Trying Out Machine Learning

[Guest mjschanne]

 

 

 

In this post, I share my recent experience with trying out machine learning including resources I used, my takeaways, and where I would want to go next with learning about ML.

 

Approaching ML

 

 

 

 

Recently, my team and I got started with exploring machine learning. We started with reading this awesome introductory blog post, ”Machine Learning for Everyone.” This served as a map of what kind of capabilities exist under the umbrella of machine learning and helped clear up some of the mysticism that often surrounds terms like machine learning and AI.

 

Our interest was specifically ways we could apply neural networks to healthcare as a follow up to our project, Chestist, a model to detect anomalies in chest X-rays. We therefore moved on to looking at another blog post, “Exploring MNIST Dataset using PyTorch to Train an MLP.” This gave a great little tutorial on building a simple implementation of a type of neural network called a multilayer perceptron.

 

 

 

To better understand what was happening behind the scenes of our code, we also played with a tool embedded in this blog post, which demonstrates how image kernels work. This was a great way to understand what we would be doing next with convolutions in a convolutional neural network (CNN) which utilizes image kernels as part of downsampling before identifying and labeling the features of the image.

 

 

 

With that under our belt, we were ready to take a crack at a more sophisticated implementation of a convolutional neural network by following this tutorial.

 

 

 

I have taken both tutorials that we went through as a team and created Jupyter notes for both that contains the results of following the tutorial as well as my annotations of the tutorial, but I highly recommend trying to complete them on your own first if you are also new to machine learning like I was. If you’re like me and are new to Juptyer Labs, you may find it helpful to watch this

.

 

 

 

What I Took From It

 

 

 

 

Based on my experience here, I would say that for beginners,

 

• You don’t need a strong math basis to get started and try to do interesting things - if you already have a programming background. If you have strong math/data analytics skills the coding to do it is pretty straightforward and only a few tutorials can get you where you need to be
  
• Just like learning code, it is helpful to annotate with comments what you’re doing and why you’re doing it as you go
  
• The actual complexity of machine learning is not necessarily in the execution of creating a model but instead in choosing the right ML technique for the right situation – that is also where having a strong math background can be helpful to understand the underlying mechanics of techniques so that you can quickly grasp their strengths and weaknesses
  

 

 

 

When thinking about how much math is involved with machine learning, I would say that for fun projects. You don’t need any kind of math. But the higher the stakes of the project, the more depth of understanding of the relevant math you can expect to need, and it scales up fast. Most documentation for methods related to ML seem to present everything immediately in formula.

 

Let me know if you agree or disagree or have any recommendations for beginners with machine learning in the comments section or join the Health & Life Science Dev Discord server to let us know there. All our content across YouTube, GitHub, and the HLS Blog will be linked there.

 

Also check out my colleague @jkarasha's brief article on image analysis using Neural Networks. Joe organized our introduction to machine learning. Or check out my teammate, @Anshika Goyal's article on using GitHub codespaces to learn ML. She was learning alongside me and chose to use GitHub Codespaces vs my own decision to run things locally so that might be helpful to you in choosing based on your preferences.

 

Continue reading...