Original Source Here
What Machine Learning Can and Can’t Do Right Now
Two years ago, when I brought an Alexa home, my kids loved playing with the new gadget. Alexa worked well for my son who was asking all about sports, Michigan Wolverines scores, Tom Brady’s age, the Los Angeles Dodgers’ team transactions and so on. However, Alexa seemed to have difficulty while interacting with my daughter, “Alexa, what is my favorite song?” Alexa answers, “Ummm I don’t know”. “Alexa, if 2+2=5, does 5+5=2?” Alexa answers proudly, “5+5 equals 10”. When my daughter finished playing with the Alexa, she said, “Alexa, you’re so dumb!”
I’m often in a similar situation at work. Many executives or clients ask me what Machine Learning (ML) can do to figure out if their problems can be solved by ML. Tesla’s self-driving car, Amazon’s Alexa, Netflix’s recommendation algorithms, Google’s AlphaGo, and IBM’s Watson are all examples of successful ML stories. Those success stories brought excitement to AI and ML, but they also taught us ML’s limitations. Especially while working at non-big-tech companies with limited resources and limited data, I learned by experience what ML projects worked and didn’t work.
Machine Learning can do anything related to pattern recognition: for example, Yelp (where you try out a new restaurant and then review the restaurant) uses ML to classify tens of millions of photos uploaded by Yelpers into different categories like food, drink, inside, outside, and menu. Each category has different patterns — for example, menu photos will have more characters than other categories. Another example, Chase Bank uses ML to detect potential fraud transactions. The fraud detection models learned suspicious transaction patterns from approximately 3.5 billion user accounts. When a transaction comes, the models will see if the transaction has a similar pattern of past fraud transactions (supervised learning), or if the transaction has a different pattern from most other transactions (unsupervised learning). If your business problem is somewhere around identifying patterns, it can get benefits from ML.
The most lucrative use case I’ve ever seen in the real world is automating tedious human tasks. Let’s use invoice processing as an example. When an invoice arrives by email, your customer service reps (CSR) open the email, download the invoice and then manually enter relevant information into an ERP system like invoice number, seller’s business name, your address, the invoice due date, product description, the quantity of products, unit price, and much more. As you can imagine, each piece of information has patterns. For example, the quantity of products will be mostly numeric and doesn’t appear before the seller’s address. The invoice due date’s format will be mostly mm/dd/yyyy in the U.S. Product description of the invoice will mostly consist of text rather than the number. ML has been improving rapidly for this kind of pattern recognition task and many businesses have already proved its efficiency.
Machine Learning cannot do anything related to reasoning: Andrew Ng, the founding lead of Google Brain and former Chief Scientist at Baidu, says if a mental task can be done in less than one second by an average person, we can probably solve it using ML. In other words, if a task requires a thinking process, it might not be ripe yet to get the full benefits of ML. An example of this is when there are four numbers; 1, 3, 5, 7. What is the next number in this sequence? Easy for you to guess, right? Ask this question to your favorite AI assistant. The reason that you get hilarious and illogical responses is that the AI is built on Machine Learning, not Machine Reasoning. Even though ML looks so intelligent at times, it doesn’t have the power to reason effectively yet.
Let’s go back to our invoice processing example. ML-powered solutions will extract the most relevant information in the blink of an eye before human CSRs open an invoice. Can ML extract 100% of the information for all your invoices? Most likely not. There are always some exceptions that ML didn’t learn from provided data. For example, there is a row where “17-inch monitor” is in the price column and “$100” in the description column. Human CSRs, from experience, will infer it is a data entry mistake and swap the two data values. Simply put, humans know that mistakes can happen anywhere. However, the exception is not that obvious to ML. ML won’t tell you it looks like an error, nor that the value looks swapped with “$100” and so on.
What is the best way to use ML in your organization?
Even though sci-fi movies have painted our future with cold-blooded sentient robots, fortunately, that scenario doesn’t seem realistic anytime soon. As its name says, today’s Machine Learning means computers automatically learn something. That something is a pattern, how people perceive and decide. ML is good at mimicking this kind of human action. Look around your organization. Are there repetitive manual tasks? In those tasks, ML will be able to deliver tangible benefits quickly. However, ML won’t replace all the manual tasks and humans still need to do what ML can’t do. ML is faster than we are but not smarter than we are…yet.
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot