Factors Enabling the Breakthrough of Artificial Intelligence – Part 2

It is a common thing these days to hear things about Artificial Intelligence (AI). But what are the factors that have enabled AI to become a household name that it is today? These are namely Compute, Data, Tools, and Momentum. Let us take a detailed look at each of them.

Quality of Tools

If you’re using Python, there is a wide array of easy to use tools at your disposal. For example, Jupyter notebooks allow researchers to write and run a code, and generate output directly within a simple browser-based interface. Additionally, you have access to mathematical frameworks (numpy), data manipulation libraries (pandas), and user-friendly implementations of standard scientific and machine learning algorithms (SciPy, Scikit-Learn). There are a large number of libraries that allow a researcher to quickly and easily visualize data, for example, matplotlib. Frameworks like Tensorflow, Keras, PyTorch etc. exist for creating and working with neural networks. Support frameworks such as OpenAI Gym provide interactive 3D environments and video game interfaces. All of these frameworks are being actively improved, and because they are all open source, they’ve been built to interoperate with, and reinforce each other. Tooling is having a great impact on AI-related research.

Momentum

Over the last few years, proven results and the exponential increase in new avenues of research in the field of machine learning have led to increased interest and investment in the field. Take for example an approximate 12,000 new papers that are submitted to arxiv.org (one of many repositories for scientific papers) every month. Researchers are attempting to creatively apply previously proposed techniques in the field of machine learning to new problems, such as the application of neural networks to problems such as language translation and image recognition. Researchers at Uber recently demonstrated that genetic algorithms could be used to create models capable of playing Atari video games. These models had previously been created using other techniques such as Q-tables and Deep Q-Networks.