Data Mining to Deep Learning

Two day's ago I posted an extremely short note to express the progress in methodology and infrastructure revolution in data and analytics space including Data Science. I got an impression from a couple of folks saying that I may be creating a confusion that Data Mining is dead and GPU emerged from CPU. I thought it is worth explaining it a bit. 

The question what is the difference between Data Mining, Data Science and Machine Learning is a well-debated one, and many eminent personalities in the area expressed their final comments on this. The general confusion on the terminology appeared because of definition issues. As the Indian systems of logic suggest, any definition should be free from over applicability, non-inclusiveness, and non-applicability. The definitions for these two terms were too ambiguous and overlapping. But people who practice knows that both Data Miner and Data Scientist is doing almost the same job. The difference is that as it is a live field, new methodologies procedures tools and technologies made the life better. Eventually, the terms became subject to over usage and PR stunts. It created confusion to many newcomers in the field. 

The other side of it is, in the early day's people who completed an academic course in quantitative fields became natural Data Mining professionals, and there was few certifications/courses to be a Data Miner by training. The open education and Massive Open Online Courses, Web2.0 revolution knowledge freedom and all helped people to acquire this divine knowledge beyond the quant world. Many became successful professionals and entrepreneurs through this process. Regardless of the size majority of IT companies benefited from this directly or indirectly from this too. Now the answer to the ultimate question is the Data Mining era over? No, not, rather it is getting a more polished practice with all the IT revolution contributing to it. Examples of such advance included, feature engineering strategies, stacking and model ensembling, parameter tuning with grid search and meta-learning concepts. 

Beyond all of these, there was some terminology came into existence called the "Unicorn Data Scientist." Earlier days people who focused only in the area of Natural Language Processing and Text Analytics and individuals who are working on Image, Speech and Video processing were considered as super specialists. They were not much considered as a Data Mining professional. Computer Science professionals gained knowledge in Data Mining was considered as versatile and unicorn IT professionals. The Big Data Revolution and immediate increase focus on R and Python for Data Mining/Science were there. The repetitive domain and range definitions for Data Science brought skills like SQL, quant knowledge, computer science, Natural Language Processing, Image Processing, Text Mining, Video Processing, Big Data technologies and IoT into the purview of Data Science. A good number of exceptional computer science, quant and other disciple professionals and students benefitted from the free knowledge available and made their hands dirty with extensive knowledge, data, and tools. Later such professional were secretly/openly referred as Unicorn Data Scientist, who can work in Image/Natural Language/Speech or Predictive Modelling with or without Big Data and various programming languages. 

Still, we didn't touch the Machine Learning and Deep Learning part. Machine Learning is the only one field which was not subject to much of definition abuse and still feeding the Data Miners and Data Scientists with sufficient delicious recipes. But deep learning which is part of Machine Learning got intensely subject to PR. The revolutionary thoughts on multi-layer neural networks started from early 90's but got much attention and emphasis by 2000's. The journey from LeNet to Google Sonnet is fascinating. Last five years we are witnessing significant outbreak in the research outcomes in the area of advanced Neural Network oriented research results. Simple predictions to autonomous vehicles, producing art to Machine Translation every field now Deep Neural networks or Deep Learning is being experimented, and satisfactory results are being reported. Many great Open Source tools are made available by giants play with Deep Learning. The larger community got benefitted and contributed back to the revolution too. So Deep Learning is still part of Machine Learning, but there is an increased emphasis on this topic now. 

Then why we talking much about CPU, GPU, and TPU. None of them are natural replacements for the other. They are the advancement in computing capability we are witnessing in tody's wold. The Data and Analytics professionals are getting significantly benefitted from this. Specifically in Deep Neural Networks innovations like CudaNN kind of inventions are worth mentioning. The introduction of GPU, a differently architected processor for increased parallel processing capability, accelerated the Deep Learning revolution of today. The error rate reduction in ImageNet competition is the best evidence for the same. But still, we depend on CPU too. So one doesn't took over the other or killed yet or not going to kill soon. The Tensor Processing Unit, which got attention from the Google Tensorlow paper is just a hint on the ASIC level innovations for accelerating Machine Learning specifically Deep Learning. And they are not the one who is thinking of the same Nervana (now Intel company) and others will also be in the race too. CPU to TPU is just denoting the innovations in infrastructure which accelerates innovation in Machine Learning. 

The forthcoming years or soon we will be witnessing the very definition confusion on Machine Learning, Data Science, and Artificial Intelligence too. The good news is as like in the braising process in cooking the end results are getting better and better over the years due to an emphasis on all the related fields.