From a discussion I lead on AI and machine learning for Product Managers. Based on the book "Prediction Machines" Sanji Agrawal et al, this talk is a non-technical intro to PMs who need to understand how AI can be utilized in products.
2. Recommended Reading
Great ML introduction
for non-technical
● Managers
● Business Owners
● Executives
Thanks, Darcy Norman, for the recommendation!!
3. What Machine Learning Is (and Isn’t)
From https://www.quora.com/Whats-the-difference-between-AI-machine-learning-and-deep-learning
8. Where ML Fits In
But what kind of tasks, exactly?
(Answer: Data-Based, Predict & Decide Tasks)
9. ML In (Fictitious)Action
Many video games
include a ‘Shop’ where
Players can purchase
needed items for in-
game or actual money.
10. ML In Action (A Thought Experiment)
Which items to
feature is often made
by the PM, who
considers:
● Previous sales
● Game conditions
● Players’ holdings
11. But what if ML did
that task for the PM?
ML In Action (A Thought Experiment)
13. ML In Action
Decision Data Prediction
Which items
should be
featured on any
particular day?
14. ML In Action
Decision Data Prediction
Which items
should be
featured on any
particular day?
● Previous sales performance
● Age of the item
● Game conditions
● Players’ holdings
● Type of item
● Day of week/month
15. ML In Action
Decision Data Prediction
Which items
should be
featured on any
particular day?
● Previous sales performance
● Age of the item
● Game conditions
● Players’ holdings
● Type of item
● Day of week/month
A list of items
predicted to
provide the
highest probable
revenue for a
particular day.
17. ML In Action
Two Key ML Components:
1. Using data sets to make a prediction about future data.
2. Update algorithm (learn) by comparing the prediction
and actual results.
18. Benefits
Companies Turn to ML Because it Provides...
● Faster decision-making which provides greater agility and faster iteration.
● More accurate predictions to provide higher return on decision (ROD).
● Freeing up humans to focus on higher-order work.
● Money saved on human resources no longer needed for this work.
19. Humans & ML Integration
Humans
Only
ML-Only
Workflows
Integrated
Human-ML
Workflows
Some cases have shown the the highest
level of predictive success (i.e.,
predictions leading to the best
decisions), comes from a joint human-
ML workflow.
● ML suggesting to Humans
● Problematic human choices
flagged by ML
20. Finally: Predictions
● ML and AI technology to become faster (e.g., neural network training within
hours instead of weeks).
● Beginnings of ‘friendly’ ML tools that will allow businesses to use the technology
without necessarily understanding the math behind it.
● Increased numbers of ML-based product and content optimization systems.
● Increased numbers of ML-based transactional security systems.
From: https://medium.com/@formulatedby/2018-machine-learning-predictions-from-the-experts-themselves-dd28d60244c1
Editor's Notes
In current form, ML employ computer algorithms in conjunction with large data sets to make predictions and recommendations. If there is a task that includes looking at large data to make a decision based on prediction, then ML can be used to either help with or takeover that step.
The combo of cheaper computing and cheaper (more abundant) data, has meant the cost of prediction has gone down.
Prediction is about filling in missing data (often, for business, making predictions that will model future data: purchase selections, sales, next video). Flagging (e.g., credit card fraud), it is when current actual doesn’t match predicted (i.e., observed data seems obherent)
Note that in most applications, ML does not take the place of decisions or Roles/Jobs.
Note: 3 types of data: Training, Input Data and Feedback Data
Workflow must be deconstructed to see where the predict and decide point(s)
Note: Input vs Feedback data...
Obviously, for this workflow, it would be conceivable for the decision itself to be automated (the system just features whichever offers are predicted to do best),
but many other workflows not only benefit from humans deciding, but SHOULD have a human decide: healthcare, military, and other life and death cases.
OR in cases where crucial data is NOT available or simply too rare to have enough data avail
There are, of course, trade-offs: all new tech requires this
Faster decision-making = Real time decision making
More accurate because wider and more precise data without lower likihood of error
Freed Humans = higher order, or simply more. In some cases, this might lead to job losses, but more often, job augmentation