ai
AI & Deep Learning
By time series in artificial intelligence we mean digitised sequences. These sequences are discretely sampled in time (sampling uniformity is not constrained), the value that can be recorded by the time series at each sampling point is a vector of certain variables, which can be either categorical or numerical.
AI and Deep Learning Solutions
Problems of AI-based time series analysis
The problems encountered when processing time series can be grouped into several categories. One may be the problem of sequence-to-sequence (seq2seq) training, where time series that are considered to be related are predicted from each other, with both input and output being sequential data. Nowadays, the time series forecasting is also an important problem, which broadly speaking tries to predict unknown points in a time series, which is usually a future point, but it is also possible to estimate a value for an intermediate missing time (e.g. a missing value due to an error or hidden information due to an external factor). The third main type of task is the labeling and segmentation of time series, which assigns descriptive values to either the whole time series or just certain points in it. Such a task can be the prediction of malfunctions or the detection of certain patterns within the sequence. Finally, there is the sub-problem of time aggregation (a type of data aggregation), which focuses on the generation of a single feature or characteristic based on either the whole or just parts of a time series. This feature compiles the information hidden in the time series data into a more easily consumable, more comprehensive form, so therefore the user can interpret and understand the information provided by the time series more easily. Such a feature or characteristic may be a quantity directly related to the time series (e.g. the signal-to-noise ratio), or maybe only indirectly related to it (e.g. the effect of a piece of text in a newspaper on the value of a stock).
Analysis of current production processes, integration of existing functions
Development and implementation of new solutions and services
Data analysis
Our AI and Deep Learning based BIG Data data analysis and data mining solutions enable efficient storage, processing, analysis and visualisation of large and complex data sets. Our systems are capable of autonomously identifying relationships and patterns within a data set as well as predicting future values for time-series data.
Our latest Business Intelligence (BI) systems also leverage our AI-based BIG Data solutions.
Natural Language Processing (NLP)
Our artificial intelligence-based language processing systems are capable of automatic recognition, understanding and analysis of spoken and/or written texts with various content.
We use our NLP solutions primarily in our AI assistants and in our text-based BIG Data data analysis systems.
Machine Vision (MV) Systems
Our Machine Vision (MV) solutions are assembled with various purpose-built industrial cameras and sensors. Thanks to machine vision based on artificial intelligence, our systems are able to autonomously identify objects placed in the field of view of the cameras and recognise certain characteristics of the objects (e.g. their orientation).
Our machine vision solutions can be used to efficiently automate various mass production processes, quality control (QC) of finished products and inventory control.
AI Assistants
Our AI Virtual Assistant solutions are able to independently interact with the speaker, interpret the instructions of the user, and then carry out the related actions.
With our solutions, various customer service processes can be automated and simplified with great efficiency.
Convolutional Neural Networks (CNN)
Convolutional solutions
(TCN, ResNet)
Convolutional neural networks (CNN) can be an effective solution for AI-based time series data analysis. In this case, the network operates on an N-channel time series of length T using the 1D discrete convolution operator instead of the usual full switching.
Convolutional networks are efficient at processing temporal patterns at a certain scale, but are incapable of detecting larger scale patterns on their own. A solution to this is the introduction of pooling layers, which perform a static (untrainable) pooling operation using a window that runs through the time series, averaging or maximizing the values in the window to an element of a new “shrunken” sequence. This allows the next convolution layer to work with patterns based on significantly more time instants even with a small kernel.
The average convolutional architecture contains alternating convolutional and pooling layers, which can thus produce a perfectly new time series under appropriate settings, or even extract a single value from them.
Time series data analysis
Image recognition, Machine Vision (MV) systems
Recurrent Neural Networks (RNN)
Recurrent solutions
(LSTM, GRU)
For processing certain time series, it may be necessary to implement dynamic behaviour and memory locally, and to establish correlation between temporally distant points, for which recurrent neural networks (RNNs) provide an excellent solution.
The use of recurrent networks is also supported by a kind of biological inspiration, as they can be used to simulate the functioning of the human nervous system more effectively, since it can also be understood as a dynamic system.
Recurrent networks are based on feedback and cyclic execution, which treats the output of a layer as input in the next step, in addition to external signals (from other layers or from outside the network) at the input of the layer. The weights of such neurons can be split into two parts, the simple input part of these can be found in all classic MLP (multilayer perceptron) networks, but these are complemented by parameters that weight the previous output of the neuron.
Most frequently used RNN network algorithms:
LSTM, GRU
AI-based dynamic predictions
Speech recognition, text analysis
Attention-based Neural Networks
Attention-based solutions (Transformers, MHA, BERT, GPT)
The attention mechanism simulates human cognitive processes. This mechanism is based on the analysis of context, since just as in human behaviour the interpretation of information depends on the situation what we pay attention to, what we emphasise, etc. In the same way, when processing a time series, it can be critical what other contextual information we take into account when processing a particular detail.
In neural networks based on an attention mechanism, we have keys, values, and a query. If we generate keys and values from the time series, we can obtain the relevant context for each time point by generating a query from the sample at that time point. In this case the weighting of the values depends on the time series itself. We do not highlight values based on a predefined fixed weight matrix, but on the query-key similarity. This operating principle gives the attention mechanism its processing power.
Text interpretation, text creation
AI assistants
Your AI and Deep Learning solution
Plan your AI and Deep Learning project with us!
Our experts are happy to help you identify the industrial and economic areas where our artificial intelligence algorithms and machine learning-based systems can effectively help your business.
Our company is in constant contact with the most successful universities and research centres in Hungary, so you can benefit from the latest state-of-the-art AI and Deep Learning solutions in your business.
We believe that every hour spent on design pays off many times over in the implementation and roll-out of our systems. Our ergonomically designed user interfaces provide our customers with a new user experience and ease of use.