Linked Knowledge Nuggets: arrow_forward "AI in Automotive Systems: Aligning with ISO/PAS 8800"
personAuthor: Sebastian Keller
How can AI be safely integrated into the vehicles of tomorrow?
ISO/PAS 8800:2024 lays the foundation for managing artificial intelligence in safety-related automotive systems. Join our webinar and learn what the new specification means for managers, project leaders, quality specialists, and engineering teams.
We’ll demystify the key concepts behind AI safety, explain how ISO/PAS 8800 relates to ISO 26262, and show how organizations can prepare for the next generation of system validation and assurance.
You will gain an overview of how to transfer AI development activities into structured frameworks such as Automotive SPICE, define roles such as AI safety manager or data governance lead, and avoid common pitfalls such as uncontrolled uncontrolled data drift.
Reserve your spot today and lead your company's AI safety transformation with confidence.
school
Webinar recording and slides
arrow_forward "What is a Tensor (in Machine Learning)?"
personAuthor: Process Fellows
Tensors are the basic building blocks of machine learning. They are multidimensional arrays that generalize scalars, vectors and matrices. They are used in deep learning, especially for data such as images, texts or time series.
Examples:
Image processing (3D tensor):
Colored (RGB) picture is stored as: Width x Height x RGB Colors
3D tensor with (256,256,3)
Training data of 32 pictures, each picture of 28 pixels x 28 pixels and one grayscale channel:
4D tensor (32,28,28,1)
# PROCESS PURPOSE
The purpose is to optimize the ML model to meet the defined ML requirements.
# PROCESS OUTCOMES
O1
An ML training and validation approach is specified.
O2
The data set for ML training and ML validation is created.
O3
The ML model, including the (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.) values, is optimized to meet the defined ML requirements.
O4
Consistency and bidirectional traceability are established between the ML training and validation data set and the ML data requirements.
O5
The results of the optimization are summarized, and the trained ML model is agreed and communicated to all affected parties.
# BASE PRACTICES
BP1
Specify the ML training and validation approach. (
O1 )
Specify an approach which supports the training and validation of the ML model to meet the defined ML requirements. The ML training and validation approach includes
entry and exit criteria of the training and validation,
approaches for the (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.) tuning / optimization,
approach for the data set creation and modification, and
the training and validation environment
Note 1: The ML training and validation approach may include random dropout and other robustification methods. Note 2: ML validation is one step of optimizing the model. If the model performance does not meet expectations, (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.) values or training and validation data may be changed. The term “validation” has a different meaning than VAL.1.
Linked Knowledge Nuggets: arrow_forward "Dropout in Machine Learning"
personAuthor: Process Fellows
“Dropout” is a regularization technique to reduce “overfitting”. NNs are often very complex and can adapt too much to the training data (overfitting).
Consequence: Good performance on training data, poor generalization to new data.
Dropout means, during training, some neurons are randomly deactivated (set to zero) to make the model more robust.
The model learns redundant structures and is less dependent on individual neurons.
Dropout is not used during prediction (only during training).
Advantages: Reduces overfitting, increases the robustness, promotes better generalization
Disadvantages: Can slow down training, does not always work well for small NNs
BP2
Create the ML training and validation data set. (
O2 )
Select data from the ML data collection provided by SUP.11 and assign them to the data set for training and validation of the ML model according to the specified ML training and validation approach. Note 4: The ML training and validation data set may include corner cases, unexpected cases, and normal cases depending on the ML requirements. Note 5: A separated data set for training and validation might not be required in some cases (e.g., k- fold cross validation, no optimization of (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.)).
Linked Knowledge Nuggets: arrow_forward "What is k-fold cross validation?"
personAuthor: Process Fellows
K-fold cross-validation is a common technique in machine learning used to evaluate model performance. It involves dividing the dataset into k equally sized folds. In each of the k iterations, one fold is used as the validation set while the remaining k–1 folds are used for training. This process is repeated k times, with each fold used exactly once as the validation set. The performance results from each iteration are then averaged to produce a final estimate. The value of k is a predefined parameter, typically set to 10, meaning the dataset is split into 10 parts.
arrow_forward "What's the difference between "corner cases" and "unexpected cases"?"
personAuthor: Process Fellows
Corner cases are extreme or rare scenarios that lie within the expected value range.
They occur rarely, but are still part of the domain.
Examples:
An image recognition model for cats recognizes a cat with a very unusual coat color (e.g. completely pink).
An autonomous vehicle encounters an unusual traffic route, but this is provided for in the road traffic regulations (e.g. an extremely tight hairpin bend).
A very high or very low value in a time series data set (e.g. a share price rises by 100% in a single day).
Corner cases often require fine-tuning or special training data to ensure that the model works reliably even in these rare cases.
Unexpected cases are scenarios that the model did not anticipate because they are outside the training domain or are not covered by the data model.
Examples:
An image recognition model for cats is supposed to recognize a cat, but the image contains a new animal species that the model has never seen (e.g. a Sphynx cat if the model only knows furry cats).
An autonomous vehicle encounters a completely new traffic regulation or an unexpected road situation that it has never seen before in training data (e.g. a sudden earthquake and collapsing bridges).
A language model is confronted with a new language or dialect that it does not know.
Unexpected cases often cannot be solved simply by using more training data.
This is where approaches such as out-of-distribution (OOD) detection, anomaly detection or unsupervised learning help to react to completely new cases.
BP3
Create and optimize the ML model. (
O3 )
Create the ML model according to the ML architecture and train it, using the identified ML training and validation data set according to the ML training and validation approach to meet the defined ML requirements, and training and validation exit criteria.
Linked Knowledge Nuggets: arrow_forward "How can I optimize the ML model architecture after the training is done without negative impact of the model outcome quality? "
personAuthor: Process Fellows
The ML model is typically considered as a black box!
This question sounds like it would be impossible? Let's see what kind of techniques could be considered and used:
Pruning (weight thinning)
Idea: Removal of unimportant neurons or connections in the network (some weights have very little influence on the final result).
Various techniques (e.g. magnitude-based pruning) can be used to eliminate weights or entire neurons.
Test: Retraining or evaluation is used to check whether the prediction performance is maintained.
Quantization (reduction of the precision of the parameters)
Idea: Reducing the precision of the model parameters (e.g. instead of 32-bit float values for weights, reduce to 16-bit or even 8-bit).
Modern hardware (e.g. TensorFlow Lite or special AI accelerators) efficiently supports such quantized models.
Test: The accuracy of the quantized model is measured in comparison to the original model.
Knowledge distillation (knowledge transfer from large to small)
Idea: A large model (“teacher”) trains a smaller model (“student”).
A smaller model is trained to imitate the output of the larger model.
As a result, the smaller model can often make similar predictions, even if it has fewer parameters.
Test: Evaluation of the predictions of the “Student model” compared to the “Teacher model”.
Tip: If TensorFlow or PyTorch is used, there are libraries such as
TensorFlow Model Optimization or Torch Pruning
that automate many of these techniques.
arrow_forward "Learning Algorithms for Neural Networks"
personAuthor: Process Fellows
There are various learning algorithms that differ in their method for updating weights. However, they are mostly based on gradient descent and its variants.
Gradient descent explained simply: Imagine a mountain hike:
You are standing on a hill (high error value)
You want to go down into the valley (minimum error function).
The gradient (slope) shows you in which direction you are going downhill.
You take small steps in the direction of the steepest descent.
Examples of algorithms:
Backpropagation, Stochastic Gradient Descent (SGD), Mini-Batch Gradient Descent, Adaptive Moment Estimation (Adam), etc.
BP4
Ensure consistency and establish bidirectional traceability. (
O4 )
Ensure consistency and establish bidirectional traceability between the ML training and validation data set and the ML data requirements. Note 6: Bidirectional traceability supports consistency and facilitates impact analyses of change requests. Traceability alone, e.g., the existence of links, does not necessarily mean that the information is consistent.
Linked Knowledge Nuggets: arrow_forward "Consistency vs. Traceability – What’s the Difference?"
personAuthor: Process Fellows
Consistency ensures that related content doesn’t contradict itself – e.g., requirements align with architecture and test. Traceability, in contrast, is about links: can you follow a requirement through to implementation and verification? Both are needed – consistency builds trust, traceability enables control. Typically, traceability strongly supports consistency review.
arrow_forward "The role of traceability in risk control"
personAuthor: Process Fellows
Traceability isn’t just about completeness — it’s about managing impact. When a requirement changes, trace links tell you what’s affected. That’s your early-warning system.
arrow_forward "The true benefit of traceability
"
personAuthor: Process Fellows
Sometimes the creation of traceability is seen as an additional expense, the benefits are not recognized.
Traceability should be set up at the same time as the derived elements are created. Both work products are open in front of us and the creation of the trace often only takes a few moments.
In the aftermath, the effort increases noticeably and the risk of gaps is high.
If the traceability is complete and consistent, the discovery of dependencies is unbeatably fast and reliable compared to searching for dependencies at a later stage, when there may also be time pressure.
It also enables proof of complete coverage of the derived elements and allows the complete consistency check.
BP5
Summarize and communicate the agreed trained ML model. (
O5 )
Summarize the results of the optimization and inform all affected parties about the agreed trained ML model.
