineering, AR research, AI project management. Okay. For radiologist learning. Okay, okay. For radiologists, I think the best thing is business analyst. This is my, my say, my expectation first we don't have AI project, we have project manager. He is the person who is able to manage any project and if he has background related to the AI he will be very let's say great project AI engineer is there this position and there's another position. But for you as your ideologist, if you are able to break the gap and to have all the technical and especially mathematical and statistical knowledge that you need, you can go for data scientists. But what's your place is the business analyst because you have clear idea about the business itself. And since you have let's say enough idea, enough information, all the AI processes, you will be the great business analyst that everybody is looking for. Nileesh, which doctor doing this question from me, I hope I also mentioned the name which doctors going AI doing AI and engineers learning medical what is the best approach to balance breath and knowledge based depth or focus in specific areas? How do you see this evolving in the next 50 years? Maybe I didn't. I will read the question once again please. Sorry, sorry. Image but with doctors doing AI and engineers learning, what is the best approach to balance free in knowledge? Okay, okay. In fact both are acceptable for sure. When we have a doctor learning about AI they will never be software architects that they can create product. For example, when you create a service to diagnosis one or more than one disease if you are implementing multitask, okay, the doctor he is able to suggest the aspect, he is able to train the model but he is unable to publish them. He needs software aspect. He can understand the networking concepts, he can understand the load concept and how it will be the data replication concept. So the software engine has been. And also if you are software engineer or software and you know about the AI, you have good knowledge about the AI and you are keep learning. But you know want to go to the medical part at the end you will never be alone, you will never be okay. Right now I can, I can give you small description about it. If I run my model for, for Covid detection, maybe it will tell you that this is Covid or not, this X ray is real or not. But me myself who created the model, I'm unable to tell you this information first. Second, I'm unable to give any kind of medical, let's say advice to face Covid. Okay, so at that time the merge that let's say we are we have to be in the middle all of it the software or the technical people and the doctors they have to let's say raise their knowledge until they reach the point that they are able to talk to each other this is the most important and then we have to create a team to create that let's say significant model that we are looking at so I'm sorry for that before I think there is the. Okay I'm sorry I go more than expected time if there is any other question I didn't answer I will check the questions once again and I will make notes on them thank you Alice for your guidance some deeplearning AI thank you for being here I hope this event will be weekly event and each week it will be with another mentor or practicing person who is to deep dive into that AI for medicine thank you for being here I spent a very nice time and I hope you enjoyed also thank you thank you.