Why So Many AI Projects Fail and What You Can Do To Improve the Chances of Success
Updated: Oct 24, 2021
Artificial intelligence (AI) and Machine Learning (ML) are quickly making their way into the daily use of businesses of all sizes and industries. From customer service to manufacturing, procurement, logistics, and marketing, AI has numerous applications that help predict trends and optimize all business aspects.
Nevertheless, many AI projects face significant challenges when being applied to reality and outside of the lab. One of the main challenges is the need to train a model so it can predict accurately. This article will dive into the difference between two models used in AI projects and examine the main advantages and disadvantages, so eventually, you can judge for yourself how to improve your next AI project.
The Challenge of Most AI Projects
In its most basic form, the challenge of many AI projects can be summarized as the "generalization problem". In other words, how to teach "the machine" to generalize a model that makes the right decision when facing new data unseen before. Generalization is one of the most common challenges with AI, and we believe that all types of AI problems stem from it.
Think about this simple example. Imagine you wish to train a facial recognition model to identify dogs around you (whatever the reason). To do so, you will need to train a model and "show" it a certain number of pictures of dogs (usually a relatively large number of images). Eventually, and after the model "sees" these pictures, it is supposed to be able to identify dogs in the real world even without knowing in advance that the object in front of it is a dog. In other words, it generalizes from previous information into new information. To do so, data scientists have several methods. Two of the best-known methods are "train & test" and "cross-validation".
I. "Train & Test"
The first method, "train & test", is pretty simple. It is common to take 80% of all data and train the model on it. The remaining 20% are left for examining the model itself and its accuracy. There are two main advantages to this method:
Speed: It enables quicker training of large data sets, eventually allowing faster time to production or time to market.
The train & test method is widely used and accepted, leading to greater trust in the product once released. Such trust is also the result of randomly selecting data, leading to a supposedly better statistical analysis.
However, there are some key drawbacks to this method:
Fore and foremost, the data for the test (the remaining 20% used for examining the model after being trained) is randomly selected. In many cases, the test fails, and then many data scientists go back to the train set and search for a potential reason to explain the model's failure.
Following the conclusion that part of the data in the train set has been flawed, that data is left out, and the model is being trained again without it. However, in real life and outside the lab, even w1hen the model now proves to be successful, what has really changed is the data set used and not the algorithm itself.
Such misconception can lead to an incorrect thinking process and to attributing the "success" of the model to the algorithm. Whereas, in fact, the whole data set was manipulated, eventually leading to failure in the long term. Therefore, the train & test method suffers from a key weakness: It cannot accurately measure the learning process and the generalization of the model in real life.
The second method for evaluating generalization is cross-validation. In this method, the data is divided into several random sample groups. The data scientist takes the first group, puts it aside, and trains a model with the remaining data. Afterward, the model is validated with the first group initially set aside. Following this validation, the same method is applied to the second group, the third group, etc.
There is one main advantage to this method over the train & test method: The model is tested with a large amount of data and without random sampling. Since all the data is used for testing, the probability of applying the model to real life is much higher.
However, this method is not commonly used, as it is much more expensive and takes longer to train the model. When dealing with a lot of data, the difference in time and cost can be pretty significant.
A Faster, Cost-Effective, and More Accurate Method
Many ML models fail to shift from the lab into actual application in real life because of the inability to validate a completely accurate model. The solution is to perform a cross-validation process with a smaller but statistically representative data set, allowing data scientists to achieve more accurate generalizations while saving time and money.
Using the methodology we have developed for smart sample selection, we can use cross-validation with smaller data. Yet, our method provides enough statistical credibility, as if the entire dataset has been examined.
This way, we can perform cross-validation training for 100,000 samples faster and with better results. It enables us to reach cross-validation accuracy while performing quick and cost-effective testing, just like the train & test method - a win-win solution for everyone.