Cracking the AI/Machine Learning Code: Essential Insights for Beginners (Part 2)

Hey there, AI/ML enthusiasts! Welcome back to our series on mastering machine learning. In my first post, we dived into building up your ML skills. Now, let's talk about how you can leverage Large Language Models (LLMs) like Claude, Gemini, or ChatGPT to help you build and run machine learning code. Spoiler alert: It's a game-changer for beginners!

Confession Time: I'm Not a Coding Whiz

Let's get real for a second. I'm not the best coder out there. In fact, I kind of hate the brain power required to master the syntax of languages like NodeJS or Python. If I had to choose, I'd rather be great at understanding the principles and applications of machine learning than be a Python guru.

This feeling was totally validated when I watched a talk called: "Large Language Models and the End of Programming". It's a fascinating look at how LLMs might take over a big chunk of basic programming in the near future.

Harnessing LLMs for Machine Learning

Some of you might already be familiar with this concept. Maybe you're using CoPilot in Visual Studio Code. But what if you're just starting out and want to focus on machine learning concepts rather than getting bogged down in code syntax? Here's how you can use LLMs to your advantage:

1. Be Specific In Your Ask

When you're asking an LLM to write code for you, specificity is key. For machine learning problems, you need to nail down three core things:

• The problem you're solving

- What type of problem is it? Classification, regression, or something else?

- Where's your data?

- What format is it in?

• How you want to solve this problem

- Do you have a specific algorithm in mind?

- Any hyperparameter optimization needed? If so, which parameters?

- Time constraints for running the code?

- Known data issues, like missing values?

• How you want to visualize or evaluate the solution

- What metrics will you use to evaluate your solution?

- How do you want to visualize these metrics?

Example Prompt

Here's a basic prompt I used for a recent machine learning problem:

Create python code that runs on a Jupyter notebook. Import the dataset from a csv file in the downloads directory of a mac book.

This code should create a classification model for a dataset we have.

The dataset has a label called priceDecrease. It also has 10 features.

Create a confusion matrix, and feature importance to show us the most important features.

Do bayesian hyperparameter optimization, because we don't know what the optimal parameters are.

One of the variables is missing 65% of the observations however, we think this variable is important so leave it in if you can.        

Use LLMs To Ask What You Don't Know

When you're new to machine learning, there's a lot you don't know. The easiest way to close that knowledge gap? Ask your favorite LLM! I added this to my prompt:

Ask clarifying questions before creating this code.        

The LLM responded with a thorough set of questions, helping me refine my request and learn more in the process. Here's what it responded:

 I have a few clarifying questions to ensure I provide the most accurate and useful solution:
1. Which classification algorithm would you like to use? (e.g., Random Forest, Logistic Regression, XGBoost)
2. What is the name of the CSV file containing your dataset?
3. Can you provide the names of the 10 features in your dataset?
4. Which feature has 65% missing observations? How would you like to handle these missing values? (e.g., imputation, using algorithms that handle missing values)
5. For the Bayesian Hyperparameter Optimization (HPO), do you have a preference for the number of iterations or the time limit for the optimization process?
6. Are there any specific metrics you'd like to focus on for model evaluation besides the confusion matrix? (e.g., accuracy, F1-score, ROC-AUC)
7. Do you have any preferences for visualizations besides the confusion matrix and feature importance plot?
8. Is the 'priceDecrease' label binary (0/1 or True/False), or does it have multiple classes?        

Don't Let Code Errors Stop You

Here's a pro tip: LLMs don't always write perfect code. You might get errors or warnings. Don't let this discourage you! Just copy and paste that error message back to the LLM and ask it to fix it.

For example, I got this error when running the Claude generated code in Jupyter notebook:

TypeError: XGBClassifier.fit() got an unexpected keyword argument 'n_points'        

I fed it back to the LLM, and it quickly provided a fix along with this explanation.

I apologize for the confusion. It seems the error is due to a mismatch between the XGBoost version you have installed and the parameters we're using. Let's modify the code to remove the n_points parameter, which isn't necessary for our purposes anyway.
Here's an updated version of the script that should work with most XGBoost versions:        

Yes, it took 5 rounds of back and forth to get all the errors fixed, but it worked!

Start Small and Grow

Starting your machine learning journey doesn't have to be intimidating. Begin with a basic ML problem you want to solve. Use the framework we discussed to craft natural language prompts for your LLM to write the machine learning code. If you've got the basics of machine learning down, you'll be well on your way to running workable code that tackles virtually any basic to intermediate machine learning task.

Remember, the goal here isn't to become a coding expert overnight. It's about leveraging tools like LLMs to focus on understanding and applying machine learning concepts. As you practice this approach, you'll find yourself getting more comfortable with both the ML principles and the code itself.

So, don't be afraid to start small, ask questions, and learn from the process. With LLMs as your coding guide, you're set for an exciting journey into the world of machine learning. Happy coding, and may your models always converge!

What's your experience with using LLMs for coding? Have you tried this approach? Share your thoughts in the comments below!