FalconAI is a VC backed start-up developing cutting-edge AI in esports and fashion domains. Our team includes top AI researchers and engineers located in Boston and Istanbul. Our aim is to make human expertise accessible and scalable to democratize domain knowledge by developing cutting-edge AI.
As of today, we have two AI-backed products in fashion and esports domains, called FashionI and SenpAI, respectively. SenpAI is an AI platform that guides low/middle skill level players for improving their gameplay and reaching their goals in two competitive video games: Dota 2 and League of Legends (LOL). FashionI is a brand-new fashion app and chat-bot service to provide the next-generation shopping journey reinforced by AI. In this article, we briefly mention the technology behind FashionI.
With the collaboration of AI engineers and fashion designers in our team, we designed an AI solution specific to the fashion domain to provide personalized journeys among different ranges of products (see three examples of simplified journeys in Fig. 1). This system contains three different modules: feature extraction, outfit generation and active learning.
Fig. 1. Examples of simplified journeys. Depending on the main intention of the user, these journeys can suggest items from a very specific subset (Smart Search), teach the user about new and potentially interesting styles (Exploration Mode), and suggest potential new matches in accordance with their individual fashion taste (Discover Weekly).
In the feature extraction process, as a first step, we perform semantic segmentation for each image to detect the clothing items and classify every pixel in the image by using a hybrid structure of deep convolutional neural networks and various image processing techniques. After obtaining such masks from semantic segmentation, we extract features by a pipeline of different CNN structures and specialized image processing methods to mask the objects placed in front of the clothing items. Such feature sets contain both meaningful (e.g., color, length, pattern, etc..) and abstract (CNN) features to represent a clothing item.
In the second module, such features are utilized to generate outfits, i.e., coherent and complementary group of clothing items. In this kind of problem setup, supervised methods are not advantageous due to difficulties in obtaining well-distributed training datasets in the feature space. To avoid recommending the same type of complimentary group of clothing for each clothing item, we focus on unsupervised learning techniques to obtain sparse solutions. In addition, we designed an algorithm to capture the complex relationship between matching products. The parameters of our system are regularly tweaked to keep track of the current fashion trends as seen in fashion-related social media and blogs.
In the last module, we present a personalized journey for each customer via an interface where a customer can like or dislike a product and the corresponding shopping journey is devised based on the intention (e.g., finding a specific product or exploring new trends) of the user. We do consider our approach beyond traditional recommender systems because we do not only estimate the mostly likable products by the user but also design an adaptive order of products, constituting a seamless shopping journey for the user. This journey requires a real-time recommendation system that responds to the customers’ rapidly changing requests and instantly updates the rest of the journey according to the user selections. Such technology is currently utilized in our mobile app, FashionI, and it is provided as an API to e-commerce or retailer brands.
This last module includes various online learning methods including collaborative filtering and multi-armed bandit algorithms, as well as offline learning methods including a similarity search method that is a hybrid structure of a Siamese network, and different metric learning techniques. It also utilizes the aforementioned feature extraction module to learn the customers' style of clothing interactively and effectively by their likes and dislikes.
We can also recommend products based on the style and mood information of users based on their shared Pinterest, Instagram, and Spotify activity. For Pinterest and Instagram, our algorithm uses the last posts or pinned images of the users to learn their current style of clothing and Spotify provides a means for forecasting our customers’ mood according to their last listened songs and the corresponding playlists. This mood estimation mechanism classifies the current mood of the customers by a deep neural network. In this way, we provide an AI solution that not only learns the general style of clothing of the customers but also makes recommendations according to the customer's current mood. By learning their style and their current mood, we present users with their own personalized journey in the fashion space, which improves user experience in the fashion e-commerce domain.
Our system stores massive amounts of information about individual clothing items, and the users who like/dislike them. This information can be used to produce a snapshot of the current state of fashion as well. Utilizing a recurrent neural network (RNN), we are now working on the problem of predicting future trends. Specifically, we aim to pinpoint the features that are predicted to increase in importance in the near future. This helps us continuously find the most relevant new items for inclusion in FashionI.
FashionI is one of the prime examples of how AI can disrupt the traditional shopping experience in fashion ecommerce. We hope that our product will enable everyone to find great, fashionable clothing that expresses their personality in style.