Researchers at Skoltech and AIRI have developed a new algorithm for optimal data transfer between domains using neural networks.

Machine Learning


Since the advent of large-scale OT and Wasserstein GANs, machine learning has increasingly embraced using neural networks to solve optimal transport (OT) problems. It has recently been shown that OT plans can be used as generative models with comparable performance on real tasks. OT costs are often calculated and used as a loss function to update generators in generative models.

The Artificial Intelligence Research Institute (AIRI) and Skoltech have collaborated on a new algorithm to optimize information sharing across disciplines using neural networks. The theoretical underpinnings of the algorithm make the output easier to understand than competing methods. Unlike other approaches that require combined training datasets as in the input-output example, the new approach may be trained on separate datasets from the input and output domains.

Large training datasets are difficult to come by, but necessary for modern machine learning models built for applications such as facial and speech recognition and medical image analysis. This is why scientists and engineers often resort to using artificial means to simulate real-world data sets. Recent advances in generative models have dramatically improved the quality of generated text and images, making this task much easier.

🚀 Join the fastest ML Subreddit community

A neural network is taught to generalize and extend from paired training samples and input and output image sets to new incoming images. This is useful for jobs that require processing many identical photos of varying quality. In other words, generative models facilitate the transition from one domain to another by synthesizing data from dissimilar data. For example, neural networks can transform hand-drawn drawings into digital images or improve the clarity of satellite photos.

Matching probability distributions with deterministic and probabilistic transport maps is a unique use of this popular tool technique. This method enhances existing models for domains other than unpaired translation (image restoration, domain adaptability, etc.). This approach allows more control over the level of diversity of the generated samples and improves the interpretability of the learned maps compared to common methods based on GANs or diffusion models. OT maps obtained by researchers may need to be corrected for unpaired activity. The researchers highlight the design of transport costs for specific tasks as a potential research area.

The intersection of optimal transport and generative learning lies at the heart of the chosen approach. Generative models and efficient transports are widely used in fields such as entertainment, design, computer graphics, and rendering. Some problems in the aforementioned sectors may be suitable for this approach. A possible downside is that some professionals in the graphics business may be affected by the use of previous tools that made image processing technology generally available.

Researchers are often forced to work with irrelevant data sets rather than ideally matched data due to prohibitive costs or difficulty in obtaining them. The team went back to the work of Soviet mathematician and economist Leonid Kantorovich and used his ideas on efficient freight transport (optimal transport theory) to find optimal data transfer between domains. We have developed a new way to plan. Neural Optimal transport is a new approach using deep neural networks and separate datasets.

When evaluated on unpaired domain transfers, our algorithm achieves better results than state-of-the-art approaches in image styling and other tasks. In addition, it requires fewer hyperparameters that are typically difficult to tune, results are easier to interpret, and are based on a more sound mathematical foundation than competing methods.

check out paper and github. All credit for this research goes to the researchers of this project.Also, don’t forget to participate Our 18k+ ML SubReddit, cacophony channeland email newsletterWe share the latest AI research news, cool AI projects, and more.

Dhanshree Shenwai is a Computer Science Engineer with a keen interest in AI applications and strong experience in FinTech companies covering the domains of Finance, Cards & Payments and Banking. She is passionate about exploring new technologies and advancements in today’s evolving world to make life easier for everyone.

🔥 Must read – What is AI hallucinations? The problem with AI chatbots How to find hallucinatory artificial intelligence?

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *