People and patterns and predictions, oh my.
For marketers, a simple way to think about deep learning is that it’s ultimately about presenting customers with exactly what they want, whether or not they know yet that they want it. That could mean an experience, a bit of information, an ad, or a suggestion for a specific product. But what is deep learning?
Deep learning is a subset of artificial intelligence (AI) derived from the science of neural networks. And neural networks are simply an attempt to mimic the way scientists think our own brains process and make sense of the world. Basically, a neural network self optimizes its performance on a desired task based on exposure to structure and unstructured data.
I spy with my AI…
For example, let’s imagine we’re creating a deep learning based image recognition system designed to spot a product––a specifically branded can of soda—in photos posted on social media because we’d like to give a shout out, through our own social accounts, to the poster for their brand loyalty.
The first thing we would need to do is train the deep learning neural network using a number of verified positive and negative sources—e.g., photos containing said soda can, pre-tagged as a hit as well as photos with no can correspondingly tagged as a non-hit. Next, the system would be fed untagged positive and negative photos. The digital patterns in those photos would be compared to whatever digital patterns emerged from reviewing the initial guided positive and negative inputs.
If the system recognizes what it has determined is the pattern for, “branded can,” it marks that photo as a positive hit. At this stage, the system will require human feedback to determine whether that positive hit was, in fact, positive and whether other photos were falsely tagged as hits or non-hits. Each iteration, every data point, refines the neural network to better identify its proper target. And with data sets that span the internet, you can imagine how refined those algorithms can get.
But here’s the interesting part. Humans generally can’t read or understand those algorithms. We don’t know what the criteria the network is using, per se. We only know it’s getting better (or worse) at identifying the branded can. And there are plenty of times the technology fails completely, not to mention offensively.
How this “portrait” was made:
- generate random polygons
- feed them into a deep learning, neural net face detector
- mutate to increase recognition confidence until the neural net is reasonably sure it is “seeing” a face
A synthetic portrait “recognized” among random overlaid polygons by deep learning AI at here, here and here, deep learning AI will likely become increasingly pervasive in marketing and advertising. If you want a far more detailed and thorough primer on the topic, Stanford university has placed online an amazing guide to deep learning.