Is Fear of Missing out on AI keeping you up at Night?

Rajkumar Bondugula, Ph.D., Principal Data Scientist / Sr. Director, Equifax
99
151
29
Rajkumar Bondugula, Ph.D., Principal Data Scientist / Sr. Director, Equifax

Rajkumar Bondugula, Ph.D., Principal Data Scientist / Sr. Director, Equifax

Hardly a day goes by without reading a breathless headline about artificial intelligence (AI) and how this technology has a huge potential and how their organization is “involved” in this movement. However, one has to be careful to notice that most of the news is being generated by vendors of the technology and not necessarily by the beneficiaries of the technology. Like any promising technology, AI/Machine Learning (ML) is going through a huge hype cycle. In this article, I will provide some practical advice about how to think about AI, how to get started in AI and what to watch out for.

What is Artificial Intelligence?

Artificial Intelligence is an academic discipline that strives to make computers behave intelligently. There are multiple ways in which computers can acquire intelligence. One way is to explicitly code the knowledge and the processes that mimic the domain knowledge and analytical thinking of domain experts into system called “Expert Systems”. 1980s have witnessed many successful commercial deployments of AI in form of Expert Systems. Another way of acquiring intelligence is through algorithms implementing machine learning. While more formal definitions exist, intuitively, machine learning can be thought of “teaching computers to program themselves”.

Artificial General Intelligence (AGI), a system demonstrating intelligence in a broad range of domains has always been the dream of AI researchers. A system that broadly imitates a three year old child would be an example of AGI. However, most successful implementations of AI systems have been in narrow domains. Examples most people are familiar with include spam filters, recommendation engines, GPS routing systems etc. Most successful implementations have one or more of the following characteristics: 1) A human being is able to make a decision in less than five seconds (Ex: determining if an e-mail is spam) 2) The basis of decision making process can be, at least, roughly described by a human being based on their experience and can be approximated using data (ex: product sorting on e-commerce website where the objective is to sort the products in the order of predicted sales) 3) Human expertise can be approximated from data (ex: recommendation systems trained using large volumes of transaction data).

Where do I start with AI?

The best place to start with AI is to attend lectures and talks that focus on foundational concepts in AI. The objective is to gain clarity on what AI can do and what it currently cannot do. In addition, read articles that discuss different applications of AI in your industry and other industries, with the intent of identifying similarities in problems that have been successfully addressed with AI. With this broad understanding of AI fundamentals, look for areas in your organizations where AI can help. The challenge is that most people who are adept at developing AI do not possess domain knowledge and are unlikely to be aware of unmet, unarticulated and upcoming needs within the organization or the broader industry. That is why, the best people that bring right problems to AI developers are senior leadership, mid-level management and industry veterans who know what they want from AI solutions.

  Non-deep learning solutions are adequate for vast majority of problems that can be solved using AI.  

It is very tempting to start with a small or a toy problem first since laying the ground work for demonstrating AI solutions for a substantial problem takes too much time and effort. However, a solution to a small problem will hardly serve as evidence that AI can help the organization. It is better to start with developing a solution for substantial problem where the problem is well realized within organization, so that all efforts can be focused on demonstrating the solution to the organization. Also, AI proponents often believe that superior results are self-sufficient for rest of the organization to embrace AI. However, it usually takes active championing by influential people within the organization to make rest of the organization realize the potential of AI.

Another pitfall organizations typically fall into is to invest in large platforms early in the process of exploring AI solutions incurring both cost of the tool and the cost of integration without fully understanding value the platform brings or if the same value can be achieved by cheaper and alternative means. Almost all major algorithms for implementing AI solutions are available as open source packages. If the size of the data that needs to be analyzed for developing AI solutions is not big, then a machine learning toolbox like Weka, along with expertise in a general purpose programming language like Python and decent computer is all you need to start. However, if the underlying data that needs analyzing is large enough to be held in a cluster, then Spark libraries accessed through Scala or Python API are all you need to start. If your organization did not already invest in a cluster infrastructure, platforms like AWS can be considered. These platforms give you the ability to try out AI without committing to a large scale platform. In addition, all major “AI platform as a service” providers work with small companies that understand AI platforms and these companies can be hired by your organization to get going quickly.

Finally, companies believe that they need deep learning to start with AI. While it is always great to be aware of the latest developments in AI, you don’t need deep learning unless your solution needs to build on information from local regions in images, temporal dependencies, or on multiple levels of abstractions etc. Non-deep learning solutions are adequate for vast majority of problems that can be solved using AI. Building deep learning solutions from scratch require huge data and computational resources and therefore should only be entertained when all non-deep learning solutions have been deemed insufficient.

In summary, invest enough time to understand what AI can do, specifically for your organization and demonstrate the power of AI by developing a solution for a non-trivial problem. It might be pragmatic to be confident of the value AI can bring to your organization before investing in large platforms. Finally, realize that not all problems need deep learning solutions.

Read Also

Getting the Most out of Big Data

Getting the Most out of Big Data

Kris Rowley, Chief Data Officer, GSA
Big Data: Separating the Hype from Reality in Corporate Culture

Big Data: Separating the Hype from Reality in Corporate Culture
Brett MacLaren, VP, Enterprise Analytics, Sharp HealthCare
Maintaining Maximum Relevancy for Buyers and Sellers

Maintaining Maximum Relevancy for Buyers and Sellers

Zoher Karu, Vice President and Chief Data Officer, eBay
Building Levies to Manage Data Flood

Building Levies to Manage Data Flood
Adam Bowen, World Wild Lead of Innovation, Delphix