I went to graduate school and got my Masters in Electrical Engineering, but that’s not what I studied. The University of Arizona had a strong program in optical sciences, medical imaging, information theory, and something called “machine learning” and I opted to take classes and ultimately complete a thesis on these topics. I remember learning the math and computing of neural networks, the computer vision algorithms behind facial recognition, and the underlying mathematics of mpeg encoding.And I remember spending countless hours in a lab testing algorithms on a Unix workstation. Would a reinforcement learning algorithm work better than a two layer neural network? Should a genetic algorithm work better, or am I programming it incorrectly? Should I apply a fuzzy controller, perform an operation in the Fourier space, focus on heuristics or prove out the underlying mathematics?
Most of what I remember is waiting for that workstation to spit out a result. There wasn’t a supercomputer that I had access or a cloud environment where I could ramp up and run several experiments in parallel. In the end, artificial intelligence back then was a lot of experimentation between what data sets to test, what algorithms to apply, what parameters to configure, and how to best program them to get better performance.
Are you Ready to Experiment with AI and Machine Learning?
I’ve been fortunate to have had some opportunities to develop artificial intelligence in business applications. I’ve developed or led teams to develop tools that enable comparing genetic and protein samples, natural language processing algorithms to extract search terms from newspaper classified advertisements, and document processing techniques for extracting building material names from construction blue prints and building specifications. What was common across all three applications is that it required significant experimentation, first to get the basic algorithm in place and then later to build up more intelligence to handle more disparate use cases with increasing quality and performance.
Is AI Today Fundamentally Easier?
The simple answer is yes, but the longer answer may be no. Today, a developer can access AI through APIs provided by
IBM Watson,
Google Prediction,
Microsoft Cognitive Services,
Amazon Machine Learning and many others MLaaS (Machine Learning as a Service) or AIaaS (Artificial Intelligence as a Service). The largest
startups in AI have been funded north of $10M and the
biggest companies are making multiple investments. There is a published
Intelligent App Stack illustrating use cases, machine intelligence providers, and data prepping tools and other
AI Marketplace overviews.
But the longer answer is maybe not. As a business person with an opportunity to apply AI or machine learning or a development team that has the priority to implement, you have many options to implement. You need developers that can normalize data sets and plug into APIs or third party services. You need data scientists that have at least some familiarity with the basic algorithms from clustering to neural networks to deep learning. You need subject matter experts that can validate the output and suggest improvements. Most importantly, you need an agile experimentation process to try approaches, configure, run, and validate results.
Lots of choices, lots of talent, lots of time to implement. But the rewards can be significant.
Leave a Reply