On the Next Frontier in Machine Learning and Computer Vision with Youseff Hosni

Sheikh Shuvo: Hi everyone. I'm Sheikh. Welcome back to Humans of AI, where we meet the incredible people building the technology that's changing the world. Today, I have a very special guest, Youssef Hosni, who's a prolific researcher, writer, and educator in all aspects of AI. Youssef, thank you so much for making time today.

Youssef Hosni: Oh, hi. Yeah, it's really nice to meet you and thank you for having me on your podcast. I really appreciate inviting me and I hope this session, or this episode, will give some benefits to the people and be interesting.

Sheikh Shuvo: I have no doubt about that. Well Youssef, I like to start the podcast with the same question for all of my guests, which is basically, how would you describe the work that you do to a five-year-old?

Youssef Hosni: Okay, that's a really nice question. Yeah, usually it somehow seems complex, but if I have to explain it not only to a 5-year-old, but also to my parents, for example, because I think they might be like...

Sheikh Shuvo: Oh, that's a much higher bar. Explaining to your parents is much harder.

Youssef Hosni: Yeah, they just see me on the computer, but they don't know exactly what I'm doing. But, what I say to them is that I try to make the computer think in a human way. So, we give them intelligence just like to understand us and do some tasks that we are doing, but in a better way, in a faster way.

Sheikh Shuvo: Nice. That's a nice, simple one. Just make the machine think and act as human. Awesome. Of course, in some tasks we are still pretty far from it, but yeah, in some tasks the machine is already doing good work.

Sheikh Shuvo: Nice. Well, Youssef, you've had such an interesting career, all throughout the world, really. Can you tell us a little bit about your career journey and what were some of the inflection points along the way that led to where you are now?

Youssef Hosni: Yeah, so basically, I'm still in the beginning of my career, I would say, because I've just been at this for 5 years. I started off being interested in biomedical engineering back in university. I had plans to do a master's in this field, and that's what I have done.

So I joined the University of Oulu in Finland for a master's degree in biomedical imaging and signal processing. I specialized in biomedical. During that, I got exposed to working with medical imaging and started applying computer vision algorithms and deep learning. After finishing my master's, I planned to first take a break from academia and do some industry work because I wanted to get this experience and work on real products that could be used by people. But then I realized there weren't a lot of opportunities in this area since it's still under research.

So I thought I still wanted to get this experience in the industry and working on products like this. So I had to shift a little bit towards data science and AI to be able to find opportunities in the industry. After that, I took a small position in a startup, and at the same time, I was still working in research. So I was working on both. At this point, I felt I had a lot of things I wanted to share with people. I started writing a bit on LinkedIn. I think LinkedIn has some restrictions on word length and people usually prefer to just read short posts.

So I wanted to share more, and that's why I went to Medium because I like to have the flexibility in writing more.

Sheikh Shuvo: Good strategic choice.

Youssef Hosni: Exactly. So it was not planned; everything happened as I needed to do it. So I searched for how to do it, and that's how it went.

And after that, I joined a company as an applied researcher. But something happened, which got me fired for some political reasons. And yeah, recently, just two weeks ago, I joined a startup as a scientist. That's my journey. During this, I've also been trying to do some mentoring for people. I've mentored around 100 people and also done some consultation for startups.

Sheikh Shuvo: Now, when you look at your background, it very much seems like you're equal parts researcher and educator. What really motivated you to initially put so much great content out there?

Youssef Hosni: Yeah, so I think, when I started, I was a bit stuck and didn't know how to proceed, especially when I tried to shift from academia to industry. There were a lot of things I wished someone had told me before I joined academia, because I think it would have made everything easier when trying to go into the industry. And I think this also applies to many people who have just finished a bachelor's degree or their master's or PhD and are trying to move into industry.

So this was the sparking point for me. I just wanted to share what I have learned through my very short and simple journey.

Sheikh Shuvo: Oh, you're being humble.

Youssef Hosni: No, no, it's true. So, this was the point. I just wanted to share what I have learned and what helped me to do what I have done, and just make it easier for those people who are trying to proceed along a similar route.

Sheikh Shuvo: In talking about the types of shifts that occurred when you went from academia to industry, what are some of the things that you had to adjust to, and what are some of the ways that you would recommend that universities adjust academic training, if at all?

Youssef Hosni: Yeah, so actually, I think, especially in programs related to data science and AI and so on, people need to focus more on software engineering skills. I think these are the main skills that people lack when they are trying to shift. A lot of database work, working with query languages, knowing how to write production-level code, and dealing a bit with cloud services and so on.

All of the stuff related to software engineering. I think people focus more on the science, the properties, the statistics, the machine learning, the modeling part, how to deal with data, which is, of course, all important. But learning how to work on something that will be put into production in a certain environment is really, really important.

And I think this is one of the main reasons people are not able to join a startup or a company, because they're usually looking for someone who can work in a software engineering environment and put the models and stuff into production. From my experience, and also from what I know from friends, people join a certain degree program but after finishing, they find there are a lot of skills they lack and gaps they need to fill, which they then have to spend more time learning. This should be taught in the academy or in master's programs.

Sheikh Shuvo: What about when you're doing the reverse of that and say you're starting in traditional software engineering and you're looking to make a shift into becoming a machine learning engineer? There are different AI courses and bootcamps out there, but do you have a recommended starting point for someone like that?

Youssef Hosni: Yeah, I think, to be honest, it's easier. So usually, when people who want to pursue a career in AI ask me, especially those trying to choose a major, I recommend choosing software engineering. Then, it will be easier to transition to working as a machine learning engineer or in data science.

For starting points, I usually recommend Coursera because they balance practical learning and theoretical knowledge well. The theoretical part is not shallow; it has good depth. The machine learning specialization by Andrew Ng is a very good start. He's been doing very good work in short courses on generative AI. For example, he recently published the Generative AI Specialization, which I think is also a good starting point if you're interested in generative AI. So Coursera is a very good starting point for everyone.

Sheikh Shuvo: Now, you also work extensively with mentoring students and aspiring data scientists. When people are just starting out, do you see any common pitfalls?

Youssef Hosni: Yes. The first one is that they focus more on the machine learning part. People usually start with machine learning, which is understandable. But what I see is that they tend to not focus first on software engineering skills. I agree that focusing on this part is really important. Also, dealing with data, like how to clean your data, do feature engineering, preprocess, and so on.

Another issue is not doing good enough projects. When I review resumes from people trying to land a job in the market, especially fresh graduates who don't have much experience, they put projects in their resumes. Usually, these projects are very common and have been done by almost everyone when they start. But these don't provide the uniqueness that you need when you're applying for a job. People want to see that you have done some good, unique projects. Don’t just stop at these very famous projects, which almost all the people starting in this field have done, like the projects in Coursera courses. Focus on building your own project. Try to find an idea that you're interested in and start working on projects in that area. Focus not only on machine learning, which is of course important, but also on how to develop a project or a product and the skills needed to do so.

Sheikh Shuvo: Moving to the research side of things, is there an area of research, either in computer vision or more broadly, that you think is very valuable right now but not enough people are talking about?

Youssef Hosni: Yeah, I think one area is efficient machine learning, or like one-shot learning for machine learning and deep learning. Most people currently are focusing on increasing the parameters to make models more capable. But I think the trend in the future might be evolving models with fewer parameters. We have seen Microsoft recently publish "The Textbook You Need" and start decreasing the number of parameters for their models. Also, for specific applications in a wide area, you need to run your models on each device, especially for cars and so on. There is some research in this area, but I think there's not a lot of hype around it, even though it’s really important.

Sheikh Shuvo: For efficient machine learning, are there any particular researchers or companies that you think are doing particularly interesting work there?

Youssef Hosni: I think companies involved in self-driving cars, like Tesla, and also Valeo, as I mentioned earlier, are doing great work there. I worked as an intern at Valeo. I was working close to the AI team in this part. So, self-driving car companies will be doing a lot of research in this area, I believe, to make their vision algorithms capable of working with natural flaws and doing the processing and so on.

Sheikh Shuvo: Nice. I used to work at Valeo as a vendor with their team in Ireland on a lot of the training data that went into the fisheye camera lenses there. The automotive world often takes the best of what's possible and tries to make it as small as possible.

Youssef Hosni: Yeah.

Sheikh Shuvo: Cool. You mentioned generative AI earlier, and I'm wondering, as various large language models proliferate, like the list of the best models on Hugging Face evaluation, what do you think are the criteria that companies and practitioners should be using to select the right model?

