Optimum transport (OT) problems have been increasingly tackled using neural networks in machine learning. Large-scale OT and Wasserstein GANs have paved the way for using neural networks as a generative model with comparable performance in real tasks. The OT cost is commonly used as the loss function to update the generator in generative models.
The Artificial Intelligence Research Institute (AIRI) and Skoltech have collaborated on a new algorithm that optimizes information sharing across disciplines using neural networks. This algorithm stands out from competing methods due to its easily understood output, thanks to its strong theoretical foundations. Unlike other approaches that require coupled training datasets, the novel algorithm can be trained on separate datasets from the input and output domains.
Obtaining large training datasets can be challenging, especially for applications such as face or speech recognition and medical image analysis. To overcome this, scientists and engineers often rely on simulating real-world datasets. Generative models have greatly improved the quality of generated text and images, making it easier to simulate data for training machine learning models.
Generative models allow neural networks to generalize and extend from paired training samples and input-output picture sets to new incoming images. This is especially useful for tasks that involve processing multiple similar images of varying quality. Generative models facilitate the transition from one domain to another by synthesizing data from different sources. For example, a neural network can convert a hand-drawn sketch into a digital image or enhance the clarity of a satellite photo.
The novel algorithm proposed by AIRI and Skoltech involves aligning probability distributions with deterministic and stochastic transport maps. This unique approach has applications beyond unpaired translation, including picture restoration and domain adaptability. The algorithm provides more control over the variety of produced samples and offers improved interpretability compared to traditional methods based on GANs or diffusion models. Further research is needed to explore the design of transportation cost for specific tasks.
The intersection of optimum transport and generative learning is at the core of this approach. Generative models and efficient transport are extensively used in fields such as entertainment, design, computer graphics, and rendering. The proposed algorithm may have implications for these industries, as it allows for the development of publicly available image processing technologies.
The team drew inspiration from the work of Soviet mathematician and economist Leonid Kantorovich, specifically his optimal transport theory, to develop this novel approach. By utilizing deep neural networks and separate datasets, the neural optimal transport method offers a more interpretable result and requires fewer hyperparameters compared to state-of-the-art approaches.
To learn more about this research and access the paper and Github repository, check out the links below. Join our ML subreddit, Discord channel, and email newsletter to stay updated on the latest AI research, cool projects, and more.
(End of article)
Paper: [Link to the research paper]
Github: [Link to the Github repository]