# OUTPUT INFORMATION ITEMS
13-52
Communication evidence (
O5 )
Evidence of interpersonal communication.Identifies:
Scope of information
Need for feedback, for example an expected confirmation within one week
Meta data, for example time when communication was done or how information was distributed.
Includes:
Personal information
Work-flows, for example within tools
Examples and References:
E-mails and other forms of memos
Verbal statements
Meeting minutes, for example in standups
Electronic media, for example webcasts, blog posts intranet forum
Chat protocols
Wiki pages
Photo protocol
Used by these processes:
ACQ.4 Supplier Monitoring
HWE.1 Hardware Requirements Analysis
HWE.2 Hardware Design
HWE.3 Verification against Hardware Design
HWE.4 Verification against Hardware Requirements
MAN.3 Project Management
MLE.1 Machine Learning Requirements Analysis
MLE.2 Machine Learning Architecture
MLE.3 Machine Learning Training
MLE.4 Machine Learning Model Testing
PIM.3 Process Improvement
REU.2 Reuse of Products
SUP.1 Quality Assurance
SUP.11 Machine Learning Data Management
SWE.1 Software Requirements Analysis
SWE.2 Software Architectural Design
SWE.3 Software Detailed Design and Unit Construction
SWE.4 Software Unit Verification
SWE.5 Software Component Verification and Integration Verification
SWE.6 Software Verification
SYS.1 Requirements Elicitation
SYS.2 System Requirements Analysis
SYS.3 System Architectural Design
SYS.4 System Integration and Integration Verification
SYS.5 System Verification
VAL.1 Validation
Used by these process attributes:
PA2.1 Process performance management process attribute
13-51
Consistency evidence (
O4 )
Evidence of information to be semantically coherent alongrelevant artifacts, ensuring completeness, purpose maturity of processes, (Task = A definition, but not the execution, of a coherent set of atomic actions.) and products throughout their lifecycle.Identifies:
Traceability information, for example hyperlinks, repository location or editorial references.
Naming conventions
Relevant artifacts
Revision and revision history information
Change documentation and analysis information
Includes:
Meta-information, for example database identifiers notes in Git commits comments
Examples and References:
Evidence of Definition of Done (DoD) adherence.
Used by these processes:
HWE.1 Hardware Requirements Analysis
HWE.2 Hardware Design
HWE.3 Verification against Hardware Design
HWE.4 Verification against Hardware Requirements
MAN.3 Project Management
MLE.1 Machine Learning Requirements Analysis
MLE.2 Machine Learning Architecture
MLE.3 Machine Learning Training
MLE.4 Machine Learning Model Testing
SUP.8 Configuration Management
SUP.10 Change Request Management
SWE.1 Software Requirements Analysis
SWE.2 Software Architectural Design
SWE.3 Software Detailed Design and Unit Construction
SWE.4 Software Unit Verification
SWE.5 Software Component Verification and Integration Verification
SWE.6 Software Verification
SYS.2 System Requirements Analysis
SYS.3 System Architectural Design
SYS.4 System Integration and Integration Verification
SYS.5 System Verification
VAL.1 Validation
01-54
Hyperparameter (
O3 )
Settings or configurations that are used to control the learning process of a machine learning model.Identifies:
Description
Initial value
Value of the results of the ML training
Includes:
Number of epochs
Scaling of network (number of layers or neurons per layer)
Examples and references:
Learning rate of training, Loss function
Used by these processes:
MLE.2 Machine Learning Architecture
MLE.3 Machine Learning Training
03-51
ML data set (
O2 )
Selection of data for machine learning training and validation, or test of a machine learning model.Identifies:
Patterns
Relationships
Features
Includes:
Annotations and labels
ML Training and Validation Data Set
ML Test Data Set
Used by these processes:
MLE.3 Machine Learning Training
MLE.4 Machine Learning Model Testing
08-65
ML training and validation approach (
O1 )
Approach describing method for evaluating a trained ML model using data to confirm its accuracy and for the intended (Task = A definition, but not the execution, of a coherent set of atomic actions.).Identifies:
Entry and exit criteria
Approaches for (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.) tuning / optimization
Approach for data set creation and modification
Includes:
Training environment setup and configuration
Resources
Provision of input data
Storage of output data
Methods to control robustness, for example random dropout
Used by these processes:
MLE.3 Machine Learning Training
01-53
Trained ML model (
O3 )
Resulting output machine learning (ML) model of the ML training processIdentifies:
Set of weights optimized during the training
Includes:
Software representing the ML architecture
Set of (Hyperparameter = In machine learning, a hyperparameter is a parameter whose value is used to control the training of the ML model. Its value must be set between training iterations. Examples: learning rate, loss function, model depth, regularization constants.)
MLE.3
Machine Learning Training
