Release notes
Version 6.1
Release Date: June 23, 2022
Platform updates
Frequency features were added to the DSP pipeline. Frequency features are based on fast Fourier transformation and may improve model quality.
Recursive feature elimination option was added to the Feature Selection. This option may reduce the number of input data columns (including deletion of features generated by the platform) and optimize the model footprint.
Version 6.0
Release Date: May 11, 2022
Platform updates
For solutions with the turned-off DSP options (only for 32-bit models), the downloadable archive with the C Library also includes models in ONNX, Keras, and TensorFlow formats which allows users to build these models into other pipelines.
DSP settings menu was updated and now users can configure DSP options in a more flexible way.
New Feature selection option was added to the platform. With this option enabled, one of the features with binary correlation above threshold (95%), will be automatically dropped.
Version 5.9
Release Date: April 18, 2022
Platform updates
Input data type selection option was added to the platform (INT8, INT16, UINT8, UINT16, or FLOAT32). This option allows to optimize DSP operations on a device and minimize model footprint.
Estimated model footprint (SRAM and Flash usage in Kb) was added to the platform.
Data dictionary was added to the C Library archive which helps users to map predicted class labels to original class names.
Remove lag features option was added to the platform.
Version 5.8
Release Date: April 12, 2022
Platform updates
Inference algorithm was updated; with the new release, the inference speed on devices may increase up to 4x times.
Windowing size specification is now available also in the number of rows.
Estimated model footprint (SRAM and Flash usage in Kb) was added to the platform.
Version 5.7
Release Date: April 1, 2022
Platform updates
New Digital Signal Processing (DSP) options were added to the platform to optimize preprocessing and improve model predictive power:
Frequency (in Hz) – the frequency of a sensor which is used for data collection.
Window size (in milliseconds) – the duration of one event (the static window approach is applied on the platform). It should be universal for all events in the training dataset (usually maximum duration), even if the duration of events is different.
Version 5.6
Release Date: March 21, 2022
Platform updates
New use cases were added to the platform. Now, users can analyze preloaded datasets and generated models. From this information, users can get inspiration on how to solve their own use cases.
Version 5.5
Release Date: February 8, 2022
Platform updates
New TinyML features were added to the platform. Now Neuton supports Digital Signal Processing (DSP). Users should enable this feature for automatic processing signals from gyroscopes, accelerometers, magnetometers, electromyography (EMG), etc.
Neuton will automatically transform raw data and extract additional features to create precise models for signal classification. For solutions with the enabled DSP option, you can download a ready-to-use archive with preprocessing code and model.
Other minor bug fixes and interface improvements were deployed.
Version 5.4
Release Date: January 20, 2022
Platform updates
The platform subscription process was modified and became more clear and easy.
Subscription plans were updated. Now a Zero Gravity plan has no restrictions on downloadable solutions or test dataset size.
New preloaded tinyML datasets and use case descriptions were added to the platform.
The text column detection algorithm was improved. Also, the logic of working with text columns and time series data was changed, now the time series option is disabled if a dataset contains text columns.
Other minor bug fixes and interface improvements were deployed.
Version 5.3
Release Date: October 14, 2021
Platform changes
Dataset validation checks were added for the uploading process. If the dataset does not meet the platform requirements, a user will see the message with recommendations on how to fix it.
With the new release, users can delete irrelevant features in the training dataset directly in the web interface. Deleted features will be automatically ignored in the test datasets with new data.
Minor bug fixes and interface improvements were deployed.
Version 5.2
Release Date: 25 September, 2021
Platform changes
We improved the subscription process, it is now easier and more clear.
Starting from this release, you can use the platform and its preloaded datasets without a subscription.
Some improvements in preprocessing and in the UI were deployed.
Version 5.0
Release Date: 4 May, 2021
Platform and UI changes
With the new release Neuton supports TinyML functionality. Using the simple and straightforward Neuton workflow you can select 8, 16 or 32-bit depth calculations and other parameters and generate your tiny model in just a few clicks. The trained model can be embedded in tiny devices like microcontrollers.
Version 4.0
Release Date: 19 April, 2021
Platform and UX changes
Model Quality Diagram indicator calculation was improved to reflect more realistic metric values for regression task type.
The EDA plotting functionality was upgraded to show the Time Dependencies plots for every detected date/time column in the training dataset.
The Correlation Matrix generation algorithm was upgraded in the EDA and now shows the top 10 features with the highest binary correlation with the target.
The Model Interpreter is now available in 1 click, for predictions which were made previously.
When Prediction is turned off, user can now enable it from the prediction results preview.
User notification was added for following cases:
when column names in the training dataset are not unique
when detected number of fields(columns) in some rows does not match the number of fields in the dataset header
when validation dataset structure does not match the training dataset structure
in the Model Interpreter when the feature values are out of bounds of the feature values in the training dataset
when prediction is enabled platform recommends to turn it off after getting the results to reduce the costs
Some small interface changes:
Processed training dataset preview is now available on the Prediction tab.
Animation was added to the Prediction button to highlight its status.
Link to dataset requirements was added on the “dataset uploading” pop-up windows.
Solution Id is now visible on the My Solution page.
System folders were hidden for “dataset selection” pop-up windows.
Close option was added to the “dataset selection” pop-up windows.
