Frequently Asked Questions

If you need technical assistance with Saimple, here are the ways to get in touch with our support team:

By email: As a registered customer with us, you can find our support email address in the welcome message sent along with your account creation.

By chat: If an agent is available, a chat widget will be displayed at bottom right corner of our website or Saimple web interface.

By submitting a form: Head to our contact page, from the drop down, make sure to select “Support” as recipient to your message.

Saimple reduces the carbon footprint of AI validation

Saimple is an AI validation solution that can significantly reduce the carbon footprint of these projects. In a single evaluation, Saimple can perform the equivalent of hundreds or even thousands of traditional and empirical evaluations. This helps to reduce the resources and computing capabilities needed, which translates into a reduction in energy consumption and greenhouse gas emissions.

In addition, the Saimple SaaS solution is hosted by a committed service provider that offsets its CO2 emissions by 200%. This means that this provider invests in projects that absorb more CO2 than it emits.

In conclusion, Saimple is a low-energy impact AI validation solution that can help you reduce the carbon footprint of your AI projects.

The training materials available for Saimple are:

• Technical documentation on doc.saimple.com
• Use Cases
• Our videos on our YouTube channel
• Custom training on demand (on-site or remote), contact the sales support

Saimple is the first tool based on formal methods that can scale and analyze industrial AI. Academic tools also rely on comparable technologies, but they do not allow for the processing of large-scale AI or in the form of a service integrated into a development chain and ultimately a certified chain. It is possible to evaluate the product by creating a freemium account.

Relevance is the measurement of the impact of each dimension of the input of an AI on its output for the entire set of all inputs considered. When the measurement of these impacts is highlighted in human-understandable reports, they provide elements of explanation about the functioning of the system.

Robustness is the ability of an AI system to perform accurately in the presence of noise or other disturbances. Dominance is a specific type of robustness that focuses on the stability of the output of a neural network around its input.

In other words, robustness is the ability of an AI system to be accurate even when the input data is not perfect. Dominance is the ability of a neural network to produce the same output even when the input data is slightly modified.

A formal method is a mathematical approach that can be used to build a proof of the correctness of a system. In the case of Numalis, formal methods are used to prove that an AI system will behave correctly for a given set of input data.

Formal methods are different from statistical methods, which can only test the behavior of a system at a limited number of points. Statistical methods can never prove that a system will behave correctly in all cases.

The formal methods developed by Numalis are based on 25 years of academic research and use the principles of abstract interpretation of programs.

There are several formats for developing and training a neural network, for example libraries such as TensorFlow, Keras, PyTorch, and Scikit-learn each offer their own format. In order to reconcile all of these approaches, Saimple uses the ONNX standard. Most machine learning frameworks allow you to export networks in this format. If the network provided is not translated to ONNX format, Saimple will attempt to convert it automatically by relying on a conversion library available in the open source community.

Saimple works by using networks described in the ONNX format. If the custom layer can be decomposed into standard ONNX operations, the model can be supported. It is important to make sure that the tool used for translation supports the custom layer and that it is decomposed into standard ONNX operators that are all supported by Saimple.

Saimple supports image files in the following formats: .png, .jpg, .jpeg, .tiff (experimental). The tiff format works as data, but the format is not supported by the Saimple web app for relevance visualization.

It is also possible to upload tabular data in csv format (space, comma, or semicolon separated) or in numpy format (npy).

Frontend: The status information for an evaluation is available in the “Details” tab of the evaluation. The “Details Execution” and “Evaluation Error” fields show the information from the analysis engine.

Backend: It is possible to query the information for an evaluation via the API using the route “/api/v2/evaluations/{evalId}/trace”.

The information is returned in a JSON document. The “err” attribute returns the logs in raw text format. The “ParsedStandardError” attribute provides a parsed version that is easier to use, but may be less complete.

Frontend: This feature is not yet available via the web interface.

API: It is possible to retrieve the uploaded network via the GET route /api/v2/models/<modelId>/file.

Local noise can significantly reduce the execution time and resource usage. If this does not work, you can contact the support team.

No, deleting the data from an evaluation requires deleting the evaluation, which will also delete the results. However, it is possible to export the results beforehand.

Saimple is actively maintained, and several versions are released each year.

The values of Delta max must correspond to the conditions of use of the system. They must represent variations linked to the intensity of the perturbations that the system may be subjected to. It is therefore context-dependent.

In addition to the application domain in the context of images, it is important to consider cases of delta that lead to effective changes in pixels (i.e., greater than 1/255 in the case of a classic RGB image).

The improvement of this result can be obtained by better training of the model, or even by improving the model.

Saimple allows you to validate the behavior of a system on an (exponential) set of data points. It is possible to manually test the behavior of inferences on isolated data points from this set (but this is very tedious). The behavior of the inferences will be contained within the envelope of behaviors calculated by Saimple.

Yes, relevance can often highlight the elements on which the network places importance, when it should not. Also, Saimple can highlight the effective separation between classes and therefore identify which ones need to be separated better as a priority in the training dataset.

Normalization is an operation that brings an input into a valid domain for the network being analyzed. Neural networks often operate on values between 0 and 1, while images are typically encoded on values between 0 and 255. A simple normalization technique is to divide all pixel values in an image by 255 before passing it to the analysis.

The architecture and training of the network determine whether or not the inputs should be normalized.

Saimple’s scalability is contingent upon the computational resources available to perform evaluations.

In Saimple’s SaaS version, these resources are tailored to accommodate any need.

For the on-premises version, our teams work with your engineers to assess the infrastructure you will require.

Saimple supports the ONNX standard format, which allows for compatibility with a wide range of model formats. If the model has been created in TensorFlow / Keras, it can be directly uploaded to Saimple. The conversion to the ONNX format will be performed automatically using the official library provided by TensorFlow / Keras (available here: https://github.com/onnx/tensorflow-onnx).

However, since most PyTorch models only save the model weights, it is necessary for the user to perform the conversion to ONNX beforehand. To convert a PyTorch model to ONNX, the procedure described here can be followed: https://pytorch.org/tutorials/beginner/onnx/export_simple_model_to_onnx_tutorial.html

The tooling provided by Numalis is proprietary. However, it can be accessed through a public API. Additionally, many resources on our website (use cases, educational materials) are freely available.

Currently, it is not possible to transfer an evaluation from one user account to another. However, you can export all the results from the graphical interface or via the API and share them afterwards.

If you’ve forgotten or lost your password, you can request a reset from the login interface by clicking on “Forgot Password?”

Follow the instructions received via email to set your new login password.

For security reasons, changing your password involves resetting it, just as you would if you lost or forgot your password.

Once logged out, you can click on “Forgot Password?” on the login page.

Follow the instructions received via email to set your new login password.