A Brief History Of Machine Learning
There has never before been as much interest in machine learning as there is now. This branch of computer science develops algorithms that allow computers to become smarter. Today, it powers modern marvels like the self driving car, Feedzai’s intelligent fraud fighting platform and Amazon’s voice activated assistant Alexa.
As part of a May 26 webinar on Machine Learning in Financial Services, Pedro Bizarro, Co-Founder and Chief Science Officer at Feedzai, spoke about the key turning points in machine learning history that have led it to the prominence and impact it’s having today.
Pedro drew on his background as a computer scientist and researcher to briefly explain history of machine learning evolution. Pointing out that roots of machine learning go as far back as the 1950s, he identified three major trends that made the current explosion of machine learning innovation possible:
Trend 1: Cheap parallel computing and storage
One shift with major implications for all fields of computer science was the availability of affordable and scalable parallel computing and storage. This revolution took off in 2004 when Google produced a research paper on MapReduce. The company used this programming model to generate and process terabytes of data split between as many as thousands of commodity computers. In 2011 Hadoop, an open-source public version of MapReduce made great strides by making this widely available to the market.
In the meantime, similarly, innovations in graphics processing units made it feasible to perform previously unheard of calculations in an efficient manner. Designed by manufacturers like Nvidia for the increasingly complex demands of 3-D gaming, the new GPUs were repurposed to allow parallel machine learning tasks on relatively cheap hardware.
In addition to cheap parallel computing, there was also progress on storage capacities and speed of data retrieval. With the open source database management system Cassandra available in 2008, NoSQL databases increasingly displaced the relational databases. Based on this design, it became practical to store terabytes of data scaled to clusters of machines and retrieve that information in milliseconds.
In 2014, the open source computing framework Spark further pushed forward the possibilities for programming clusters, offering an interface specifically designed for data parallelism and streaming analytics.
Trend 2: Big data
According to study done by IBM, 90% of the world’s data was generated in the last two years. In 2016, a third of online payment transactions will happen on mobile. The lines between offline and online are also blurring, what was done offline before is now moving online. Everything is measurable, from explicitly shared data such as social network information (Twitter, LinkedIn, Facebook, Youtube, etc.) to implicit data collected via mobile or other wearable devices – location, accelerometer data, heart rate, footsteps, and much more.
A big challenge of previous machine learning initiatives was the sparsity of data, a problem that no longer exists. AI systems are now able to automatically understand an image of people playing football on field, because they’ve been able to be trained by millions of such images. Fraud patterns can be thwarted more easily because of the increasing amount of historical profile data for individual customers, merchants and even point of sale devices.
Trend 3: Better and more accessible machine learning algorithms
Over time, the algorithms involved in machine learning improved and became more widely accessible. In 2001, Leo Breiman and Adele Cutler presented their trademarked version of the random decision forests ensemble learning method which they called Random Forest. This was a far speedier improvement over the widely used decision tree model that also made it easier to clean up the resulting datasets.
In 2006, Geoffrey Hinton and his collaborators garnered attention for their work in the deep learning movement. This school of thought is concerned with creating neural networks based on layers of nonlinear processing units. New hardware made this powerful approach practical, if still expensive, and the algorithms are now used in applications like computer vision and speech recognition.
Open-source options, like the generalized boosting model package or the scikit-learn data analysis tools for Python, made it possible for researchers everywhere to experiment with machine learning algorithms. From 2007 on, the field leaped forward, with the ongoing refinements embraced by a broad range of private enterprises.
Along with other advances in technology and shifts in business practices, the work of computer scientists set the stage for today’s proliferation of machine learning. Financial services, banking, retail and insurance organizations alike increasingly realize the tremendous value of data and analytics. More than the latest tech trend or buzzword, the explosion of machine learning points the way for the future of big data computing and informed business strategy alike.