Youssef Hosni: Yeah, there are a lot of factors. One of them is the performance of the model on the task it will be used for. Another is the capability of pre-training this model; whether it is open source and can be pre-trained on their own data. Also, the technical requirements to run the model are important, because we know that something like OpenAI consumes a lot of resources, and not all companies are capable of this. Also, the availability of data to perform these tasks – if they need to get the model and do some pre-training on it, do they have the data to do so?

Sheikh Shuvo: Yeah. You mentioned OpenAI, which definitely stole the attention of the world this week with their Dev Day. Were there any particular announcements there that you were most excited about, if any?

Youssef Hosni: Yes, the one that caught my attention is the GPTs, where people can now create their own GPT and just publish it for the community, and there will be a GPT store. I think this will somehow close a lot of startups, especially those that have recently started based on the OpenAI API and provide some AI service. People can now just buy this GPT and use it for their own problems. Also, I read that they can use GPT on their internal data. I think this will decrease the need for LLM engineers and AI engineers because you don't have to hire a lot of engineers to do the fine-tuning for you. So it's really interesting. It’s promising, of course, but it's also a bit of a danger, becoming more and more prevalent.

Sheikh Shuvo: Yeah, especially with the built-in reach, it automatically puts a lot of companies out of business right there.

Youssef Hosni: Yeah.

Sheikh Shuvo: When will the use of GPT be coming out?

Youssef Hosni: Yeah, I'm not sure about when the use of GPT will come out. Let's see if I have an idea about that. Most of my friends have just started to publish their own GPTs, which is really interesting.

Sheikh Shuvo: Amazing. Thinking about the tools and resources you use both as a researcher and as an educator, are there any AI tools that you use in your own workflows, like Copilot, for example?

Youssef Hosni: Of course, I use Copilot. Also, ChatGPT, I use it a lot. It has a lot of capabilities; every time I use it, I discover something new. It's like chatting with a friend, but one who has a bit more knowledge and can do things faster than me. I use it a lot and always find new ways to chat with it. I was using a lot of tools, but eventually, I found that ChatGPT is doing most of what I used the other tools for. So now, I mostly use only ChatGPT. Like at the beginning of this year, I was trying different tools, but in the end, I found I could use ChatGPT for all the things the other tools were doing for me.

Sheikh Shuvo: When you're using ChatGPT or Copilot, say, for writing software, how often are you using the code that's generated directly and accepting the suggestions?

Youssef Hosni: To be honest, I don't use it directly like that. What I do is I usually write code first and start using it in debugging because it helps me a lot. Sometimes it helps me understand error messages and find the error, or trying to understand what the problems are, or to improve my code. For example, sometimes I just try an initial version and give it to ChatGPT to help me improve it. But I only use the code generated by ChatGPT for minor things, because it's like an example of certain things instead of searching through documentation or the official library package. But usually, I use it for debugging. I think it's really useful and saves a lot of time when you use it this way.

Sheikh Shuvo: Does ChatGPT work equally well in English as well as Arabic when producing content?

Youssef Hosni: Actually, it's doing a very good job in Arabic. I haven't used it much for Arabic, because I generally don't write a lot in Arabic for articles and so on. But I used it a bit for translation and it was really good, way better than Google Translator. When you ask some questions in Arabic, its answers are really good. I asked different people in different languages and they also mentioned that it works really well in other languages.

Sheikh Shuvo: Awesome. The very last question I have for you is, with the pace of AI accelerating the way that it is, with so many new groundbreaking research papers and companies emerging every day, what are some of the ways that you use to stay on top of the latest and greatest?

Youssef Hosni: I follow Hugging Face's daily papers, which publish the good new important research papers almost daily. I also follow a lot of research institutes, labs, and researchers on Twitter to know what they are doing. I try doing small projects to use new models and fine-tune them, but on a very small scale, just to keep my hand always with the new advances. I also watch tutorials and videos on YouTube. There are a lot of channels that talk about the new important models and features, maybe in ChatGPT or new releases from ZeptoPremium and so on. And of course, I have my learning plan for long-term understanding, but in the short term, on a weekly basis, I just try to follow the new papers. I don’t read all of it, but if something catches my attention, I might read a bit or just see the Hugging Face and GitHub pages to stay in shape.

Sheikh Shuvo: Awesome. Well, thank you so much, Youssef, for joining and sharing about your world. Lots of super practical advice and exciting news. Hopefully, we'll talk to you again soon.

Youssef Hosni: Yeah, it was really nice talking to you. And I hope to talk once again.

Sheikh Shuvo: Thanks, Youssef. Thanks.