Steps of an AI project

The idea

An AI project cycle goes in order: define the problem, collect data, test the AI tool, then reflect and improve, because each step depends on the one before.

Where do you start?

A class wants an AI tool that guesses how many plates the school canteen will need each day. What should they do first?

• Collect as much data as they can find
• Decide exactly what the tool must predict
• Try the tool out on the canteen staff

Build the project in order

Anil's class is building that canteen tool. Tap the four steps in the order they must happen. Each right tap tells you why it belongs there.

Each step feeds the next

You could not skip ahead. The goal told you which data to collect. The data gave the tool something to learn from. The test showed whether it worked.

The last step is not an ending. After you improve the tool, you test it again, so the steps go round in a cycle.

An AI project cycle goes in order: define the problem, collect data, test the tool, then reflect and improve.

Your textbook splits this into six stages: define the problem, collect and prepare data, develop and train, evaluate and refine, deploy, and monitor. Here, training sits inside the test step, and launching and watching come after.

Notes

An AI project cycle goes in order: define the problem, collect data, test the tool, then reflect and improve.

Think of cooking: decide the dish, gather the ingredients, taste it, then adjust. Skip a step and the result suffers.

Check yourself

Farhan's team collects thousands of plant photos first. Only then do they ask what the tool should find. What is the likely problem?

Tenzin's train-delay tool keeps guessing wrong on rainy days. He studies those mistakes and decides it needs rainy-day records. Which step is he doing?

Tenzin has improved his tool. What is the best next move?

• The tool will pick a goal that fits the photos. A tool cannot choose a goal. People decide the problem, and they must do it before anything else.
• Many photos may not fit the goal, so effort is wasted — correct. Without a clear problem, nobody knows which photos matter. That is why defining the problem comes first.
• Thousands of photos are too many for any tool. Plenty of data is not the trouble. The trouble is that nobody can say whether it matches the goal.
• Define the problem again. The problem, predicting train delays, is already clear. Tenzin is working on a tool that exists.
• Collect the rainy-day data. He will collect rainy-day records soon. But right now he is studying the mistakes to decide what is missing.
• Test the tool on new days. He has already tested it. That is how he found the wrong guesses on rainy days.
• Reflect and improve — correct. He looks at the mistakes, finds the likely cause and plans a fix. Then he will test again.
• Test it again on new cases to check the change — correct. A change can fix one mistake and cause another. Only a new test shows whether the tool really got better.
• Launch it at once, because it has been improved. Improving is not a finish line. The cycle goes round again, starting with another test.
• Start again by defining a brand new problem. The goal usually stays the same. You redefine the problem only if the goal itself turns out to be wrong.