Dmitriy Gizlyk / Profilo

Transformer architecture-based models demonstrate high efficiency, but their use is complicated by high resource costs both at the training stage and during operation. In this article, I propose to get acquainted with algorithms that allow to reduce memory usage of such models.

When working with time series, we always use the source data in their historical sequence. But is this the best option? There is an opinion that changing the sequence of the input data will improve the efficiency of the trained models. In this article I invite you to get acquainted with one of the methods for optimizing the input sequence.

In this article, we continue the implementation of the approaches of the ATFNet model, which adaptively combines the results of 2 blocks (frequency and time) within time series forecasting.

The authors of the FreDF method experimentally confirmed the advantage of combined forecasting in the frequency and time domains. However, the use of the weight hyperparameter is not optimal for non-stationary time series. In this article, we will get acquainted with the method of adaptive combination of forecasts in frequency and time domains.

We continue to explore the analysis and forecasting of time series in the frequency domain. In this article, we will get acquainted with a new method to forecast data in the frequency domain, which can be added to many of the algorithms we have studied previously.

By studying the FEDformer method, we opened the door to the frequency domain of time series representation. In this new article, we will continue the topic we started. We will consider a method with which we can not only conduct an analysis, but also predict subsequent states in a particular area.

All the models we have considered so far analyze the state of the environment as a time sequence. However, the time series can also be represented in the form of frequency features. In this article, I introduce you to an algorithm that uses frequency components of a time sequence to predict future states.

In an attempt to obtain the most accurate forecasts, researchers often complicate forecasting models. Which in turn leads to increased model training and maintenance costs. Is such an increase always justified? This article introduces an algorithm that uses the simplicity and speed of linear models and demonstrates results on par with the best models with a more complex architecture.

Forecasting plays an important role in time series analysis. In the new article, we will talk about the benefits of time series patching.

We continue to study timeseries forecasting algorithms. In this article, we will discuss another method: the U-shaped Transformer.

In this article, I would like to introduce you to a new complex timeseries forecasting method, which harmoniously combines the advantages of linear models and transformers.

We already know that pre-processing of the input data plays a major role in the stability of model training. To process "raw" input data online, we often use a batch normalization layer. But sometimes we need a reverse procedure. In this article, we discuss one of the possible approaches to solving this problem.

This article introduces the Conformer algorithm originally developed for the purpose of weather forecasting, which in terms of variability and capriciousness can be compared to financial markets. Conformer is a complex method. It combines the advantages of attention models and ordinary differential equations.

In this article, we will discuss another type of models that are aimed at studying the dynamics of the environmental state.

In previous works, we always assessed the current state of the environment. At the same time, the dynamics of changes in indicators always remained "behind the scenes". In this article I want to introduce you to an algorithm that allows you to evaluate the direct change in data between 2 successive environmental states.

In this article, I will get acquainted with the GTGAN algorithm, which was introduced in January 2024 to solve complex problems of generation architectural layouts with graph constraints.

In the previous article, we got acquainted with one of the methods for detecting objects in an image. However, processing a static image is somewhat different from working with dynamic time series, such as the dynamics of the prices we analyze. In this article, we will consider the method of detecting objects in video, which is somewhat closer to the problem we are solving.

In this article, I propose to look at the issue of building a trading strategy from a different angle. We will not predict future price movements, but will try to build a trading system based on the analysis of historical data.

In our models, we often use various attention algorithms. And, probably, most often we use Transformers. Their main disadvantage is the resource requirement. In this article, we will consider a new algorithm that can help reduce computing costs without losing quality.

This article continues the topic of predicting the upcoming price movement. I invite you to get acquainted with the Multi-future Transformer architecture. Its main idea is to decompose the multimodal distribution of the future into several unimodal distributions, which allows you to effectively simulate various models of interaction between agents on the scene.