|Giovanni Strocchi, @gstrocc, CEO ADmantX|
Have you ever heard of deep learning? Earlier this month, re/code posted an interesting article that goes in depth on what they’re billing as “tech’s latest craze.” It defines deep learning as “the subset of artificial intelligence focused on teaching machines to find and classify patterns in mass quantities of data.”
If you haven’t noticed, AI is popular again, most likely thanks to increasing interest shown by Google, Facebook and IBM (Watson). As the article highlights, three of the field’s leading scientists have joined those companies over the past few years: Geoffrey Hinton joined Google to expand their deep learning division in 2013; Facebook appointed Yann LeCun to head AI later the same year; recently, IBM announced that it is working with Yoshua Bengio to infuse its Watson supercomputer with deep learning.
What we’re seeing today seems to confirm what the AI experts have been saying all along: soon, even the products that we all use on a daily basis will have an AI backbone.
In fact, it’s already happening. Digital advertising is already using statistical models and techniques in line with AI and machine learning. Audience profiling uses a range of techniques, from neural networks to “black-box” systems that learn from data, such as those described in the article. There are many techniques, and they can be combined in a variety of ways.
From more “computational intensive” tools to those that seek to derive more information that can be interpreted from a strategic point of view before starting a discovery and optimization process (fundamental in the world of online advertising), this is essentially advanced audience profiling based on semantic technology. The difference between many audience profiling based on neural networks, and our advanced audience profiling based on semantics is less about the technical aspects and more about its vision, even if the solid consumer’s interests understanding that come from our unique “human like” understanding of the consumed media gives a unique foundation to our data analysis.
Automatic optimization and recommendation processes can be supported with similar tools, but differs in two important ways:
1. Without a logical model that supports your business processes, input data can be informationally weak. Therefore, your planning must include a phase where you ensure that the initial data is meaningful relative to your objectives.
2. When the goal is data management, being able to interpret data (for example combining statistics and machine learning algorithms to look more deeply into user behavior, or leveraging semantics to enrich the analysis of the set of data statistics analyzes to define propensity to clicking) will go much further than strict data optimization in providing a strategic understanding of phenomena.
In practice, it’s not enough to simply plug the data into a powerful algorithm. You must also establish a process that integrates business experience and needs, methodologies, algorithms and technologies to generate content with high informational value that meets your goals for information optimization (which often flows automatically), as well as goals for comprehension and decision making.
A powerful algorithm is important, but it’s not everything. A winning approach is one that is more holistic, which requires serious reflection of business aspects in both the input and the output processes.