September 30, 2020

Lean AI: How “Much” AI Does Your Company Need?

Being humans, we have a tendency to anthropomorphize objects around us. Thus, in the 20th century many technology inventions carried the names of people who invented them, like Ford cars and Boeing airplanes. At the same time, pet owners love nothing more than attributing human characteristics to the behavior of dogs, and who doesn’t love the talking animals in Disney movies? Even the God’s of Greek, Chinese, and Egyptian mythology resemble humans.

This trend holds true for AI, the field of
creating an electronic mind that knows no fatigue or stress. Whatever huge promises
that the technology gives us, we often unconsciously expect more from AI than
it can do.

Many companies have experimented with AI adoption, but have found it challenging to prove the value of AI solutions. In its report Three Barriers to AI Adoption, Gartner names a lack of skills, fear of the unknown, and data quality as three most common ones. And surely there are others reasons also – one may be the high expectations on the technology’s ability to be, in HR terms, a “self-starter” and a “team player”. Companies should not take what they desire for reality and take full control on AI adoption in their hands.

The hype surrounding AI has certainly put it in the spotlight—companies understand the benefits that an artificial mind can offer. But at the same time, they need to know more about how to put it in practice. This article outlines three recommendations on AI adoption:

1)      Ensure data quality and volume. Data has to go a long way from entering the storage system to creating value. Data discovery, tagging, and organizing are tedious activities, especially when today’s expectations center on fast innovation.

However, one kilobyte of structured data may bring way more value to an enterprise than terabytes of data that are messy. Ironically, when companies are drowning in data, they tend not to have enough of it. Modern enterprises are likely to find that they lack some pieces of information that they need to navigate in a changing environment. For example, a supermarket may not know what customers are searching online, but this potentially represents an opportunity for the retailer to grow sales. Companies need to take responsibility to put their data in order and evaluate what information is missing to the full picture. A systematic approach to data collection and cataloging is the cornerstone of the overall enterprise data initiative.

2)     Find the balance between predictive accuracy and required computing power. AI is basically a series of math problems, but the number of problems grows at an accelerating rate as the precision of the model increases. Complex AI models can consume a tremendous amount of computing power, resulting in mounting costs. At the same time, increasing model complexity means that incremental improvements in algorithm efficiency become smaller with every step. For example, in a recommendation system that isn’t critical, it might be the case that implementing a model with 50% accuracy can pay off better than a beefier model with 90% accuracy, as the costs of hardware will be a dimension lower.

Another reason for the increasing demand for massive computing power is the increasing amount of data. For example, by doubling the resolution of a picture, the number of pixels also doubles. Certain calculus calculations, however, can show how much AI is enough for the organization.  This is a matter of trade off, and every company should decide where the sweet spot is.  Hardware can be costly for an enterprise upfront, but with the salaries of experts and the cost of power consumption, early adopters who decided to implement full blown machine learning initiative may lose their motivation very fast. AI is a real thing but it requires some thoughtful consideration.

Source: Two Distinct Eras in Compute Usage in Training AI Systems

3)       Companies should start with perceptual AI – natural language processing and computer vision, the most mature technologies in the AI domain. While AI has a broad scope of applications, NLP and computer vision are well understood largely due to their implementation in smart city scenarios around the world. Many companies invest in AI to explore its possibilities. And intelligent camera recognition, document scanning, automatic video, and image tagging should be some of the first applications on an IT department’s AI agendas. The industry has accumulated a lot of experience in this domain, so enterprises just need to build on industry blueprints, best practices, and previous experiences. In this area, the threshold for implementation is the lowest.

While the potential of the technology is very
strong, decision-makers should take full responsibility for AI adoption.

The best approach is some
rational thinking on how AI can bring positive changes and then developing a
step-by-step plan.

Read More

Leave a Reply

Your email address will not be published. Required fields are marked *