About 700 million people around the world send each other gifs through texts, messaging applications, and social media every day. With all that messaging, people flocked to GIPHY, an online database and search engine for animated images. Using GIPHY, individuals can quickly and easily search for and share gif files — think Google — and share to platforms like Facebook. On May 17th, 2020 Facebook announced the acquisition of GIPHY. GIPHY will be fully integrated into Instagram, allowing IG users to easily find gifs and stickers for their stories and personal messages.
GIPHY was founded in 2013. GIPHY is one of the largest online databases of gifs, containing billions of GIFs. In addition to gif search, GIPHY supports the GIPHY Arcade gaming platform, a video section of gifs with sound (GIPHY Video), and integrations with Facebook, Instagram, and TikTok.
In 2017, GIPHY partnered with Proxet to develop different parts of GIPHY’s API, the internal system for managing application clients, search engine optimization algorithms, recognition of celebrity faces, and other technological solutions.
Proxet has a team of ten engineers dedicated to GIPHY— four in machine learning, two PHP specialists, two in Scala, one in Python, and one in JavaScript. They work on one of the most important areas for GIPHY—integrations of the service into other platforms. GIPHY and Proxet teams gather in the same office from time to time to synchronize and work together on the technical part of the solutions
GIPHY is a search engine, like Google, but users search differently on GIPHY than on Google. Google searches are for unknown information —”what is the diameter of the earth?” might be common. GIPHY searches are for gifs that capture a feeling or message —“hungry” or “hello” are popular. GIPHY search uses artificial intelligence algorithms to select the most suitable gifs. Proxet creates and optimizes custom models that provide the most accurate search for GIPHY.
gifs are frequently animated images of pop culture references and memes. Because the specific meaning and relevance of gifs are given by social groups, trends, and common language usage, it is often difficult to accurately describe a gifs when loading them into the database. Inaccurate or imprecise descriptions lead to irrelevant search results. When users do GIPHY searches of celebrities' names, they expect to get gifs associated with those celebrities.
Proxet developed a system that uses ready-made data from GIPHY and other sources. When users do GIPHY searches on celebrities’ names, the system presents the best matched gifs to the user; if the user selects one of the gifs, the system tags the selected gif with the celebrity name or names put in the GIPHY search. The system then searches the entire GIPHY database for gifs with the same celebrities using image recognition and tags those gifs with the celebrity name or names as well.
Proxet’s system is part of GIPHY Celebrity Detector, a machine learning model that identifies celebrity faces in gifs. GIPHY released Celebrity Detector in 2019.
Dmitry Voitekh, a Machine Learning engineer at Proxet, explained the capabilities and development of Celebrity Detector: “GIPHY Celebrity Detector can detect over 2,300 celebrities in gifs with 98% accuracy. First, it searches the gif for the face, then matches the face to a possible celebrity. The first step involves the analysis of all gif image frames. The second step uses the ResNet neural network, which is trained on millions of people from its own dataset of images. Celebrity Detector creates informative vector representations (so-called embeddings) for celebrities. These vector representations generalize and condense multiple images of a celebrity, sort of like how initials generalize and condense a name. Celebrity Detector reduces gifs to vector representations, then uses a clustering algorithm to compare the vector representations to samples where the celebrities have already been identified. Our team worked on every stage of Celebrity Detector’s development, from data collection, to model building, to implementation and support within the GIPHY platform."
GIF Recommendation Platform
A goal of any content website, or for that matter, most businesses, is to build long-term user engagement through return visits. To achieve this, Proxet created a recommendation platform that offers users similar gifs depending on their interests. This platform makes 30 million recommendations a day.
Using machine learning algorithms, the recommendation platform studies user interactions with gifs to “understand” user preferences and the meaning of those gifs. The platform then uses those preferences and gif meanings to offer users similar gifs they may like.
“Tagging is one of the main parts of GIPHY engine. When a user searches for gifs, the user is actually looking for tags for this gif. Automated marking can quickly sort and tag huge numbers of gifs; automated marking then assigns these gifs to the appropriate ‘groups’—gifs sorted by certain criteria, such as ‘laughing’ or happy birthday’— for analysis,” comments Igor Kroosh, a machine learning engineer.
Proxet designed the recommendation platform to grow with GIPHY and the GIPHY user base and gif set. The platform includes lifelong learning technology that updates the platform to automatically process new data.
.webp)
