It is still impossible to develop a system based only on a neural network, according to Oleksandr Savsunenko, the head of AI Lab at Skylum Software. On November 14, the expert will speak at AI Conference Kyiv on the topic: AI for visual content. How it works and where it is used.
In an interview with AI Conference Kyiv, the specialist talks about his machine learning projects and reveals how AI is used to improve the quality of photos.
Interviewer: AI Conference Kyiv (AICK).
Respondent: Oleksandr Savsunenko, Head of AI Lab at Skylum Software (OS).
AICK: Tell us about your ML project – Let’s Enhance. How did you improve the quality of photos and what challenges did you face during the project?
OS: To train the neural network, we used high-quality images that we turned into compressed and noisy ones. The neural network’s task was to restore a picture in good quality. The result was successful: the neural network learned how to remove pixilation, compression artifacts, and other defects.
The most complex was the production support of this system. After the publication of articles about our service on TechCrunch and Mashable, we obtained a huge amount of traffic and had to process about 200 000 images in 24 hours.
AICK: Let’s Enhance 2.0 was launched last year. How does it differ from the previous one and how does the service look like today?
OS: The second version of Let’s Enhance has different training methods, loss feature, and network architecture. These aspects should be constantly changed in order to enhance the quality of products.
I left Let’s Enhance.io, the startup entered the Techstars London project in July 2018, attracted investments, monetized, and became profitable.
AICK: Share the most interesting and remarkable AI solutions that you were involved in.
I worked on one of the first nutrigenetics projects – Titanovo. Using ML, my colleagues and I were trying to predict physiological predisposition based on the whole genome analysis and chip genotyping.
We collected information from articles, created datasets, and taught models. Based on them, we prepared recommendations for people and predicted their health condition in the future.
AICK: What functions do neural networks perform in photo editing services, which you are working on now?
OS: AI performs the main role in Photolemur by Skylum Software: the program improves images in a single click using machine vision.
Initially, the software recognizes a type of image as well as detects people, buildings, and the time of day. Afterwards, it segments the image and emphasizes corresponding zones.
The image is improved without AI. It is conducted by stored algorithms as photos are processed by photographers.
AICK: What databases and algorithms were used to teach this system?
OS: As to the framework for neural network development, I prefer an uncommon but becoming popular MXNet.
The key advantage is the speed of calculations. I cannot name datasets and algorithms because it is a commercial secret.
AICK: What challenges did you face while designing an intelligent graphics editor?
OS: If an image is complicated, neural networks are frequently mistaken, recognizing patterns and making segmentation. Therefore, we had to conduct the analysis of results and correct mistakes, using conventional methods and standard algorithms.
There is still no system performing everything using just a neural network.
AICK: Tell us about the audience of these products. Do you manage to entice Adobe users?
OS: Luminar is an alternative to Adobe Lightroom. New technologies are integrated into the product faster due to the small and solid team. Besides, we always gain new users.
Luminar is an appropriate solution for both beginning and professional photographers: the product includes tools for editing images in a single click and a set of features allowing to process photos more thoroughly.
Photolemur was developed a little more than a year ago. This product is actively used, as its target audience is people who do not want to explore Photoshop. It is designed for those desiring to make their photos wonderful.
AICK: Describe your projects evolving neural networks in order to optimize landing pages.
OS: It is a common task when one requires A/B testing of a landing page. Generating separate pages for all possible options, we can obtain millions of versions as the result. To achieve a meaningful result, one should conduct pairwise A/B testing of all versions: it needs a huge amount of traffic. Such major testing is suited only for large companies.
If a small company is going to check a great deal of options, it can carry out A/B testing using neural networks based on reinforcement learning. In this case, the neural network’s goal is to fill the page with elements and increase its conversion. Thus, the neural network will be taught during the traffic flow. As the result, the neural network will find the best possible landing version faster.
When complicating, AI can learn how to show options of landing pages customized to certain users. They will see a page shown by the neural network. Such an approach requires less traffic. Therefore, the page will provide the good results more rapidly.
Discover more about artificial intelligence and machine learning from the expert at the conference!
You'll know more than your colleagues and business rivals.