How can you train a machine learning system?
Geometric artificial brain with chip, gears and circuit board (miakievy, Getty Images)
Geometric artificial brain with chip, gears and circuit board (miakievy, Getty Images)
How does this align with my curriculum?
Create a set of rules that a machine learning system could use to identify a specific icon from an icon library.
Consider using the Design & Build Process with this challenge.
This activity will help build skills related to the Plan, Create, Test & Evaluate, and Reflect & Share phases of this process. It also highlights the iteration required during this process.
Materials:
- Icon Library (see below)
- Something to write with
- A partner
What to do!
Your job is to develop rules to train a machine learning system to recognize icons. In this activity, your partner will pretend to be the machine learning program. We have provided you with a library of icons to use.
- Plan – Choose one of the icons from the library. Do not let your partner know which icon you chose. Look at the icons and think of different ways you could describe and sort them. This will help you when creating the rules that your partner will use to find the icon you chose. For this activity,
- A rule cannot use the name of the thing (e.g., “It has four legs” and not “It is a cat”).
- Each rule can only involve one characteristic (e.g., “It is alive” and not “It is alive and has legs”).
- Create - Write down the rules that your partner (the computer) will use to find your icon.
- Test - When you think you have a good set of rules, read them to your partner. See if they can figure out which icon you picked from the icon library.
- Evaluate – Was your partner able to identify your icon? If not, revise your rules and try again.
- Reflect & Share – What types of characteristics did you use for your rules? What rules helped your partner the most? Were any rules not useful? What are the fewest number of rules that could identify your icon? Would your rules have been different if the icons had been in colour?
Image - Text Version
Shown is a grid with black and white line drawings in each square. The grid is five squares across and five down. Each square contains a simplified line drawing of an object. From left to right, starting at the top, these are: the Sun, a bicycle, an umbrella, an airplane, a basketball, a shoe, an apple, a music note, a tooth, a flower in a pot, a cloud, a dog, a fish in a bowl, a bird, a squirrel, a ball of yarn, a cat, a bone, a paw print, a comb, a pair of scissors, a moon, a star, and a tree.
Discovery
A machine learning model is a program that can find patterns and make decisions based on a set of data. Machine learning is a form of artificial intelligence.
In a supervised machine learning model, people need to help the model figure out a pattern to correctly identify something. First the model is given labelled pictures of something, like an object. These pictures and their labels are called input data. The model uses the input data to come up with rules. Then it uses them to get the right output data. The output data is the identity of the object. Machine learning models need lots of input data to come up with accurate rules.
In an unsupervised machine learning model, the model is given unlabelled data. Then it is asked to find patterns. A human can then use those patterns to build a model that does the task needed. Sometimes machine learning models find unexpected patterns in data. This is why humans need to be in the loop. They help build models that work the way we want them to.
It’s important to know how and why a machine learning model comes up with its output data. If the input data is flawed or incomplete, the output data could be incorrect, biased, or unethical.
Computers do not “think” the same way as people do. A task that humans think is easy can be very difficult for a computer, and vice versa!
- Try your rules again with other icons. Do they still work? If not, what needs to be changed or added?
- Can you come up with rules that only use shapes to identify the icon?
- Read your rules to someone, without showing them the Icon Library. Can they tell what object you’re trying to define? Can either of you think of other objects that would also fit your rules?
What’s happening?
A machine learning model is a program that can find patterns and make decisions based on a set of data. Machine learning is a form of artificial intelligence.
In a supervised machine learning model, people need to help the model figure out a pattern to correctly identify something. First the model is given labelled pictures of something, like an object. These pictures and their labels are called input data. The model uses the input data to come up with rules. Then it uses them to get the right output data. The output data is the identity of the object. Machine learning models need lots of input data to come up with accurate rules.
In an unsupervised machine learning model, the model is given unlabelled data. Then it is asked to find patterns. A human can then use those patterns to build a model that does the task needed. Sometimes machine learning models find unexpected patterns in data. This is why humans need to be in the loop. They help build models that work the way we want them to.
Why does it matter?
It’s important to know how and why a machine learning model comes up with its output data. If the input data is flawed or incomplete, the output data could be incorrect, biased, or unethical.
Computers do not “think” the same way as people do. A task that humans think is easy can be very difficult for a computer, and vice versa!
Investigate further!
- Try your rules again with other icons. Do they still work? If not, what needs to be changed or added?
- Can you come up with rules that only use shapes to identify the icon?
- Read your rules to someone, without showing them the Icon Library. Can they tell what object you’re trying to define? Can either of you think of other objects that would also fit your rules?