Challenges of Machine Learning | AI Fundamentals Course | 2.4

By Tech Gee on May 27, 2025

So, you’ve started diving into the world of machine learning (ML), and everything seems kind of… magical. You feed a computer a bunch of data, and boom…it can detect spam, predict the weather, or recommend your next favorite movie.

But here’s the reality check: machine learning isn’t magic…it’s math, data, and a whole lot of problem-solving. And like any powerful tool, it comes with its own set of challenges. Three of the most common (and most important) ones you’ll come across are:

Overfitting
Underfitting
Data Bias

These issues can make or break your ML model. In fact, even if you’re using the fanciest algorithms and state-of-the-art hardware, your model will fall flat if you don’t understand and manage these challenges.

So, with that being said, let’s walk through each one together.

What an Machine Learning Model Actually Does

Before we dive into the problems, let’s remind ourselves what machine learning models actually do.

Imagine you’re trying to teach a machine how to predict house prices. You give it a bunch of examples: square footage, number of bedrooms, location, and what those houses sold for. The goal is for the model to find patterns and learn the relationship between features (input) and prices (output).

But here’s the kicker: the model doesn’t really “understand” houses or money. It just crunches numbers and finds patterns.

Sometimes it finds patterns too well. Other times it doesn’t find any meaningful pattern at all. That’s where overfitting and underfitting come into play.

Overfitting

What is Overfitting

Overfitting happens when your machine learning model learns the training data too well. Like, obsessively well.

It memorizes every tiny detail even the random noise or errors in the data. It becomes a straight-A student on the training data, but completely bombs the test (new, unseen data).

Imagine this: you’re studying for a history test by memorizing every single word from the textbook. Then on the test, the questions are phrased differently, and you have no clue what to do. That’s overfitting in a nutshell.

Signs Your Model is Overfitting

High accuracy on training data
Terrible performance on test or validation data
Very complex models (too many parameters or too deep a neural network)

A Real-Life Example

Let’s say you’re building a model to identify spam emails. You train it on thousands of messages. But your dataset includes a weird pattern, some spam messages are always sent at 3:00 AM. Your model learns: “If it’s 3:00 AM, it must be spam.”

Problem? In the real world, lots of legit emails are sent at 3:00 AM. You’ve trained your model on noise, not real patterns.

How to Avoid Overfitting

Use More Data: A larger, more diverse dataset can dilute the noise.
Regularization: Techniques like L1 / L2 regularization help prevent the model from going overboard.
Cross-Validation: Train your model on different subsets of the data to ensure generalization.
Early Stopping: Stop training when performance on validation data stops improving.
Prune the Model: Reduce complexity…fewer layers or features.

Underfitting

What is Underfitting?

Underfitting is the opposite of overfitting. It’s when your model is too simple to learn the underlying patterns in the data. It’s like trying to use a single crayon to draw a realistic portrait. No matter how hard you try, you just can’t capture the details.

Your model performs poorly on both training and test data because it hasn’t learned enough.

Signs of Underfitting

Low accuracy on both training and testing data
Very simple models (linear models when relationships are nonlinear)
Model can’t capture the trends or patterns

A Real-Life Example

Say you’re trying to predict house prices again, but you build a model that only looks at square footage and ignores all other variables (like location, number of bathrooms, etc.). The model might draw a straight line through the data when you actually need a curve to reflect market trends.

That’s underfitting. Your model’s just too basic to get the job done.

How to Fix Underfitting

Use a More Complex Model: Try decision trees, ensemble methods, or deep learning.
Add More Relevant Features: Include other useful data like neighborhood crime rates, school ratings, etc.
Train Longer: Sometimes underfitting happens when the model hasn’t had enough time to learn.
Reduce Bias: Make sure your model isn’t overly constrained by assumptions (like linearity).

The Balance: Bias-Variance Tradeoff

At the heart of overfitting and underfitting is something called the bias-variance tradeoff.

High bias = underfitting (the model is too simplistic)
High variance = overfitting (the model is too sensitive to noise)

The goal is to find a sweet spot with low bias and low variance, which leads to good generalization. Think of it like a tightrope walk…too far in either direction, and your model topples.

Data Bias

What is Data Bias?

Data bias happens when the data used to train your machine learning model isn’t representative of the real world. It could be:

Skewed
Incomplete
Discriminatory
Outdated

This means your model will make predictions that are inaccurate, unfair, or even harmful. Bias in data leads to bias in decisions. And when ML is used in areas like hiring, healthcare, or criminal justice, the consequences can be serious.

Types of Data Bias

1. Sampling Bias

When your training data doesn’t reflect the diversity of the actual population.

Ex: If you train a facial recognition model on mostly white male faces, it will struggle to identify women and people of color.

2. Label Bias

When the labels used for training are flawed or subjective.

Ex: In hiring, resumes from women might have been historically rated lower, which the model then learns as a pattern.

3. Measurement Bias

When features are measured inaccurately or inconsistently.

Ex: Using zip codes as a proxy for income can be misleading and discriminatory.

4. Confirmation Bias

When the data confirms a preconceived notion or hypothesis and ignores contradictory evidence.

Ex: Only including data that supports your prediction that older people don’t use smartphones when many do.

Why Does Data Bias Matter?

Because machine learning doesn’t have morals. It just mirrors the data we give it.

So, if your data is biased, your AI will make biased decisions. It doesn’t mean the algorithm is evil, it means the input is flawed.

Remember: garbage in, garbage out.

Real-World Examples of Bias in AI

1. Amazon’s Hiring Tool

Amazon built an AI to help with hiring. But it was trained on resumes submitted over 10 years, mostly from men. The AI began penalizing resumes that included the word “women” (like “women’s chess club captain”).

2. Facial Recognition Fails

Studies have shown that facial recognition systems have higher error rates for people of color, especially black women. Why? Because the training data wasn’t diverse enough.

3. Predictive Policing

Some law enforcement tools use crime data to predict where crimes might happen. But if that data reflects historical over-policing in certain neighborhoods, the AI ends up reinforcing those patterns.

How to Detect & Mitigate Bias

Audit Your Data: Who’s represented? Who’s missing?
Diversify Your Training Set: Include a wide range of ages, races, genders, & other groups.
Test Fairness Metrics: Evaluate model outcomes across different demographic groups.
Involve Domain Experts: Ethicists, sociologists, and community leaders can provide insights algorithms can’t.

Be Transparent: Share your data sources & model assumptions.

Quick Recap

Challenge	Description	What It Looks Like	How to Fix It
Overfitting	Model is too complex, memorizes training data	High training accuracy, poor testing accuracy	Simplify model, regularization, early stop
Underfitting	Model is too simple, can’t capture patterns	Poor training and testing accuracy	Use better features, more complex models
Data Bias	Training data isn’t fair or representative	Unfair predictions or discrimination	Audit data, test fairness, diversify input

So, What Should You Take Away From This?

Machine learning isn’t just about building cool models, it’s about building reliable, fair, and useful ones. And that means understanding:

Overfitting isn’t a flex…it’s a flaw.
Underfitting means your model is basically guessing.
Data bias is the hidden enemy of fairness and accuracy.

Getting your model to work in a lab is one thing. Making it work in the real world, where things are messy, diverse, and unpredictable? That’s where the real skill lies.

Final Thoughts

Here’s the deal, building ML models is exciting, but it comes with big responsibility. Your models can influence people’s lives, reinforce or break stereotypes, and shape decisions.

So don’t just aim for accuracy. Aim for awareness.