Real world Machine Learning in Ret[AI]l
The future of Ret[AI]l cannot be envisioned without thinking of AI as an integral part of it! There have been many trends that changed the landscape of retail, here are some of the ones that we would have experienced in one form or the other.
- AI Chatbots for customer interaction: IBM conducted a survey that showed 60% percent of millennial demographic prefer interacting with bots vs talking to human beings. Top retail brands like Burberry, Sephora, and 1800-Flowers use chatbots to connect with customers to provide an immersive experience leveraging social platforms like (and not just limited to) Facebook, WhatsApp, WeChat, Slack etc. AI Chatbots leverage Machine Learning algorithms that uses Natural Language Processing (NLP) as a foundation block.
- Smarter Inventory Management: A typical retailer loses four percent of total sales due to items being out of stock according to a Harvard Buiness Review (HBR) study. Getting real time status updates in the retail supply chain enables proactive inventory planning based on demand. On a related note, market basket analysis helps which product gets a boost when adjacent (or related) products go on sale.
- Image Recognition: Improvements in image recognition using Deep Learning has enabled novel ways of interacting with the consumer. Neiman Marcus launched a Snap.Find.Shop feature that allows consumers to take a picture of a product and find matching products from the store inventory.
- Customer churn and Retention: It is a challenging task to win a customer back once you lose them in retail. AI helps to predict customer churn and uses rewards and incentives to retain their business
- Enhanced shopping using Augmented Reality: There is an increasing pressure for the brick and mortar stores to stay relevant and competitive with the online shopping giants. One of the ways to provide the niche experience of shopping in a store is to provide enhanced shopping experiences. Neiman Marcus launched a compelling feature in their store to improve the customer shopping experience using their Memory Mirror. It allows to take a 360 degree video of the customer and helps them to visually compare different outfits they have tried side by side (it remembers all the different trials), customize using different colors without physically trying them on, and optionally shared on social media. The retailer has also launched a smart mirror used in their cosmetics department for enhanced AR experience for customers shopping for cosmetic products.
- Smart Product Placement: Retailers have been able to increase product sale by monitoring the foot traffic patterns and identifying how much time customer spends at different areas of the store which also takes into other factors like time of the year, seasonal events, weather, and time of the day into account. The in store cameras track the customer movement in the store and the data collected over time (as shown in the picture below) helps to tune the product pricing and placement. Such heat maps have shown to improve product sales from as low as 5% to as high as 20% percent.
We will not be covering each of these trends in this article but will review two interesting examples that leverage cutting edge AI technology.
Next Generation Fashion using Generative Adversarial Networks (GAN)
Human beings can easily recognize different classes of visual images. For example, even a toddler can distinguish a bag from a shoe. Humans are also good at finding associations between different classes of images. For instance, it is easy for someone to pick a shoe or a sports coat to go with a pair of jeans. But this is a complex problem for a computer program to solve. This has been addressed typically using Supervised Learning and labeled training data. Obtaining a corpus of labeled training data is a time consuming process. Recent research has enabled discovering relations between two visual domains without any labeling or explicit pairing step. The example we show here is referring to a model that “discovers cross-domain relations with GANs” or DiscoGAN for short.
We have discussed GANs in previous articles. A GAN typically consists of a Generator and a Discriminator. The DiscoGAN model here uses 6 different neural networks, 4 Generator (GAB and GBA for xA dataset and GBA and GAB for xB dataset) and 2 Discriminator ones (DA and DB). The details of the implementation can be seen in the paper listed in the references.
The results of the model are compelling.
- Image (a) shows how the model captures the specific features of each domain.
- Image (b) shows the style transferred from handbags to shows
- Image (c) shows the style transferred from shoes to handbags
The team also were able to train the model to map from sketches to a finished product (and vice versa). The results are shown in the image below.
The ability to go from sketches to the end product and being able to see in in such a flexible way will help to reduce the time to market for new products and services being released to the consumers.
Another interesting application of GANs is to use them for personalized fashion recommendation.
In this case, the model can build new outfit designs by taking in a couple of inputs. The first one is a natural language query (for example: a purple dress in the above illustration)that are submitted by user. The second input is a training data set of existing clothing. The output is a personalized fashion recommendation tailored (pun intended!) to user preferences.
Smart Retail and self-checkout
Amazon Go released its pilot store that enables customers to buy products from a brick and mortar store without checkout, thus creating a seamless shopping experiences without any long lines or waiting. The advances in image recognition using Convolutional Neural Networks (CNNs) and Long Short Term Memory (LSTM) algorithms enable detection of different products with high accuracy. These images show how the algorithm is able to show dynamic inventory and label products with a confidence factor in such automated brick and mortar shopping experiences.
One concern with self-checkout stores is to identify and avoid shop lifting scenarios. The image recognition algorithm (specifically referred to as Action Recognition) identifies moving objects dynamically. Once the object recognition is done successfully, the program can be enhanced with features to alert the shop staff.
Conclusion
We looked at a two examples which use state of the art Image Recognition using one or more variations of GAN, CNN, or LSTM architectures to build compelling solutions in Retail sector.
