Day 1 (45m)
A.I. Party Tricks (5m)
Let's start with a friendly art competition:
Hint: Start drawing with your mouse and click the suggestions that appear at the top.
Notify your cohort when you finish. First place prize. Last place prize.
Quickly share with your cohort your intuition/guesses at how those web apps learned to draw.
A.I. Intro (10m)
Automation is often thought of as a replacement for highly repetitive mundane or dangerous tasks. It's comforting to think of a robot endlessly stuffing envelopes or dutifully vacuuming your carpet.
Artificial Intelligence (A.I.) is the evolution of automation to critical thinking tasks. Automation is done getting coffee and pushing papers. It has its eyes on the corner office and wants to make the decisions.
This course focuses on a specific type of A.I. known as machine learning. Machine learning is the practice of transforming data into decisions. Instead of programming a system with an ever-growing list of explicit rules, machine learning leverages data to create an evolving set of features, parameters, solutions, and policies that manifest themselves as intelligent behavior.
Imagine trying to program a system how to play with a dog. You could try to come up with a series of rules. If the dog is asleep, don't wake it. If the dog's tail is wagging, throw a ball. If the ball is not returned, throw a stick. If the dog barks, say 'good dog' ... Exhausting and futile, no?
Machine learning would feed an algorithm video and audio from thousands of hours of interaction of owners playing with pets. The algorithm would sort the data into groupings based on features and fine-tune parameters that can pick up on cues from the dog and optimize its behavior.
This "ad-hoc" sort of behavior can convincingly emulate "human" characteristics. They can create art. Try bold strategies in competitions. Invent scientific theories.
Setup
You're going to create your own A.I. utilization plan to harness A.I. for the good of an organization you participate in (e.g., work, community).
Start and share your own A.I. Utilization Plan
Click the image below to open the template / Sign into your google account / File / Make a Copy / Replace "Copy Of" with your First Name and Last Initial and remove "Template" / My Drive / Shared With Me / "A.I. Utilization Plans" / OK
Throughput (10m)
Before we get into the weeds of machine learning, consider why you would use it. Every organization has an overarching purpose that can be expressed as a rate of throughput. A business measures how quickly it can produce something it can sell (e.g., refrigerators/hour). A charity could measure the rate at which it can deliver meals to people in need (meals/day). Remember the throughput should be for your organization as a whole, not a sub-system (e.g., not frames welded/hour but ordered cars finished/day).
Please pick an organization and fill out the "Prepared for:", "Prepared by:" & "Throughput is..." fields.
Bottlenecks and the Theory of Constraints
A bottleneck is the one or two steps in your process that can't keep up with the rest of your organization. Your throughput can never be faster than your bottleneck because it limits the rate work can flow through your organization. Any delays in a bottleneck cascade down the dependent events in your process. Simply put, a bottleneck's capacity is less than or equal to the demand on it. Consider where work is typically pilling up, what is not available when you need it, and what seems to always increase the time it takes to complete the end product of your organization.
Please fill out the "Our bottleneck is..." field.
A bottleneck is more formally known as a constraint. Constraints should be the focus on any automation effort or A.I. implementation because creating automation to save time/resources on any other step would just be "activating" it. A.I. targeted at helping alleviate a bottleneck would be genuinely "utilizing" A.I. because it improves the throughput of your organization. A formal way to handle constraints is prescribed by Eli Goldratt's Theory of Constraints:
-
Identify system constraint(s)
-
Decide how to exploit constraint(s)
-
Subordinate everything to decision(s)
-
Elevate constraint(s)
-
Go back to step 1
A theoretical example from a pomegranate juice maker:
-
Separating the seeds from the pomegranate is the constraint in producing bottles of pomegranate juice.
-
Note other workers and spoiled seeds being found later on in the process.
-
Communicate to all employees the top priority of and plans for improving the "seeding" step. Make every possible resource available to help.
-
Pull extra workers from other steps in the process to help manually peel the pomegranates. Inspect the fruit before seeding and throw it out instead of wasting time on a spoiled fruit. Design a new process that "seeds" the pomegranate in 25% less time.
-
Notice the new bottleneck is now the bottle filling step.
Take 3 minutes and share your "Throughput" and "Bottleneck" fields with someone in your cohort.
Data and Features (5m)
Machine learning requires multiple sets of (good) data. The multiple sets are needed because you'll use the majority of your data to "train" your machine learning system, and a smaller set to "validate/test" your machine learning system.
What makes for (good) data? Accurate (garbage in = garbage out). Clearly segregated and relevant "features" that the system can use for prediction (e.g., height, weight, wingspan, 40-yard dash, vertical jump, bench press, etc. could all be "features" to a machine learning system to predict success at a given sport). The data must be pre-processed into something quantifiable (e.g., "9/10" instead of "fantastic") so it can be interpreted.
A.I. Fit (10m)
Keeping in mind the Theory of Constraints and the data requirements for machine learning, you can now pick a realistic task for A.I.
Please fill out the "The task for A.I. to perform is..." field.
The capabilities of A.I. warrant carefully considering and accounting for its place in your organization as you would any human employee. It will depend on help from others and others will depend on it. To facilitate this, it's useful to personify the machine learning system.
Please fill out the "A.I Profile" section.
A.I. Rewards (5m)
While machine learning produces useful artificial intelligence, you don't have to worry about the A.I.'s morale. You do however have to worry about the morale of your fellow humans. If members of your organization view the A.I. as a job, money, and/or power grab, they may become hostile to the automation. It will then be hard to get the data the A.I. needs, it is unlikely to be maintained, and its work is likely to be discredited. To achieve buy-in, the planned use for the value generated by the A.I. should be committed to and communicated up front.
A useful way to frame these plans is to outline how the A.I. will be considered a public good within your organization. By definition, public goods cannot force someone to pay to use them (e.g., firefighters put out all fires, indiscriminate of who pays taxes). Also, their use by any one person doesn't prevent another from using the public good (e.g., breathing clean air). Machine learning algorithms have the potential to create a virtuous cycle where not only does one person's use not inhibit another's, but it actually improves the next person's utility from the A.I. because you've generated additional data to teach the system.
Please fill out the "A.I. Integration" section.
You did it!
In the next section you'll get to pick a learning stategy for your new A.I. peer.