Myths and Jargon
What people get wrong.
Machine learning attracts plenty of hype, and quite a lot of nonsense. After this course, you'll be the calm voice in the room. Let's go through four myths that you'll hear often, and why each one is wrong.
- Lots of hype and nonsense
- Be the calm voice in the room
- Four common myths
Myth versus reality
Myth one. More data always fixes it. Reality. Not if you have high bias, or if the extra data is bad. Myth two. Deep learning is always better. Reality. On tables of data, boosting often wins, and it's cheaper. Myth three. A model is objective, because it's maths. Reality. It learns whatever bias exists in its data. Myth four. Higher accuracy means a better model. Reality. It depends on what mistakes cost.
Myth
- More data always fixes it
- Deep learning always wins
- Models are objective
- Higher accuracy is always better
Reality
- Not with high bias or bad data
- Boosting often wins on tables
- They learn their data's bias
- Depends on what mistakes cost
Words you will hear
Let's also decode some jargon. An epoch is one full pass through the training data. A batch is a small group of examples, handled together. Inference means making predictions. A feature store keeps features in one shared place. Ground truth means the correct answers. And a hyperparameter is a setting that you choose, and that isn't learned.
| Word | Plain meaning |
|---|---|
| Epoch | One pass through the data |
| Batch | A small group of examples |
| Inference | Making predictions |
| Feature store | A shared home for features |
| Ground truth | The correct answers |
| Hyperparameter | A setting you choose |
Summary
Now you can spot the hype, and speak the language. You've reached the final lesson. Let's talk about where you go from here.
Key takeaways- More data isn't always the answer
- Models copy their data
- Deployment is the beginning