Meet Dr. Padma Polash Paul, a man whose efforts in the field of artificial intelligence and machine learning led to the development of mlOS – the machine learning Operating System.
mlOS is a platform to build, deploy, and manage machine learning models. The ease-of-use and speed at which it can build and deploy models are helping developers and businesses with rapid modeling. What started out for Dr. Paul as an undergraduate course in computer science 20 years ago has become a career dedicated to machine learning.
Padma was first introduced to machine learning in the late 1990s. He taught himself about ML through trial and error. It was a time-consuming approach, but it was his only option at the time. This learning experience provided a strong foundational knowledge about ML, which has turned out to be a tremendous benefit throughout his working life.
In his early days, he had ~250 algorithm combinations – 10 for feature extraction, 5 for feature selection, and 5 for ML. That’s 250 combinations (10 x 5 x 5) that had to be tested by copying and pasting code over and over. The process was as repetitive as it was tiresome. Dr. Paul committed himself to constantly improving his techniques.
Problems of a Researcher
Over time, Dr. Paul transitioned from a student to a researcher. His work changed significantly. More importantly, the combinations of algorithms became more complicated.
His ML experiments grew to 22,500 combinations (10 x 3 x 10 x 5 x 15) – 10 for feature extraction, 3 for dimension reduction, 10 for feature fusion, 5 for feature selection, and 15 for machine learning. Manually testing each one was the only way to learn and to build upon existing work.
There were several problems that Dr. Paul had to contend with as a researcher. Publishing research meant that he had to combine several thousands of algorithms in completely unique ways, as well as get the best results, save, and then recreate those results during thesis presentations. The algorithms had to work on the data given by other researchers, like to compare results. He had to deal with infinite possibilities from the outcome of thousands of combinations that could apply to several problem domains. His research in the field had to be outcome-based and involved many comparisons, reproductions, and generations. All the data also had to be tracked regularly and continuously. Dr. Paul learned through the entire process.
Eventually, he went on to work in industrial application development. Designing commercial software meant making machine learning models that were production-ready.
As a researcher, his primary focus was on performance and model output. Now, as a developer of industrial applications, he had to focus on APIs, scalability, security, and interoperability. His job also involved managing and coordinating both code and a team of people.
Problems of a Developer
Machine learning is complex. There is no doubt that takes a lot of time and effort to build and deploy models. And developing models at scale requires an army of people that are not often available or feasible for businesses.
Developers don’t always have all the coding abilities and data science skills required to use the available tools. Too much time is spent on a trial-and-error where developers need to repetitively copy and paste code. The process leaves them with very little time to deploy and manage models.
There are paid tools that make the process easier. But these tools are unaffordable for many. Free tools come with minimal functionality.
Dr. Paul co-founded Braintoy to solve the pains he faced as a student, a researcher, and as a developer. Today, he is the Chief Technology Officer. He performs his jobs with the same diligence but with a different focus – to help students, researchers, and developers like himself overcome the banes of machine learning.
If there is one takeaway for all developers from Dr. Paul’s experience as a machine learning expert, it is this: the traditional approach to machine learning is time-consuming and involves significant trial and error.
It sounds crazy, but it’s normal to have to go through several thousand models to get the right one. Revisiting your models to compare, contrast and reuse can help improve your techniques and make incremental improvements, but the time and effort are still too much for many to handle.
mlOS is a production machine learning platform that provides developers with a fast, uniform, and repeatable way to produce and manage models at scale. Teams love the collaboration features. It made AI accessible to businesses.