In simple terms deep learning is a method of abstraction through which an algorithm solves a complex problem with little or no instruction. Basically you can throw some tough problems at the algorithm and it will spill out the answers. Of course it has its limitations and the quality of answers depends on several factors but as a marketing specialist, there are reasons why you should pay attention.
To understand the impact that deep learning can have on ecommerce and marketing, let us take a specific example. For example, let’s say Jane is visiting an online store and is looking at a brown women’s handbag. She likes it but it is not exactly the type of handbag she is looking for. She is looking for something, well … prettier. Yes, there are the best selling item recommendations and highest rated products but these are not necessarily items that Jane would consider pretty.
Luckily for Jane, she can click on the image of the bag and get visually similar recommendations. These recommendations have been made based on analyzing the image of the bag that Jane clicked. As a result, the algorithm has recommended Jane images of other handbags with brownish color and similar shape. For Jane it is a simple way of discovering pretty bags to buy.
Behind the scenes it is deep learning that is deciding what type of bags to suggest Jane. On an abstract level the network works as an input-output system. You give the network some inputs and it produces outputs. The outputs are predictions made on the inputs. In this case the input is the image of the brown handbag that Jane is clicking.
More specifically the input is the collection of pixels in the image. The picture is broken down into many small fragments of the picture called pixels. As the original image is 200 pixels wide and 200 pixels high, it contains 40,000 pixels. Each of those 40,000 pixels serves as an input into the network. Inside the network an algorithm analyzes each pixel in the image of the bag and as an output, the network attempts to recreate a picture similar to the input picture. A more detailed overview of the way in which this happens can be found in this excellent ebook.
You might ask why would the network be trying to create something as an output that we inserted as input? The reason is that we, humans, understand that the input is a brown handbag and the output is a handbag. However, the machine lacks our intelligence and the network only “sees” the image first time when we put it into the network as an input. Essentially the network is the visual sensory system for the machine. The network is to the machine what the eyes and the brain are for us.
So based on analyzing all the pixels on the image of the bag, the network has created an algorithm by which it can assess what kind of a bag it is that is on the image. At first, the algorithm is rather poor as it produces bad output images. However, when we feed the network tens of thousands of images, it modifies the algorithm and in the end can do a good job in identifying the images.
To be able to provide the necessary outputs the network is trained with large datasets and recent advances have sped up the process of optimizing the network. As such it will be easily able to tell a difference between a small red handbag and a backpack. So when Jane clicks on the image of a brown handbag, the network recognizes the image and recommends Jane similar images because the network knows what images are similar to the image of the handbag that Jane clicked. And this does not only work on images of handbags. It can be literally anything — a jacket, a shoe, a table or sunglasses.
The are many benefits to such recommendation. Users like Jane can now find products that they likes more easily. All they have to do is click on the images of the bags and they are served similar images of handbags. It is a powerful way of discovering products and becomes increasingly more and more relevant as efficient conversions on mobile displays become more important.
As a marketer this provides an opportunity to increase your marketing conversion rates as visitors like Jane are more likely to discover fashion products that they like and complete their purchase. It is a clever form of personalization as the recommendations are optimized based on the behavior of the consumer. As results for each click are different, each recommendation is personal as it depends on what the person has clicked. This type of personalization brings opportunities for targeted personalized content that so far has been difficult to do.