Machine learning is a certain type of artificial intelligence that allows the development of systems that can learn without being programmed to do so, or with only minimum startup logic. The primary way in which machine learning works is that the system develops knowledge or intelligence in response to continual exposure to new data.
The system learns by identifying and absorbing patterns present in the data. Machine learning has a strong association with AI and robotics, and was hence confined for a long time to the scientific side of things.
With the existence of high-scale business computing becoming mainstream with the Internet and the arrival of cloud computing, machine learning is in a position to drive business benefits on a large scale.
If you have ever seen book recommendations on Amazon, driving directions on Google Maps, or have run the spell checker function in Word, you have benefited from ML. Let us look at some ways in which machine learning can drive big business benefits.
How do organizations use machine learning?
With the ubiquitous presence of computing, the Internet and cloud computing, ML is anywhere and everywhere. Email programs use machine learning patterns to identify if an email message is spam or a regular email.
The key fact there is that the learning is adaptive, which is why with continual absorption of data, the program is able to identify newer forms of spam. The same goes with antivirus programs. When you are using something like Google Translate, there is an algorithm in the background running on the principles of ML that is reading the text and performing the translation.
Facebook uses machine learning tools to implement face recognition from photographs and suggestions for tagging members’ faces. If you take Amazon, there are programs running on the server side that are able to observe the patterns related to your browsing and prior purchases to provide you with product recommendations.
Challenges related to implementing machine learning
As exciting as the prospects of implementing machine learning are, there are enormous challenges that come along with its implementation. Just understanding what type of algorithm to use for the problem at hand itself can be a difficult task. There are several types of machine learning algorithms that can be used, and each one can be used for specific types of applications.
It is like finding the cure to a disease. Which route do you want to take? Which route should you take?
If you take a clustering algorithm, for instance, it can be used to differentiate customers on the basis of whether they are more likely to desire a take-out order or would instead dine in. However, such an algorithm cannot be used to assess the impact of a change in the menu prices on sales numbers. Similarly, if you take a regression algorithm, it can be used to assess the impact of price changes on the sales. However, it cannot be used to predict one of a fixed set of outcomes.
One more risk with machine learning is that of overfitting the data. What that means is the act of training a system so extensively that it loses the capacity to learn, generalize, and predict results based on new data. In such a scenario, the model tends to become worthless as it starts to make inconsistent predictions. Kind of like Matthew McConaughey’s character did in the average movie "Two for the Money"! No one cares for inconsistent predictions. Now he did redeem himself at the end, but that is another story.
Finally, some problems may simply not be solvable with ML at all. It cannot always be predicted what type of problem cannot be solved, so what can happen is that the process of applying data to the algorithm continues for a long time without ending. In such a scenario, a functional model is never really developed.
Will machine learning be a good choice for your organization?
When implemented accurately, ML can help you solve large business problems with the potential of adding revenue or reducing dollar spending in a significant way. Just about every organization around is drowning in data ranging from purchases, customer demographics, location data, search data, pricing data, inventory information, and delivery-related data.
Given the humongous volume of data, machine learning can potentially provide a solution for many of the problems that involve the processing of large sets of data on a continual basis. Like, how many donuts Homer Simpson eats in a day or how many times Michael Kelso (“That 70s Show”) makes a fool out of himself! No, those pursuits are not really that important!
To implement machine learning, a business needs:
- A thorough understanding of the machine learning process.
- An understanding of the various types of algorithms available and the types of problems to which they can be applied.
- An understanding of the data flows.
Realizing big business benefits
Consider a scenario at a FedEx-like company where several packages are being routed to the wrong destination. The situation is this. The packages come in with a certain barcode indicating the right destination. The routing machines are supposed to receive these packages, read the barcode, and paste another barcode sticker to its left.
Given the fact that many of the packages do not have properly printed barcode stickers, the barcode is not being read properly, and consequently the new stickers are being pasted with incorrect barcodes. This situation is ripe for ML. Even Joey from “Friends” could figure that out!
These could be the steps in implementing a ML-based solution for this problem:
- Put in changes in the code running the routing machine that take in a new file with the list of “bad barcodes” in the output.
- Make some processing runs in which a count is made of the number of packages that are being fed into the routing machine that are being read wrongly.
- Gather a visual pattern for the barcodes on all the packages that are being identified wrongly, thereby causing the routing machine to print the wrong new barcode.
- Apply changes to the image processing software present in the routing machine that would recognize such patterns over time of the improperly printed barcodes and be able to make adjustments by using machine learning logic.
- Once the changes are in place, each time a case is encountered where the input barcodes are being printed improperly, the case is recognized and the correct output barcode is printed on the new sticker.
- Once test runs are completed, the changes in the image processing software are productionized.
A solution like this can literally save millions of dollars every year for the company.
Photo credit: 20th Century Fox