Machine Learning for Product Managers: Use Case Exploration
Recently I met up with a friend (Prem) from university and he asked what some opportunities are in the market for machine learning to be applied and what are some amazing highlights + advances in the area, below tries to explore this.
Let’s start by looking at some of the largest leaps for machine learning in 2017 and 2018 so far.
Machine Learning Advances (last 18 months)
AlphaGo - games
Why amazing? Shows progression of machine learning to self taught systems
One of the most talked about is the AlphaGo from DeepMind and Google, why does this work so well, it is a perfect set of inputs for a machine learning algorithm there are only 3 states (white piece, black piece, empty) that can exist on a fixed area (19x19 squares), these fixed constraints let a computer optimize and have limited ambiguity enter the situation, so a natural fit. Not to mention the goal of the game is to win, so a machine learning model can optimize around this desired final output.
The big leap was it was self taught, no input from human’s, just an understanding of what success was.
Tesla — self driving cars
Why amazing? Tesla demonstrated that existing systems can be supercharged with AI/ML tools to improve a consumers experience by 10x all over the air!
Being a leader in technology enabled cars, Tesla was the perfect candidate to show how a car could be updated with a self driving car via an over the air update (Tesla upgraded their software which made use of the onboard sensors and allowed for a self driving to be enabled).
Tesla’s machine learning is applied in a few different manners; lets focus on self driving on highways.
This use case for a car is ideal as there is a constant speed with tapered acceleration and deceleration, there are limited unexpected objects that interfere with a drivers condition and usually clear lines of sight for the onboard sensors to detect when cars are coming close in-front, on the sides and from behind.
Machine learning helps these sensors work together to have a set of known signals which cause different classifications to be triggered. For a camera sensor, machine learning is applied to detect objects (e.g. cars). Machine learning is essential to help, this article by David Silver goes into greater detail on how.
MIT — early lung cancer detection
Why amazing? Maintaining your health and being able to proactively fight issues will be amazing; this is just a first but important step in the field of healthcare.
Given large sets of image data small nuances can be discovered using neural networks and teams are setting out to do just that. For cancers this makes sense a doctor can have a patient benchmarked against tens of 1000’s of people to see how their body is doing versus others with an illness. With this a patient could see how they may progress or how to fight it from what has worked for others.
The challenge here is that human biology is so complex that what might have worked for a 30 year old male may not work for a 40 year old female. Each person is their own and has their own health history+genetics so quite a ways to go.
Scientific Advancement — Using lasers to compute faster
Why amazing? Shows that computing power is progressing and 10x for machine learning processing is vital to make it more accessible for businesses and consumers
One interesting technology leap was accomplished by a UK firm Optalysis who is using lasers to compute neural networks at a faster rate 10x faster! This should be interesting to watch and see how it progresses.
Where are the largest opportunities in machine learning right now that can provide real value for a product manager?
Here are a few categories we can break the opportunities into:
Type of use:
- Classification (saying if a series of data is one thing or another for example detecting cars and people in a self driving car)
- Prediction (can you guess what the right thing to do for a consumer or customer is, think googles smart reply)
Areas of use:
- Conversation – chatbots, take a common series of questions and track the conversation flows so 80% or more of it is covered and automated
- Automation – take an object or task and automate it for a computer to do it for you; self driving cars etc.
- Behavioural prediction – Google’s smart reply or shopping suggestions
- Assistants – specific use cases where suggestions, instructions or answers are provided for a niche area of expertise
Given these areas there is a lot of misconception as to what is possible given limited subsets of data or where the mathematics sits today.
Lets use classification as a base example; this is extremely difficult given the current mathematics to get past detecting more than let’s say few states with a greater than 99% confidence rate in the real world, look at self driving cars they look for cars, road lines, people, sidewalks, animals … and there are still errors with teams of 100s of engineers and billions of data points available.
That being said great leaps have been made in focusing the datasets with improved inputs (including lidar mapping and proximity sensors with cameras).
Where are the opportunities?
For Consumers:
Niche applications that automate a specific set of questions or instructions for them to get to an end goal.
Examples:
- Assistant for finding out more about a medical disease or learning about one.
- Clothing suggestions based on a price point you set and finds clothing for you when it goes on sale
- Personal call / message assistant; auto reply to messages and call for you when you are busy
The aim of the game here is to be as specific and niche as possible, since machine learning models have trouble with a large amount of ambiguity given small training data sets.
For Businesses:
Optimize existing processes such as customer support (chat bots!) or reduce review time of documents or data (filter with a high degree of confidence good or bad information — so you can focus on only what will drive value for the business)
Examples:
- Chatbots when you get to a website intercom is a leader here doing a great job, however specific conversation flows for your sales on support cycle could be automated
- Trading or finance (filter through large data sets to find outliers in patterns given a set of rules you define) think excel on a large scale or could crawl through the internet to find unique opportunities that match your conditions for investment e.g. P/E ratio is …
- Video — image analysis to find points of interest whether it is people or things that are happening can help many businesses who have image or video components and need to review a lot of data — AI can help streamline that process
- Document processing — optimize where data entry and review is tedious in documents whether it is in HR for candidates or legal for reviewing language, huge gaps that exist here.
For businesses there are HUGE opportunities to use machine learning and data science to help streamline or provide EXCEL 3.0 to existing data collection, review and analysis processes.
Thanks to the cloud revolution a lot of the data is already in server or database format so PM’s just need to come in with the right machine learning tool kit to optimize and make their teams more efficient in the long run.
I will follow up this post with limitations in machine learning today and deriving real value from a machine learning model.
In summary the tools are available today; the starting code to process or build models is mostly open source, the challenge lies in finding the business case where a process could be optimized or an experience improved given a limited set of data.
Try to see what is repeatable and has data available in your product.
Apply some simple data science to the problem in excel and see if it works
Extend it with machine learning if all your checks pass and let your machine learning computer do the work for you!
