Manager’s Guide to Demistify A.I.— Part 1
“Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don’t think AI will transform in the next several years.”
~Andrew Ng, Adjunct Professor of Computer Science At Stanford University
My curiosity kept asking me…
How does it work? How do I know if this is can be done by AI? How long does it take? What are the first steps? How much data is needed? Does it require PhD level expertise?
As an innovation manager in technical talks day in and day out, it just bothered me that I did not know enough…that I did not know what I did not know. And decided to change it.
After some self-study, and conversations with ML practitioners I decided to go for a Nanodegree as AI Product Manager. This helped me further de-mystify ML for myself to a degree where I feel much more confident in evaluating the possibility and feasibility of AI business cases — and can also build a simple AI model by myself.
For business managers who want to adopt AI in their businesses and want to know more, I have prepared this 3 step guide to accompany you along the journey of demistification:
- Understand the steps involved in developing an AI model
- Build your own AI prototype using Google AutoML
- Building the right AI model, avoiding bias
This blogpost will be focused on Step #1. Steps #2 and #3 will be discussed in the upcoming posts.
Understanding the steps involved in developing an AI model will make you more productive and confident in suggesting AI solutions to problems, and evaluating them. Let’s start with the steps in building a machine learning (ML) model (adapted from “7 Steps of Machine Learning”):
- Identifying the problem
- Gathering data
- Preparing that data
- Training the model
- Evaluation
- Prediction
There are two ways in which you can make a ML model: 1) Use an existing, ready to use model (e.g. Google AutoML) or 2) Create your custom model. The existing, ready to use models can be utilised for the most common use cases. If your case is more specific, it may be better, though more resource intensive, to go for a custom ML model.
For the sake of simplicity, we will focus here on using an existing Auto ML model. For a custom model we would need to add additional steps. These would be “choosing a model” which would come right before step #4 “Training the model” and “hyperparameter tuning” which would normally come right before step #6, “Prediction”. In Auto ML, these steps are taken over by the application itself.
- Identifying the problem
Before you start with your ML prototype, it is important that you are going after a real customer or organisational problem.
If you have not yet identified it, you can follow the methodology described in my previous blogpost to come to this stage.
2. Gathering the data
In ML you teach everything to the machine as if you are teaching a child — see a ball, point at it and say ‘ball’ and after a while the child points at it and screams in excitement ‘ball!’. Gathering data is all the reference points you are collecting to teach a machine — in this example — what a “ball” is.
Now the child knows that thing that bounces, kind of elastic, colorful and round is a “ball”. Here a machine is much more primitive than a child, in which it can neither feel nor smell nor interact with the objects around. All it can do is learn to relate from the images and the “context”, environment surrounding it. So you need to feed the machine with a lot of images of balls with different colors, sizes and types so he can learn (program himself to) what a ball is, in various contexts.
2. Preparing the data
This is the process of preparing the data to tell the machine ‘ball, ball, not ball, ball, not ball,…’.
To do this, you need to mark all the images with balls ‘yes’ and not ball ‘no’. The process of marking the items on a picture or labelling the picture as yes / no is called ‘data annotation’ or “labelling”. Currently most of the data annotation process is being done by humans and is one of the highest cost items in building AI models.
3. Training the Model
The process of teaching a machine certain parameters so that it can make the necessary predictions is called in ML ‘training the AI model’.
In order to train the model, you need to upload the labelled data into the platform (e.g. Google AutoML, Amazon ML,…). Depending on the platform, you may need to provide additional instructions regarding the training of your data. The platform automatically chooses the right learning algorithm for you.
4. Evaluation
This is like a small quiz where you ask the machine the questions where you already know the answers — “Is this a ball?”
Depending on the percentage of the right answers, you can decide if you want to further train the model or move on to the prediction phase.
5. Prediction
Prediction is the step where the machine learning starts providing you the answers, where you start benefiting from the results.
Important to take care that in many cases the learning may need to take place continuously, also after deployment, if the input data is evolving.
Now some examples of how ML will enable business innovation:
Autonomous vehicles
- Identify the problem
As the cars drive around by themselves, they need to understand their environment to decide on the next action- what is around me, is this a traffic sign, a car, a pedestrian…etc. How can ML help?
All these need to be taught to the machines, like to a small child. Applying the next steps of the process:
2. Gather Data
You find a lot of pictures from the streets including cars, traffic lights, pedestrians, etc. It is important that you have a balanced, diversified data set, for instance showing objects with different characteristics and from different angles.
3. Prepare all the data
Mark all the objects on it -> Car, human, traffic light, people, pavement, stop sign, etc. The more examples the machine has, the better he can evaluate what is what.
The most used technic to label object in the pictures is called “Bounding boxes” here more specifically “2D Boxes”.
4. Training the model
If you are using an existing model, this is pretty easy. You just press the “Train the model” button! After several hours, your data will be trained.
This is the way the car will “learn” to identify the surrounding objects, which is a vital input for decision making.
5. Evaluate
For evaluation you can see make a small test to evaluate the reliability of the model — show a picture and ask the machine to identify the cars. Depending on the results, you may need to refine your model.
6. Prediction
It is “deployment”, where you basically bring the model to life. The machine now should be able to identify the different objects in a given image.
Other popular use cases of ML:
- The way our mobile phones detects the faces to focus on while making pictures
- X-Rays in which the machine can detect certain diseases
Using sentiment analysis to protect brand reputation
- Identify the problem:
There are a lot of comments on the Facebook channel of your company. You want to know if there is a problem so you can react on time not to damage brand reputation. How can ML help?
This can be done through sentiment analysis. How you go for it:
2. Gather data
Collect a lot of comments from users expressing their feeling towards a brand or product.
3. Prepare all the data
For the collected data, label if the sentences indicate “positive”, “negative” or “neutral” sentiments.
4. Repeat the steps 4–6 above.
Next time a user feedback is posted, the machine will be categorising it accordingly and you can only focus on the negative sentiments that could potentially damage the brand reputation.
Knowing these basics how a machine learning model is actually being trained, and what is the input data vs. the output helped me answer some of the questions I had in my mind and demistify AI. Needless to say that the whole AI process is not as simple as shown here. This blogpost aims to give you a basic understanding, eliminating the complexities you may come across on the way.
Here some more self-study resources that can help you further demystify AI:
- AI product manager from Udacity (ca. 2 months, no coding required)
- AI Online course from EdX (basic Python knowledge as prerequisite)
- AI Adventures videos from Google Cloud
Follow me to hear about more practical tips in my next blogposts, and enhance your abilities to adopt AI in your business:
