7 AI Myths That Can Get in the Way of Business Competitiveness
We can all agree that the use of AI in business is at its infancy and may be long until it becomes widespread. However, businesses of all sizes may find it easier than thought to run early AI experiments to clear their vision on how to accelerate their competitiveness. However, several myths will be on the way and need to be reflected upon. Let’s dive into the most common ones.
1. My business isn’t sophisticated enough to require AI.
So what’s the difference between Artificial Intelligence (AI), Machine Learning (ML) and Deep Learning (DL)?
AI is humanity’s attempt to simulate our brain’s intuition and put it on steroids to experience and interpret the world for us. Its development started back in the early 50s, but the bold ideas didn’t resist reality back then. We didn’t have enough data and computational power to make practical applications. The infamous “AI Winter” started off as pure skepticism from researchers in the late 60s.
In the early 90s, the development of very narrow applications using AI concepts gave birth to what we now call machine learning (ML). Chess master Garry Kasparov was defeated for the first time by a computer program — not using ML — in 1994. Two years later IBM’s Deep Blue used what can be interpreted as an ML algorithm to consolidate computer’s superiority in chess. Today’s state of the art ML algorithms are used as tools to train and validate human strategies for the game as no human being would stand a chance against them.
Deep Learning (DL) is the 2010’s comeback of ideas pioneered in the early 50s. Researchers saw the feasibility of multi-layered artificial neural networks due to the abundance of data and computational resources. You may have heard that DL gained traction with the extra performance provided by graphic processing units (GPUs).
DL is part of ML, but there are two important distinguishments: 1. Instead of containing a purely mathematical algorithm, an artificial neural network is used and 2. several layers of artificial neurons will be put between your input and output signals. The way information goes through the layers will depend on the problem you want to solve, but in general terms, DL is a killer way to map input data (x) to some desired output class label (y). So when you have new input data, you can predict the output with a certain accuracy. It’s kinda spooky that a very loose representation of the brain can lead to accurate prediction results. But the cost for this ability is the incredibly large amount of labelled data you will need to train the neural net.
In summary, we say that DL is a subset of ML which is a subset of the broad field we call AI.
Your business can — and eventually, will — use AI. The reflection about which approach to use, DL or a very specific ML algorithm, will depend on your objectives and the data you can bring to the table. Just because there are companies building self-driving cars, it doesn’t mean that AI is restricted to such breakthrough projects.
2. AI is a magic box, just throw another problem at it!
While there is something magical about predicting an outcome from an input that the machine learning model never saw, the magic ends there. If you try to use machine learning without minimally understanding the problem you want to solve, you will fail miserably. It’s very important to think of your AI strategy as a portfolio of approaches to solving very hard problems you can’t solve with traditional programming. Each problem may require completely different datasets and approaches to achieve meaningful results.
3. Only the big companies have enough data
While it’s true that whoever has the data will have an advantage in solving certain problems, no business should be trapped in the analysis paralysis around the question “do I have enough data?” Maybe you don’t, but that doesn’t mean you shouldn’t try to attack a business problem using AI. There are some scenarios to keep in mind:
- Sometimes you can augment the necessary datasets with public or purchased data
- By creating the first version of your application, you may get your users to generate the data you need to improve your ML model
- Depending on the problem being mapped, you can hire people to generate the data you need (crowdsourcing, Mechanical Turk, etc.)
- It’s common to use computers to generate data that can be used to augment your dataset. Imagine that your dataset is composed of photos of faces taken with smartphones. You could use standard algorithms to simulate different lighting conditions.
4. “AI is always learning” or Models improve with new data, ‘automagically’
Most of the machine learning models are trained offline. Surprised? The thing is, things can get widely out of control if you just feed more data to your model. By keeping humans in the loop, you can make sure your models will keep performing well.
So, every time Siri, Alexa or the Google Assistant tell you they can’t help you, but they are learning, it doesn’t mean they are learning with you right then. However, the collection of inputs that didn’t map to any result is highly valuable data to help you fill the important gaps with users. You will need to use them to retrain your model.
Online machine learning models are possible, but not advisable. An example would be a self-driving car training itself as it navigates the roads. Just imagine if an autonomous vehicle is left on its own without any human control or interference. If your robocar runs over a dog and it doesn’t feel anything, it may consider this normal.
5. A low accuracy model can’t be used
We recently had a ‘Fortune 50’ company ask us to add a clause in our contract that would guarantee 95% accuracy in the ML model. We obviously didn’t accept it, but the reason wasn’t because we didn’t believe our model would be good enough.
The graph above is very common. During training, a typical machine learning model will have an accuracy that asymptotically increases with the number of data used to train it. After training, you will test the model with your evaluation set (a subset of the data you had at the beginning) and see how the model performs. Note in the graph that the training set achieved 100% accuracy, but the evaluation set is slightly above 90%. This means that the model is “bad” or, more precisely, “overfitted”. You want a model that behaves well with both training data and new data.
Sometimes accuracies above 80% will be more than enough for practical applications as long as you have a good plan to work out the situations where the model doesn’t work well.
But now think of the variety and representativeness of new data that you can gather over time. You will have what you need to constantly improve your model. Building a machine learning model should be seen as a journey that tends to bring you increasingly valuable returns.
6. UX is irrelevant for Machine Learning
The image above is from a mobile application that implements the imagenet model for image recognition. As you can see, the photo on the left, from above the mouse, led to an unexpected result. By tilting the camera I managed to catch the right category, albeit at a small confidence percentage.
Now imagine if the mobile application used the device sensor information like gyroscope data, and it told me that I should tilt the camera in order to get a better result. It would’ve guided me to a better experience because it would’ve provided the machine learning model a better input.
Depending on how you design your application, you can also get valuable information from users that will help improve your model.
In a different example, imagine how UX could improve your experience with a chatbot. They can be very frustrating because chatbots don’t often answer all of your questions. User experience design in this case goes beyond visual cues or feedback buttons. You will need to consider the conversation design, the sentiment analysis and the tone of the answers.
7. I don’t have budget for an AI project
The cost of building your first AI project is equivalent to the cost you had when you built your first mobile app, just to give you a tangible reference. If your company can afford a mobile app, you should be able to set aside some money to get started with AI. The biggest advantage of an AI project is that it will force you to constantly analyze the data and adapt. How many mobile apps have you built that got stuck in time?
In contrast, the cost of not building your first AI project soon, rest assured, will be much higher as time goes by. You wouldn’t want to be left behind learning the basics, while your competitors are already building and creating.
Companies who will treat AI as part of their portfolio of problem-solving tools will probably achieve compounding gains over time. They will have, however, to manage internal expectations around early results and consider experiments as bets worth making.