How to break into machine learning technologies as a young professional
Aria Chang30:522020
Three years before an IEEE paper award, she struggled with a for loop
Three years before her reinforcement learning model for a smart grid won a best IEEE conference paper award and went into production on a real power grid, she struggled to understand a for loop. Aria Chang, a software and data engineer at MediaMath with a string of hackathon wins, retraces that climb for anyone who feels behind. What actually turns a beginner's classwork into a machine learning portfolio that recruiters notice?
Get a four-step plan for your first ML app and a hackathon route to a stronger portfolio
Aria Chang, a software and data engineer at MediaMath, retraces how she went from struggling with basic programming to award-winning reinforcement learning research. She argues that a slow start is no barrier to machine learning and gives a four-step plan for building a first ML application, plus hackathons and competitions as the way to build a visible portfolio.
“So it's important to understand your learning curve. If you're also like me, where you have to spend weeks, months, even years to even get started in a field that you don't have any previous experience with, do not give up because there will be light at the end of the tunnel.”
“But do keep in mind that finding the training data isn't the hard part. What's harder would be finding data you could use to test how it works after you have trained your model.”
What you will learn
Recognize that a slow start in computer science is normal and plan around your own learning curve.
Plan a first machine learning application in four steps: support system, team and users, training and test data, and deployment.
Frame a reinforcement learning problem by defining its state, actions and reward.
Spot data access limits early, such as privacy rules that block testing a medical model on real patients.
Build a visible ML portfolio through hackathons, Kaggle competitions and side projects.
Details
Level
Beginner
Format
Talk
Technical depth
Introductory technical
For
Student, Early career, Career changer
Length
30:52
Published
June 2020
Questions this talk answers
How can a complete beginner get started in machine learning?
Find the part of the field that speaks to you, take relevant classes and do side projects even without pay. Keep applying what you learn to new projects.
Who do you need on a team to build a machine learning application?
In her view at least four people: someone who knows the full app stack such as MEAN, a back end and database person, a front end developer, and someone who knows the ML libraries, usually in Python.
What is the hardest part of getting data for an ML project?
Training data is easy to find in open-source datasets; the harder part is finding real data to test the trained model. Her team's seizure prediction app could not be tested on patients because of HIPAA.
How do you find teammates for a hackathon?
Collegiate hackathons run team-finding channels where you share your LinkedIn profile, experience, project interests and the skills you need. She also suggests women-only hackathons and contacting organizers.
How can an associate data analyst move toward machine learning?
Work on more machine learning projects, especially if the day job involves little model training, and take training courses on the side to fill knowledge gaps.
About this talk
Are you a passionate engineer looking to break into machine learning and AI technologies? Are you looking to increase your profile visibility? Do you feel like a little guidance will ignite the flame of desire in you to conquer the tech world? Look no more, vote for me and I’ll show you the door to opportunities and help you become a better version of yourself.
Test data, not training data, is the hard part (61 s)
The thrill of a first hackathon win (68 s)
Speaker 1: Welcome everyone to this session. So I have really enjoyed the every presenters talks in the past day felt very inspired. And some of you reached out to me on LinkedIn with questions you have. So I will try to address all of them in the talk today. And do feel free to go to the chat. We'll go to the link at the top to put your questions there. Or something you want me to talk about. If I'm not covering it, you know, be easier for me to see the questions since I'm doing the notes on a separate screen. All right, so I assume everybody who's looking at this session right now, you all have different backgrounds.
Speaker 1: But you are here because you want to learn more about machine learning and how you could also get into the field, but doesn't matter if you're a completely new beginner or even if you're just new to coding in general or if you're proficient in programming and also familiar with some of the machine learning concepts, AI concepts, or if you have done some machine learning research and projects, but you want to create a more attractive, Speaker 1: portfolio, then you're at the right place. So I am here today to show you how I went from zero to where I am today so that you could also achieve your goal. So let's get it started.
Speaker 1: Within my last year, during my bachelor's at UC Davis, I was awarded for one of the top 50 MLH most inspirational hackers award and the paper based on a model I built using reinforcement learning for a Speaker 1: smart grid model was awarded one of the best IEEE conference paper awards. And this model is currently used in production at a real power grid. And also I won five hackathon awards in a row for the projects. My teammates and I built using applied machine learning technologies since the first hackathon I've ever attended. But can you imagine that just three years before this, I struggled to even Speaker 1: understand the most basic computer science to understand the most basic computer science concepts such as a for loop and what is even a function input and outputs. I had a very hard time figuring everything out and I was very disheartened.
Speaker 1: I tried to learn the data structures and algorithms in the lower division computer science classes, but I just felt like I was so far behind my peers. I felt I was lost. Like I was left to survive on my Speaker 1: own. But toward the end of the year, I was able to find an internship doing web development at a nonprofit human rights organization working on content management management system. Even though this internship wasn't related to anything I learned in school the previous year, I was still able to make an impact in tasks such as content management system, version migration, using file transfer protocol softwares. And I, this was a defining moment in my tech career because my peers and my manager, they gave me very good feedback.
They gave me the confidence I needed to continue pursuing the career and not give up. So it's important to understand your learning curve. If you're also like me, where you have to spend weeks, months, even years to even get started in a field that you don't have any previous experience with, do not give up because there will be light at the end of the tunnel. So let's dig in to see how I got started in my machine learning journey. When I took my first machine learning class after the summer internship, I learned about the major areas that fall under the big umbrella, artificial intelligence. So natural language processing is used in conversational AI or sentiment analysis.
Algorithms that are able to determine the sentiment of the text, classify it into anger, happiness, sadness, different sentiments. And there are great cloud computing platforms, which can help you integrate this technology into the absolute world. Predictive analysis. This is for things like, there are examples like stock and weather prediction, not saying how well current research is doing in predicting stock prices. But it's an area where you could use a historical time series data to attempt to find some pattern within it. And then there is computer vision, which is heavily used in autonomous vehicles.
So, you know, the self driving cars, their ability to determine if there are pedestrians or moving cars in the surroundings, or tell if an image is a cat or a dog. And then planning and optimization. This is, there are the examples would be real time strategy games, AI games like Starcraft, which I will talk about later in the later slides. And also such as AlphaGo, AlphaGo, who, which was able to defeat the best to go player in the world. And when I was learning all of this, I just fell head over heels in love with the probability proofs in the machine learning theory class, and even implementing a neural network using Python from the ground up. So the takeaway from this is, find the one thing that speaks to you. It might not be in the intrinsic beauty of writing math proofs,
like how I found it, but whatever it is, just find your passion, and don't let it go. So now we're going to talk about some action steps that you actually care about on how you could create your very own machine learning application. So the first step, you need to have a team. It might be possible that you want to create an application from the ground up, all on your own, and you don't want anyone else to be a collaborator in your GitHub repo, that's completely fine. But it's always good to have a team that you want to create an application from the ground up, all on your own, and you don't want anyone else to be a collaborator in your GitHub repo, that's completely fine. But it's always good to have a team that you want to create an application from the ground up,
that you want to create an application from the ground up, all on your own, and you don't want anyone else to be a collaborator in your GitHub repo, that's completely fine. But it's always good to have a support system. So in the case of creating a machine learning application, this would be a mentor, a professor, or some industry professional, someone who is a little more knowledgeable than you, who can provide you with some insights, or if they know more about the work that you're currently working on, the existing state of the art applications or research that you can learn something from, before you develop your application. And then the second step would be it would be to decide the type of application, if you decide to proceed with developing the application, that you would need a team of four, at least, in my opinion, you need someone who knows the entire app development stack, such as the mean stack, so that you could actually develop an application.
And then you could actually develop an application, which can be used by users. And then you need someone who is familiar with the backend server and database side of things. And then someone, another person who is a front end developer who knows the design aspects of things, and with front end programming languages, such as JavaScript, CSS, HTML. And then, most importantly, it is an ML application.
And you'll need someone who knows the machine learning libraries, the tools to use, or someone who is very good at learning them, who is passionate about them, which is very likely that that person will be you. So you need Python, usually it would be the language to go, and the passion to learn the technologies is very important in this, because there are always new tools coming out.
tools coming out, and you have the ability to learn if you have enough passion, and you want to put in invest in enough time into making it happen. tools coming out, and then you have the ability to learn more. And then you have to determine who are your intended targets, who will use this model after you develop the application. This is very crucial in evaluating the model performance.
