Software Dev
Jun 17, 2020

Troubleshooting Artificial Intelligence

Jenalee Janes

It’s no lie that AI has the potential to vastly boost the efficiency and production of data analyses for your company. That being said, the AI-producing companies that have begun to pop up throughout the business world have a tendency to oversell exactly how pleased you’re going to be with the implementation of their AI--particularly in the early stages of its use.

Abhinov Somani, the CEO of an AI company called LEVERTON, says that a number of AI companies aren’t giving the full truth about their product. So before you hop onto the AI bandwagon, there are a few things you should probably know.

An AI Breakdown: How It Works

One thing that many vendors tend to gloss over when they’re pitching their AI is that it isn’t actually a fully-formed system that works exactly in the way you need it to the first time around. Instead, it’s made up of a hierarchy of mechanisms that all need to function in tandem at their full capacities for the AI to have a chance at producing the most effective results--and even then, it’s not guaranteed that you’ll see the best results right away.

Machine Learning

The first in this descending hierarchy is Machine Learning. Machine Learning is the mechanism that drives the “Intelligence” of AI. Without it, AI would not be able to keep producing better and better results without the help of a continuously thinking human.

Artificial Intelligence works quite similarly to the way human intelligence does. Just as human intelligence isn’t something inherent, but rather something that is gained through continuous, active learning, AI has to be able to continuously adapt--to learn--if it’s going to efficiently and effectively update its system with each new influx of data.

Machine Learning, then, is the mechanism by which AI is able to accomplish what would be unimaginable levels of learning for humans, but levels of learning for Artificial Intelligence that bring about the most effective processing for your company.

Machine Learning doesn’t have any kind of inherent knowledge. Just like people, it has to partake in a kind of studying to see success.


If Machine Learning is the equivalent of human learning, algorithms are the human equivalent of studying. The algorithm of an AI system is the mechanism that sifts through all of the new data that is being input, and sorts it according to what the algorithm already knows.

As new data comes in, the algorithm cross-references it with the data it has already processed and sorted according to its code. Once it has done this, it can then send its new understanding of all the data to the Machine Learning mechanism, which records that understanding as new information it has learned.

When either of these mechanisms doesn’t function at its full capacity, however--when there’s a bug in the algorithm that makes it analyze the data incorrectly, which makes the Machine Learning remember the wrong information--an AI system can yield inefficient, and sometimes even inaccurate, results.

Trying to work out AI training can sometimes feel a lot like you’re looking at all the pieces-parts of an actual machine at once--a little overwhelming.

Training Can’t Always Produce the Most Accurate Results

Testing your AI’s performance is the most effective way to know if and how it’s going to work for you, but chances are more than likely that your vendor won’t let you test it out before implementation. Before you make any decisions about implementing a new AI, you should be aware of some of the reasons your results may not be as accurate as the vendors claim.

At the core of testing any AI performance is training the algorithm. Before an algorithm can do what you want it to do, it has to undergo all of the processes necessary to work out any kinks it might have. Once your AI’s algorithm has been set up and adjusted for your company’s needs, training begins by feeding one of your relevant datasets to the algorithm.

It sounds simple enough, but the tricky thing that no one really tells you about is that you need a lot of data in order for your algorithm training to produce the best performance. For huge corporations like Google and Facebook and Twitter--companies that need data for their business objectives--this isn’t a problem.

For smaller companies, however--enterprises that have data, yes, and need help sorting and cataloguing it--algorithm training might not be so simple. While corporate giants have access to billions of public data points, something like a private practice family practitioner might have a harder time training an algorithm because the data might be limited.

Say, for instance, that a family doctor wants to set up a new AI system for their patient records. The number of patients, those patients’ medical records, and the types of conditions the doctor might come across in their daily appointments are all going to be limited in scope. Even if the doctor uses every bit of data they have, there may not be enough for the training to bring back the desired algorithmic function.

This doesn’t mean that the algorithm will never be successful. But it may take much longer than your company plans for, which, when you have limited time and funding, can throw a wrench into the plans you’ve laid out to help your company achieve its goals.

Research is a key part of making sure that your AI is, or at least can be, catered to your specific needs.

How to Make Your AI Work for You (It’s Not Always Easy)

Training is necessary to get your AI to work for your needs, but that doesn’t mean you have to suffer through the hardships of manual training. There are some other options out there--albeit, not all easy--so do some research before you exhaust yourself of all your company’s time and resources.

Choosing Your Training

While your results will probably be most catered to your specific needs if you carry out your AI training locally, it’s not always necessary to do so. Some AI’s come pre-trained, meaning all you really have to do with them is implement them and let them do their jobs. Whether or not this will be successful mostly depends on the specific goals of your company.

The more specific your goals are, the less accurate the training will be for your specific needs. If you’re only implementing AI for a more basic business process that every business goes through, however, a pre-trained AI model may be exactly what you are looking for.

Employing Purposeful AI

Your business goals should play the most important role in how you go about choosing the best AI for your needs in almost every capacity. Another way you can look into how to make your AI work the best for you is to consider what type of business you have. The more niche your business is, the more niche you will need your AI to be.

You might want to consider looking for AI vendors that are specifically focused in your niche field to get the best AI you can. If your business goals are fairly general, however, you can search for AI systems from pretty much any vendor and have your pick for which one you find the easiest to work with.

Testing AI Performance

Most companies measure the performance of their AI through the F-score--a classic metric measurement in the field of computer science. With the F-score system, a “1” represents certainty in the algorithm’s performance. Unfortunately, however, F-scores rarely come back with anything below a 0.55, which would suggest that nothing is really uncertain.

In terms of real life experience, this doesn’t quite add up. Almost everything in the real world is uncertain, so the AI seems to be spitting back an overly optimistic point-of-view. Somani recommends using an efficacy score instead. The efficacy score looks specifically at whether the pattern the algorithm produces is right or wrong in terms of what your company is looking for.

Practicing Patience

Like most new experiences, AI has a learning curve. Make sure you factor this learning curve in when you think about the steps you will need to take to start implementing an unknown technology. Be patient. It might take much longer than you would have anticipated, but if you wait it out, the rewards will be more than worth your time. Algorithms get better with age and practice, which mean Machine Learning gets better with age and practice. And once your Machine Learning is as good as it can be, your AI will be well on its way to success.

If you’re think implementing some new AI will be beneficial to your business, it’s completely achievable. Keep in mind, however, that the primary goal of most AI vendors is to sell you their product. You might not be getting all of the information you’ll need when you hear their pitch, so come prepared with questions about the things we’ve mentioned here, and you’ll be well on your way to getting the best AI for your business.