Common ML problems

Page 1

Three types of learning:

Supervised learning: Model is provided with data (features) for specific labels. The model will then be able to make predictions by linking unseen data to labels.

Unsupervised learning: Model is provided with data without any labels or hints, and it creates its own rules to categorize the data.

Reinforcement learning: No data or labels is given to the model (called agent in this case). It is given a specific task as well as a reward function that it receives every time it performs the task. It has to either be either a physical agent able to interact with the real world or be virtual in a virtual word.

Types of ML problems:

Classification: pick one of N labels depending on prediction.

Regression: predict numerical values.

Clustering: Unsupervised. Group similar examples.

Association rule learning: Unsupervised. Infer likely association patterns in data.

Structured output: Create complex output.

Ranking: Identify position of a scale or status.

The ML Mindset:

ML is different than traditional software engineering, where you can use mathematical models and logic to go from requirements to a workable model. In ML, you need to experiment.

ML interprets signals in a different way than humans which makes it very difficult to understand sometimes. It can also make mistakes, because of skewed training data or false interpretations, etc. Figuring out what the model is doing is difficult and very time consuming. It is difficult to know which features the model uses to get to its conclusions.

Experimental Design Primer:

Get comfortable with some uncertainty:

ML is different than traditional SW engineering as we don't have set parameters and don't really understand exactly how the code behaves. The non-coding work can be very complicated but there is less code to write.

Scientific Method:

We have to think of ML process as an experiment.

Set a research goal

Make a hypothesis

Collect data

Train the model and test the hypothesis

Analyze results

Reach a conclusion

Refine hypothesis and repeat

Page 2