So then after that, you will develop you, sorry, I'm on the third step that you have to find the training data. That's important. Because machine learning the model, you can't have one without data. There are a lot of open source databases on the on the web, where you can find in pretty much any machine learning tasks you can think of.
But do keep in mind that finding the training data isn't the hard part. What's harder would be finding data you could use to test how it works after you have trained your model. I'll give you an example. During one of the hackathons I attended, my teammates and I, we created an application to predict seizure for patients who experience epileptic seizures on a regular basis.
So we were able to create the application. The accuracy was very well in to predict the seizure on the training data that we used, which was open sourced. But we weren't able to find any patients who that we could actually test our model on because for the medical research, this concerns the privacy law privacy laws in the case of medical research, the HIPAA.
Compliance will have to be applied. So just keep this in the back of your mind when you conduct your in the first research period, what kind of data you will need and where you could possibly get your data after you create your model. So then the last step will just be develop and then deploy your trained model, decide on your platform, if it's the web application, the mobile application, where you want to host your services, where you want to host your services and you're going to use the cloud platform.
This is for an application. But if you decide to go into the research route, you'll have to do a lot of literature review of peer reviewed peer reviewed research papers so that you know what the existing state of the artwork is. So yes, let's move on to an actual example. I also took a an AI class during my undergrad career for during this class, my teammates and I, we created a reinforcement learning model for the real time strategy game called Starcraft.
Maybe some of you have played it before I have very addicted to it. So for this project, we were able to make some marginal improvement for the performance of the AI based on the best performance model we could find at the time. But later on, Alpha Star, as many of you may know, Alpha Star, which was created by DeepMind was able to defeat the best human Starcraft 2 player in the world.
So you see how powerful AI reinforcement learning can be. And looking at all of this, it gives me so much thrill to continue pursuing in the field of AI using machine learning technologies. So this is how I was able to get my summer internship in machine learning research at a smart grid institute. I was able to compare the two systems, the similarities there are between the two systems.
So let's talk about the basics of creating a reinforcement learning application. You have to define three things very clearly. In the case of the game AI for Starcraft, the stage will be whatever you can see on the screen when you play the game. The actions will be building, training new units, new, new, everything you can like for the structures and the upgrades, everything action or go scout your enemies camp and then attack them.
So all of these fall under the action variable. And the third thing will be the reward, the reward for the reward function in the Starcraft project. So we defined it as for every iteration of the game, if the AI we have is able to win the game, then the AI would be rewarded by accumulating the points. Otherwise, the points will be deducted from the system. And in the case of the smart grid, it was very similar. The part of the power grid would be like everything that defines a power grid system status will be the state.
And then the actions will be the actions that a human operator is able to to fit contingencies in the power system, such as if there's an explosion nearby blows up the blows up the buses and the within the power system, and you would be able to take some actions to change the settings in order to navigate around the contingency.
And the reward function will be defined as if the action you perform during this iteration is able to fix the contingencies in the power system. And this one achieved a very good result. So the lesson, the takeaway from this story is that you'll have to apply what you learned to new projects and keep taking on new challenges, never be afraid, and never be afraid.
You cannot do it and never be afraid that the projects are not important enough. If you just slap it onto your resume for all the meaningful projects you do that it'll add up. And chances are the recruiter when they look at your resume, they'll, they'll see that you have this experience and likely you are able to take on a similar challenge at work, even though you might lack industry experience previously.
And also it doesn't matter what industry you are in, you could always employ machine learning technologies into your domain. I've talked to several people who approach me on LinkedIn, I know that some of you are in the legal, legal domain and the healthcare domain. It is possible to use the extremely powerful machine learning technologies to make your system better.
You just have to have the willingness to learn. It might be very hard in the beginning, but the longer you work in this field, the easier it becomes. So how should work in your portfolio if you wish to be known for your machine learning expertise? My suggestion is to attend the hackathons. If you are a student, there are collegiate hackathons out there, very famous ones.
You can look them up. It's a very good way to give you confidence. I have a little anecdote to share with you. So after I took the machine learning and AI classes at college, I had this strong desire to just make cool applications at hackathons using what I had learned and also to win prizes. I remember how unbelievable I felt the first time at my first hackathon when my team's name was announced at the closing ceremony.
I couldn't believe when my team's name was called the first time or the second time in that same closing ceremony. It gave me so much adrenaline rush and the confidence I needed, even though it's bad to seek approval from the outside world. But when your values are being affirmed from an external organization, it just makes it official somehow.
It's like when you get an offer for an internship for a full time position, makes you feel good. So from hackathons, I also learned to stay up when I have to. Usually it would be staying up for two nights, three days straight, and then probably only sleep like 30 minutes each night. But it just felt so incredible when you actually create your application in the end, something that other users can use and benefit from it, especially if it's in an area that you're passionate about, such as benefiting for the social good or for people with health needs, anything like that.
Don't forget to ask questions. I don't see any questions. Maybe you're posting in the chat. I'll check in a little bit. So yes, or competitions. If you're interested, it's a great way to increase your profile visibility as well. There are Kaggle competitions. They're very famous and hard to beat, but it could be your short term goal. If you're motivated.
So keep applying what you learn to new projects always. And I will end my presentation with this, something which I hold dear to my heart. It doesn't matter how long it takes you to get to where you want, as long as you get there in the end, even if it might take you 10 years. Okay, so let me see if anyone had any questions for me.
Okay, I see quite a few questions. I see some of you are completely new beginners. If this talk isn't enough to answer all your questions, do feel free to, this is a QR code for my LinkedIn profile. So scan it and connect with me there. So one question I'm seeing is, okay, I'm currently working as an associate data analyst.
Could you help me how to elevate myself and the resources? How can I start? Firstly, congratulations. You already stepped into the door of data science. And I hope you are enjoying your work. So yeah, continue excelling at your work. And if you wish to advance in your career, it's important to work on more machine learning projects. If your work doesn't involve too much model training.
And if you think you need more knowledge, more knowledge in the domain, then you could take some training courses on the side. There are a lot of them that are really good. I'm not going to promote them in the video. But yeah, I'll talk to you after. And some other questions. How do you find teammates for the hackathon?
Very good question. So if it's a collegiate hackathon, I'm going to take it that probably Sharon, you don't have much experience with it. So for all the collegiate hackathons, the organization would create channels using specific softwares where you could look for teammates. There will be channels for like team teammates finding. So that's a good way where you can like put your LinkedIn profile and tell them what kind of experience you have and what kind of projects you're interested in creating and what skills you need, you're looking for in your team and potential teammates.
If it's an industry backed hackathon, then I'm sure they'll have, I haven't attended too many of those, but I have seen in some of the talks during the conference that some organizations host women only hackathons, which sound really incredible to me. So you should definitely look into those and always reach out to the organizers of the organizers of the hackathons if you have any questions.
And then can I talk a little bit about the tech stack that I've used for machine learning projects. So I have mostly worked on research projects for these ones. Because the except for the hackathon projects, then the research projects don't necessarily involve deploying the models using the using some of the cloud technologies, not going to name names.
So for the research projects, it's usually just finding the data, and then you do a lot of feature engineering, train your model for the supervised machine learning tasks, and then evaluate and evaluate, validate and test it during your training process. And then test against real data, and then continue improving your model. And don't forget to tune your hyperparameters.
