Follow our latest advances in machine learning.

Generative models

Computer vision

Distributional shift

Graph machine learning

Nearest neighbor search

Uncertainty estimation 

Machine learning theory

Natural language processing 

Tabular data

Bayesian methods

Gradient boosting

Large-scale machine learning

Blog

Yandex Research Blog

large_dfs_16_9.png

medium_dfs_16_9.png

small_dfs_16_9.png

thumbnail_dfs_16_9.png

dfs_16_9.png

Neural algorithmic reasoning focuses on building models that can execute classic algorithms. It allows one to combine the advantages of neural networks, such as handling raw and noisy input data, with theoretical guarantees and strong generalization of algorithms. Assuming we have a neural network capable of solving a classic algorithmic task, we can incorporate it into a more complex pipeline and train end-to-end. For instance, if we have a neural solver aligned to the shortest path problem, it can be used as a building block for a routing system that accounts for complex and dynamically changing traffic conditions. In our work [ref1], we study algorithmic reasoners trained only from input-output pairs, in contrast to current state-of-the-art approaches that utilize the trajectory of a given algorithm. We propose several architectural modifications and demonstrate how standard contrastive learning techniques can regularize intermediate computations of the models without appealing to any predefined algorithm’s trajectory. 

Neural algorithmic reasoning without intermediate supervision

Yandex Research team regularly contributes to the computer vision research community, mostly in the field of image retrieval and generative modelling.

Generative models in computer vision are powerful tool for various applications.

Diffusion models are currently the most powerful approach to generating images from a text description. One popular application is the adaptation of such models to new concepts: usually, it is quite difficult or even impossible to write a prompt that accurately describes the target object.

At the same time, existing methods are too slow for practical purposes. In “Is This Loss Informative? Faster Text-to-Image Customization by Tracking Objective Dynamics”, our paper accepted at NeurIPS 2023, we propose DVAR  — a lightweight early stopping criterion for gradient-based adaptation methods. Our experiments with Stable Diffusion show up to 8 times acceleration with minor losses in the quality of identity preservation and less overfitting. We release the source code of DVAR so that anyone can speed up their adaptation pipeline.

Faster customization of text-to-image models with DVAR

Text-to-image synthesis models like Stable Diffusion, DALL-E 2 or Imagen have increased in popularity over the past few years. The research community keeps developing new approaches, with RAPHAEL, DeepFloyd and DALL-E 3 being some of the most recent ones. Still, assessing the language understanding capabilities of such models remains a challenging problem. In our latest work, "Hypernymy Understanding Evaluation of Text-to-image Models via WordNet Hierarchy", we present a method for analyzing the linguistic abilities and world knowledge of text-to-image models by measuring how well they understand subtype relations between concepts.

How much do text-to-Image models know? A hypernymy-based approach

Eight papers have been accepted for publication at the Conference on Neural Information Processing Systems (NeurIPS 2023).

Papers accepted to NeurIPS 2023

Nearest neighbor search is a long-standing problem arising in a large number of machine learning applications, such as recommender services, information retrieval, and others.

This competition focuses on developing indexing data structures and search algorithms for practical settings of the Approximate Nearest Neighbor (ANN) or Vector search problem. More details are on the competition’s website. We encourage everyone to join by August 30th, 2023!


Blog

Posts

Neural algorithmic reasoning without intermediate supervision

Faster customization of text-to-image models with DVAR

How much do text-to-Image models know? A hypernymy-based approach

Papers accepted to NeurIPS 2023

NeurIPS 2023 Competition Track: Big-ANN

Filter by: