“Aim to make decisions, not just predictions.” – Google ML

AI has been a moving target. When the term “artificial intelligence” (AI) was introduced to the industry in the year 1955, a lot of speculation surrounded it. Most importantly how it would take over a lot of the jobs carried out by humans. The idea is to mimic the functioning of the human brain. But it is the most complex organ in the human body. Can it really be replaced by a machine, so soon?

Machine Learning is a method of using a combination of algorithms and data to train machines. As per Richard Freeman, the difference between Machine Learning (ML, a branch of AI) and traditional programming is that ML trains the machine and forms the rules based on the data provided. In my experience, those who have lesser understanding of the inner workings of AI, tend to think AI is magic. They expect AI to be the solution to all their issues. Partially, pop culture is to be blamed for that — The Matrix, J.A.R.V.I.S. in Iron Man and so on. But it is a misunderstanding. AI models are only as capable as the algorithm and the supporting data provided to it. Whichever model you choose, there should be measurable data that you can collect. AI is a tool to enhance the day-to-day functions.

How do we decide that AI is the best solution?

AI would be extremely useful for increasing process efficiency. One important use would be to measure data while day- to- day operations are occuring.  This would ensure more realistic data and resulting AI models.

Here are some questions you should ask before deciding to use an AI model (Richard Freeman). Will the model:

• Be used for matching, recommending, predicting, or suggesting options to users?
• Increase revenue and decrease costs?
• Compare when humans execute the task versus the machine untrained and trained by the model.
• Be easy to maintain and monitor the model?
• If not, it is possible to incur more costs when using such a model. And that would defeat the purpose of using the AI model.
• Be using the right data set?
• Are the input values measurable and can be recorded?
• The model is expected to provide a viable output using this dataset. This output is what will assist with decisions.
• Is there a general pattern visible when using  sample data? This is one of the checks before deployment of the model.
• This model will be used to predict possible output therefore it is best if this pattern  is valid even in abnormal situations. (Cassie Kozyrkov) Otherwise the model is of no use.
• Please note that this can be also managed by using a dataset from sites like Kaggle,  if data is not readily available.
• Ensure to consult SMEs (subject matter experts) before delving further.

The 2 common training methods in ML include supervised and unsupervised learning.

Supervised Learning

Supervised learning would use data that is labeled and categorized. Because it is labeled, it would be easier to predict using a graph plotted with their data points.

Example:

You are depositing a cheque and you upload an image of it to your bank app. The AI model running in the background will then be able to extract key information from it accordingly by supervised learning. When depositing a cheque, upload an image of the cheque to the bank app. An example of supervised image learning - The AI model is provided various images that covers the required security features of the cheque, digits, and alphabets to be detected at marked areas on the cheque. Here are examples of  features to be recognized:

• Bank logo and name
• Customer name and address
• Customer signature
• MICR – this includes bank, account no. and the cheque leaf no.
• Dollar amount (digits and text)
• Endorsement (at the back of the cheque leaf)

Unsupervised Learning

Alternatively, unsupervised learning will contain data with no labels. The goal when using this method is to identify any underlying patterns in the data.

Example:

When Amazon would like to identify associations between what all you buy when you buy a mattress, fitted sheet, pillow covers, duvet, side table.

Combination

A third method is a combination of both, called semi- supervised learning. Real- world problems fall under this category. Only part of the data is labeled.

Example:

Small grocery stores usually function using this method. They have price labels on some products and use that information to estimate prices of similar products.

Fun Fact:

As a new mother, I can see similarities between a newborn and an AI model. At 8 months, she cannot communicate in a language I understand yet. I try all permutations and combinations – milk, diaper change, sleep, releasing gas, uncomfortable position, etc. If I record the time and action completed, I can load this into a model. Assuming there are no identified health conditions, I should be able to predict what she would be doing at the specific time of the day (until she is 10 months old, where she is expected to undergo sleep regression, which is not a statistical term in this case).

I train her using a supervised method. There are the phrases and actions I use:

• “Are you hungry?”
• Waving a milk bottle
• Using ASL sign for milk

Depending on a 2-hour gap, I can predict when she will start feeling hungry.

In the coming weeks let us learn together what you need to build a basic model for your purposes.

For further reading:

https://towardsdatascience.com/when-to-not-use-ai-or-use-it-based-on-my-experience-abb58c063aba

https://medium.com/hackernoon/ai-reality-checklist-be34e2fdab9

https://www.youtube.com/watch?v=2ePf9rue1Ao