Is there any other question? Any other resources for the rising sophomore? yes major league hacking is good to go about it if you're a college student or even a high school student they have some summer hackathons going on every weekend these days so do them virtually great way to make experience and other questions okay what about getting over the lack of experience i have industry experience but not a non-experience that is great if sharon if your industry experience is in a software-related field then you already have the very important skills because a lot of the machinery in the data science positions like data engineer machine engineer engineer you will have to code if you're a software program it makes everything so much easier so from the machinery expertise you could take courses online
and create projects on your own time find a platform to collaborate with your with teammates there are organizations i will share the resources if you send me a private message trying not to cover too much um reset could you elaborate a little on getting started in hackathons it looks difficult um i would say if you're new to hackathons just very important thing to keep in mind is do not be scared do not be scared to ask for help ask for the mentors that are in the hackathons um tell them you're a beginner ask for a beginner ask for advice and actively participate in all of the workshops that they have um if they will have workshops on how to create your first web application and what kind of tech stack
you should use um do not be afraid um even if your first application you try to you attempt to finish uh continue working on it or do not work out continue working on it even after the hackathon or during another one um and also um there would be teams who have more experienced engineers and they are looking for someone who could give them a different perspective even from a non-traditional computer science field it'll be okay okay so do not be scared that you are you lack experience or anything like that how do i start an internship if you're just beginning so hannah asked this question um it's important to yes uh feel free to go attend your next session i'm just going to try to finish the latin answering the
last few questions um as a beginner without previous experience how do you're gonna aim gonna gain and so like i said um if you're in school take the relative class relevant classes um do side projects um even it doesn't give you a even if it doesn't have any monetary gain it's always good to gain experience just like even software related experience will be great i wish i had more time otherwise i would be able to talk more about how the the other uh non-machine learning experience that have contributed a great deal um okay questions are around i will definitely uh talk to you more in pm add me on linkedin um what different algorithm do you suggest that a beginner's profile must contain um that's a good question uh for a beginner's profile i would say if you just get started on a basic
classification task um then you would be able to touch on the basic algorithms uh related to machine learning um so it's hard to answer it in one sentence but um feel free to chat with me after um yes how do you get in contact with me for more resources uh scan the qr code on the slide if you can still see it um i will definitely answer connect with all of you and answer all of your questions um i think that's pretty much it great i'm really happy that you guys attended my session i am uh really thrilled to be sharing my experience with all of you and i am an advocate for women's success um so i would love to contribute whatever i could if you just send me a message on linkedin i will
schedule a video talk with you if that's something you need um or however i could help with that i will leave the presentation i could figure out how um great thank you guys so much i was very nervous it was my first virtual presentation cool i hope you guys enjoy the rest of the uh the i almost was going to say hackathon the rest of the conference tomorrow i think it's networking day you can talk to a lot of the a lot of the recruiters ask them what skills they need uh what kind of experience you should aim toward getting if you are interested in the position in the data science machinery field great bye guys
Transcribed automatically; names checked against the speaker record. Select a time to jump there.
Article based on this talk
Why a slow start won't keep you out of machine learning
Aria Chang shows how classes, one internship and a run of hackathons took her from basic programming struggles to award-winning reinforcement learning work.
In short
Aria Chang believes a slow start does not rule anyone out of machine learning. Drawing on her own path, from struggling with for loops to an award-winning reinforcement learning model used on a real power grid, she gives young professionals a four-step plan for building a first ML application and argues that hackathons and competitions are the quickest way to build a portfolio recruiters take seriously.
Key takeaways
Recognize that a slow start in computer science is normal and plan around your own learning curve.
Plan a first machine learning application in four steps: support system, team and users, training and test data, and deployment.
Frame a reinforcement learning problem by defining its state, actions and reward.
Spot data access limits early, such as privacy rules that block testing a medical model on real patients.
Build a visible ML portfolio through hackathons, Kaggle competitions and side projects.
A slow start is not a verdict
Chang opens with what she achieved in her final undergraduate year at UC Davis: a place among MLH's top 50 most inspirational hackers, a best IEEE conference paper award for a reinforcement learning model of a smart grid that she says is now used in production on a real power grid, and, as she told the audience, five hackathon awards in a row. Then she turns the story around. Three years earlier she struggled to understand a for loop or what a function's inputs and outputs were, and in her lower-division data structures and algorithms classes she felt far behind her peers.
The turning point was modest. She found a web development internship at a nonprofit human rights organization, working on a content management system migration with file transfer tools. The work had little to do with her coursework, but the feedback from her peers and manager gave her the confidence to keep going.
"So it's important to understand your learning curve. If you're also like me, where you have to spend weeks, months, even years to even get started in a field that you don't have any previous experience with, do not give up because there will be light at the end of the tunnel." (3:44)
Find the part of AI that speaks to you
Her first machine learning class introduced the main areas under the AI umbrella: natural language processing for conversational AI and sentiment analysis, predictive analysis on historical time series such as stock or weather data, computer vision for self-driving cars, and planning and optimization, as in game systems like AlphaGo. What hooked Chang was the probability proofs of machine learning theory and implementing a neural network in Python from the ground up. She does not expect everyone to love proofs, only to find their own spark and hold on to it.
"So the takeaway from this is, find the one thing that speaks to you." (5:35)
Four steps to a first ML application
Chang's practical plan starts with a support system: a mentor, professor or industry professional who knows the current state of the art and can point you to existing applications and research. The second step is deciding what to build and who builds it. In her opinion a team needs at least four people: someone who knows a full app stack such as MEAN, someone for the back end server and database, a front end developer with design sense and JavaScript, CSS and HTML, and someone who knows the machine learning libraries, usually in Python, which she says is very likely to be you. The team must also define its intended users, which she calls crucial for evaluating model performance.
The third step is data. Open-source datasets exist for almost any task, but she warns that this is not where projects stall. At one hackathon her team built a seizure prediction app that performed well on open-source training data, yet they could not test it on real patients because medical data falls under HIPAA.
"But do keep in mind that finding the training data isn't the hard part. What's harder would be finding data you could use to test how it works after you have trained your model." (9:57)
The final step is to develop and deploy: choose web or mobile, decide where to host the service and which cloud platform to use. Those choosing the research route instead need a thorough literature review of peer-reviewed papers.
Reinforcement learning, from StarCraft to the smart grid
In an undergraduate AI class, Chang's team built a reinforcement learning agent for StarCraft and made a marginal improvement on the best model they could find. DeepMind's AlphaStar later beat the best human StarCraft II player, which only deepened her interest. That project helped her land a summer research internship at a smart grid institute, where she saw the parallels. Both problems need three clear definitions: the state (what is on screen, or the status of the power grid), the actions (building and attacking, or the steps an operator takes to work around a contingency such as an explosion knocking out buses) and the reward (points for winning, or for fixing the contingency).
Her lesson is to keep applying what you learn to new projects and to list them, because meaningful projects add up and signal to recruiters that you can handle similar challenges even without industry experience. She also stresses that the domain does not matter, mentioning people in legal and healthcare who had contacted her.
"And also it doesn't matter what industry you are in, you could always employ machine learning technologies into your domain." (16:08)
Hackathons, competitions and the long view
To build a portfolio known for machine learning, Chang recommends hackathons. She recalls disbelief when her team was called twice at the closing ceremony of her very first one, and the confidence that external recognition gave her. Kaggle competitions, which she calls famous and hard to beat, can serve as a short-term goal. In the Q&A she adds practical tips: use the team-finding channels that collegiate hackathons provide, look into women-only hackathons, try Major League Hacking's virtual weekend events, ask mentors for help, join the workshops, and remember that experienced teams often want a different perspective from a non-traditional background. For a beginner's profile, she suggests starting with a basic classification task.
"It doesn't matter how long it takes you to get to where you want, as long as you get there in the end, even if it might take you 10 years." (19:36)
Watch the talk
"How to break into machine learning technologies as a young professional" by Aria Chang (31 minutes): https://www.womentech.net/video/aria-chang-how-break-machine-learning-technologies-young-professional
About the speaker
Aria Chang is a Software/Data Engineer at MediaMath who won a Best IEEE Conference Paper Award for a smart grid AI using reinforcement learning that she built from scratch, and who has a winning streak in hackathons since her first in November 2018. She advocates for women's success and lives to empower women and be surrounded by women leaders from around the world and across all backgrounds.
Questions this talk answers
How can a complete beginner get started in machine learning?
Who do you need on a team to build a machine learning application?
What is the hardest part of getting data for an ML project?
How do you find teammates for a hackathon?
How can an associate data analyst move toward machine learning?
Drafted with AI from the talk transcript. Editor review pending.
Frameworks and ideas mentioned
Reinforcement learning (state, actions, reward)
MEAN stack
Resources mentioned
Major League Hacking (MLH) hackathons
Kaggle competitions
Collegiate hackathons
Women-only hackathons
Cite this talk
Aria Chang. "How to break into machine learning technologies as a young professional." WomenTech Network, 2020. https://www.womentech.net/video/aria-chang-how-break-machine-learning-technologies-young-professional