Version 3.0
Release Date: February 4, 2021
Fixed Issues and Notes
New data preprocessing methods and best practices for categorical variables are now available on the platform. The new methods allow to increase the model predictive power and training speed.
Preprocessing result (prerocessed training dataset) is now available in 1 click on the Training tab.
Calculation of Model-to-Data Relevance Indicator was accelerated for all task types.
Several bug fixes and interface improvements were also implemented.
Version 2.7
Release Date: December 10, 2020
Fixed Issues and Notes
An expansive list of preloaded datasets was added to the platform, grouped by business industry. These preloaded datasets are now also available from the "Business uses cases" tab on the left side of the screen.
Prior web prediction results are now available from the Prediction tab. To access prediction results hover the mouse over the "Start" web prediction button.
Model Interpreter is now available only for solutions with the time series option disabled.
Version 2.6
Release Date: October 27, 2020
Fixed Issues and Notes
Support of Cyrillic symbols was added to the platform.
The design of the left side of the menu was changed.
In the new release users can turn off the Model-to-Data Indicator and Confidence Interval calculations.
Test datasets were added for the preloaded datasets.
Version 2.5
Release Date: August 25, 2020
New Features
The Neuton Neural Framework was updated on the platform. The Neuton Neural Framework is the core of the platform, and the next generation of the framework allows users to create models up to 10 times faster and with enhanced predictive power.
A Model Quality Diagram was added to the platform. It is a graphical representation of model quality in relation to metric indicator values that are scaled in the range [0-1], where 1 is the ideal quality of the model, and 0 is the minimum quality of the model. It also allows users to understand metric balance. When the figure displayed is close to the shape of a regular polygon that conveys a perfect balance between the metric indicator values.
Version 2.4
Release Date: August 5, 2020
New Features
Model-to-Data Relevance Indicator was added to the platform, showing how trustworthy the predictions are. It is determined by evaluating the statistical difference between the data used for training and the data used for predictions.
Text notifications are now available on Neuton Platform. For your convenience, Neuton can inform you about training completion via text notifications, which can be especially useful for trainings with long durations. You have the opportunity to provide the preferred phone number for receiving updates in a pop-up window when training is initiated.
Bug Fixes
Class labels are now displayed as they looked in the training dataset for numeric labels.
Feature Importance Matrix is now available for solutions with enabled “Continue Training” option.
Version 2.3
Release Date: July 27, 2020
New Features
The Model Interpreter allows you to change the feature values and see how prediction result will change in real time. It also shows the feature influence for numeric features, which can help you to understand the influence of feature value on prediction results. And using the interactive settings for the threshold value of the target variable, you can see the values of features for which prediction results are above or under the selected threshold value.
In the new release you can use caret, tab or pipe as a separator for csv files. The full list of supported separators is now: comma, semicolon, pipe, caret and tab.
For preloaded datasets all training settings, except the target metric, are now preconfigured.
Automatic ticket notification was added for all workflow stages.. If you have in issue while working with Neuton, the platform will automatically create a ticket for the Neuton Support Team, and you will be notified via email. In the new release terms of use are also updated for support requests.
Prediction results are now displayed in the same window, instead of in a pop-up window.
Version 2.2.0
Release Date: June 15, 2020
New Features
Confidence interval column was added to prediction results for regression task type for all type of prediction (Web, REST API and Downloadable Solution).
Solution_ID was added to “Solution Details” settings on “My Solutions” page.
New descriptions for preloaded datasets were added to Dataset tab.
Metrics lists were changed for binary and multi class classification task types.
Smart task type detection algorithm was updated and now it detects task type for train dataset in more advanced way.
Version 2.1.0
Release Date: April 27, 2020
New Features
The New Feature Importance Matrix shows the ranking of the most important features based on both the original dataset and the dataset after preprocessing/feature engineering.
Exploratory Data Analysis shows dataset analytic information: data overview, feature categories relations, correlations, time dependencies, missing values, outliers and many other
Neuton Online
Release Date: December 1, 2019
Neuton Neural Network Framework release__version 3.4
Automated Neuton neural network model architecture creation
Automated hyper parameters configuration
Regression and classification (binary/multi-class) problem solving
Support for large datasets for training
Overfitting prevention
Automated training completion upon reaching optimal model and predictive power
Proprietary automated training and validation sampling
5-fold cross-validation for training
Neuton Auto ML
Intuitive and user friendly workflow for Machine Learning model creation
Automated and manual problem type definition (regression or classification) based on dataset analysis
Automated dataset preparation (preprocessing)
Time series support
Text fields support
Automated feature engineering
Automated training
Automatic training completion upon one of the following conditions:
Creation of the best possible model without overfitting
Time limit (user input) reached
Number of coefficients limit (user input) reached
Manual stop and resume training capability
Support for CSV datasets
Ability to preview uploaded datasets
Ability to download datasets previously uploaded to the platform
Preloaded experimental datasets
Display of other relevant metrics in addition to the target metric
Model feature importance matrix visualization
Downloadable models in binary format with Python script-calculator for local use without Neuton
Embedded models capability
Predictions via Web interface
Predictions via REST API
REST API examples in various programming languages s (C++, C#, Python, Java, Scala)
No limitations on input data size for predictions*
*Excluding the Gravity Plan
Cloud Based solution
SaaS solution
GPU training support
Automated and seamless provisioning of infrastructure necessary to perform training and predictions
Stay updated, join the community