Almost 20 years ago, I was implementing a Fraud Management product for Telecom operators around the world. It included wireline long distance call operators as well as Mobile operators. The product created accumulators that triggered alerts when pre-determined criteria were met. For example: numbers of calls in the last hour/day, calls from two or more locations in a short period of time. However, we soon realized fraudsters were getting more creative. Implementing new accumulators and rules to combat them was not easy to achieve that quickly.
That’s when we started exploring options for implementing a Neural Network-based solution (AI/ML were not the buzzwords) that could learn, based on existing fraud data, and then detect fraud cases. This presented its own challenges. We had sufficient data (call detail records were huge and fraud data was already available). However, we didn’t have enough machine power to process the data and create sufficient network learning repeatedly. On top of this, no math libraries were easily available to implement such a network—and the ones available exceeded project costs. This is probably what happened to most AI/ML projects for the next 10-15 years as machine power was still evolving and there weren’t enough tools available for any eager developer to download and put to use.
So, how is this related to pinball? I was at a Pinball Gallery for a kid’s birthday last month (below photo taken there) and it gave me the idea for this blog post. Pinball–Space Cadet was a fairly famous Windows game around that time. And while I was researching Neural Networks this game stuck in my head as a neural network in action. The pinball launch is like an input to a neural network and every mission target it hits is the neuron it “triggers” as a hit. After a number of launches, the network will be trained based on how each neuron was hit by the input set.
Last year, I completed a 5-course package from coursera.org, and tools like TensorFlow and Keras have taken Math out of core programming and languages like Python and R have great support for Matrix computations, which greatly helps in processing.
Fast forward ~15 years from old challenges, and AI/ML was advancing a lot. The industry was picking-up speed with a lot of AI interest, data availability, data storage capabilities and machine power, and all that helped by freely available developer tools.
Recently, however, general interest in AI/ML has been slowing down. It’s still hot but there seems be a slowing of demand. The main reason why is that AI/ML is reaching the next limit of computing power. While the amount of data to process is huge, the processing power to handle it still lags behind. Additionally, the personalization of AI/ML-based products is limited by privacy concerns, specifically in the US. Therefore, such products are failing to launch at their full potential.
In the context of the current Covid-19 situation, Israel is able to use a person’s mobile location to alert the authorities when it determines if anyone was close to an infected person. However, the same technology may be limited in other countries due to privacy concerns, and such data may not even be shared for such applications. Therefore, AI/ML may be slowing down a bit now, but once other factors become more favorable, it will be back to pick up speed.
If your organization is interested in learning what Machine Learning can do for them, our ML Kickstart arms your organization with a cohesive ML strategy in just two weeks to maximize your organization’s potential through faster and smarter decision making.