© Fraunhofer IESE 1 Dr. Dominik Rost Dr. Matthias Naab 14.10.2020 Software Architecture Gathering 2020 Machine-Learning-basierte Systeme: Eine Einführung für Architekten

© Fraunhofer IESE 2 Recommendation Systems

© Fraunhofer IESE 3 Traffic Prediction

© Fraunhofer IESE 4 Email Categorization & Spam Detection

© Fraunhofer IESE 5 Image tagging and filtering

© Fraunhofer IESE 6 Automatic ID Verification

© Fraunhofer IESE 8 Huge number of examples ML as building block in applications

© Fraunhofer IESE 9 Information Sources Company Websites, Success Stories, etc. Architecture ? Tech Tutorials, Math. Foundations etc.

© Fraunhofer IESE 10 Goals of this Talk We elaborate the topic for software architects ◼ Create ◼ “Big Picture” for architecture of ML-based systems ◼ Architecture language for ML-based systems ◼ Foundation for ◼ Structured thinking about and designing ML-based systems ◼ Talking to ML experts and data scientists ◼ Judging existing concepts and technologies and filling the own toolbox

© Fraunhofer IESE 11 Approach Top-Down Bottom-Up System Decompostion according to Architecture Views Functions, Data, Deployment, … Explanation & Classification of Major Concepts ML Concepts, Process Steps, …

© Fraunhofer IESE 12 Example: Autonomous Driving Information partially based on „Tesla Autonomy Day“, https://www.youtube.com/watch?v=-b041NXGPZ8

© Fraunhofer IESE 17 ENGINEERING TRADITIONAL VS. ML-BASED SYSTEMS

© Fraunhofer IESE 18 Traditional ML-Based Engineering Traditional Systems vs. ML-based Systems Computer Input Data Program Output Computer Input Data Expected Output Model Mix of Dimensions :(

© Fraunhofer IESE 19 Engineering Traditional Systems vs. ML-based Systems Software System Traditional Software Engineering (Methods & Tools) Requirements Input Data Software System Output Data Traditional DevTime RunTime SE for ML-Based Systems (Methods & Tools) Machine Learning Data Science (Methods & Tools) Software System based on ML ML Component Requirements Output Data Software System based on ML ML Component Input Data ML-Based ML-Training ML-Inference Input Data Expected Output

© Fraunhofer IESE 23 FUNCTIONAL & DATA DECOMPOSITION

© Fraunhofer IESE 24 Data Flow through Software System based on ML Software System based on ML ML Component Input Data Output Data Data Pre- Processing Data Post- Processing

© Fraunhofer IESE 25 Multiple ML-Components in a System Software System based on ML ML C1 Input Data Output Data ML C3 ML C4 Architecture Decision: How many ML-Components and which ones? ML C2

© Fraunhofer IESE 26 Example Autonomous Driving Alternative Functions Software System based on ML “Driving” Sensors, Cameras, … Driving Actuators Data Pre- Processing Data Post- Processing Alternative A Software System based on ML Driving Area Detection Sensors, Cameras, … Driving Actuators “Steering” Obstacle Detection Roadsign Detection Alternative B Software System based on ML Road Marking Detection Sensors, Cameras, … Driving Actuators … Alternative C

© Fraunhofer IESE 27 Logical Structure of a ML Component (Generalized, Neural Network) ML Component Model Config Parameters (e.g. NN Weights) Learning Method / Algorithm (e.g. NN Topology) Hyperparameters (Training Config Params) Config Data Basic Inference Logic Learning / Training Logic Code / Logic ML Model (fixed in inference) Training Data In-Model State during Inference Data Input Data Output Data Config Data The ML Component can be treated as a black box, architecturally The ML Component is the unit of training State: E.g. in Recurrent Neural Networks with feedback relationships

© Fraunhofer IESE 29 ML Component Example Topology of a Convolutional Neural Network (CNN) Topology (Layers, Nodes, Relationships) Decisions about the topology of the Neural Network are mainly done by data scientists. Architects need a basic understanding to judge external implications. https://www.easy-tensorflow.com/tf-tutorials/convolutional-neural-nets-cnns

© Fraunhofer IESE 30 MODEL TRAINING

© Fraunhofer IESE 33 Overall Lifecycles / Workflows and Data Involved ML-Training (DevTime) Data Collection Data Preparation Model Selection & Training Model Evaluation Model Persistence ML-Inference (RunTime) Data Ingestion Data Preparation Inference Model Deployment Large amounts of data Computing-intensive training Exploratory approach Concrete input data Inference is comparably cheap “Just computation”

© Fraunhofer IESE 34 Feedback Data and Optimization ML-Inference (RunTime) Data Ingestion Data Preparation Inference Model Deployment New Training Data from Live Operation Deploy optimized model ML-Training (DevTime) Data Collection Data Preparation Model Selection & Training Model Evaluation Model Persistence

© Fraunhofer IESE 35 Example Autonomous Driving Tesla: Data Collection from Current Fleet – Driving Real World, not Autonomously Yet Deploy optimized driving functions model New Training Data from Live Operation Camera images Driving situations Data labeled from driver behavior / steering Data labeled from explicit user feedback Data labeled from additional sensors (e.g. radar) Central Data Collection and Learning Model Selection & Training & Evaluation Data Preparation Model Persistence Instruct cars, which data to collect Partially human pre-processed data ▪ Architects need overall system perspective ▪ Strong integration between runtime system (cars) and development time (learning and improvement) ▪ Continuous improvement and deployment ▪ Learning from the pre-phase of autonomous driving and continuously during operation

© Fraunhofer IESE 38 Feedback Data and Optimization (Online Learning) ML-Inference (RunTime) Data Ingestion Data Preparation Inference ML-Training / Retraining (RunTime) Model Training / Optimization Model Persistence Model Deployment New Training Data from Live Operation Learning can happen at defined points in time (rather not after every inference) (DevTime) el on & ng Model Evaluation Model Persistence Model Deployment The data science work is still done at DevTime Model is selected and training is done At Runtime, only optimization of the model

© Fraunhofer IESE 39 DATA

© Fraunhofer IESE 40 Data Aspects ◼ Large amounts of data for training needed ◼ Amount depends on application area, available data and on ML models / algorithms ◼ Very different types and formats of data ◼ Text ◼ Images ◼ Video ◼ Audio ◼ … ◼ → require very different treatment ◼ → result in very different computational load

© Fraunhofer IESE 41 Example Autonomous Driving Data Aspects in Autonomous Driving ◼ Data needs ◼ Large data ◼ Varied data ◼ Real data ◼ Collect data from the fleet ◼ Create simulation data ◼ Cover edge and unusual cases Image: https://www.youtube.com/watch?v=-b041NXGPZ8

© Fraunhofer IESE 42 DEPLOYMENT

© Fraunhofer IESE 43 Design Alternatives Deployment Options Model Training / Optimization Model Persistence ML-Inference (RunTime) Data Ingestion Data Preparation Inference Model Deployment (DevTime) el on & ng Model Evaluation Model Persistence Model Deployment Training HW Powerful Server ML Component Client Server ML Component Client ML Component Server Design Alternatives Client Server ML Component Client ML Component Server ML-Training (RunTime)

© Fraunhofer IESE 44 Example Autonomous Driving Multiple Instances of Systems (Cars) ◼ Learning strategies ◼ Online-Learning in each car? ◼ Batch-Learning in a central system, only? ◼ Can cars communicate? ◼ Compare: ◼ Learning of typing recognition on mobile phone Training HW Powerful Server ML Component ML-Training (DevTime) ML-Inference (RunTime) New Training Data from Live Operation Deploy optimized model Software System based on ML ML Component Software System based on ML ML Component Software System based on ML ML Component Software System based on ML ML Component Software System based on ML ML Component Software System based on ML ML Component

© Fraunhofer IESE 45 TECHNOLOGIES

© Fraunhofer IESE 46 Available Technologies as Services / Libraries for ML Different Level of Reuse ML Component In-Model State during Inference Model Config Parameters (e.g. NN Weights) Learning Method / Algorithm (e.g. NN Topology) Hyperparameters (Training Config Params) Basic Inference Logic Learning / Training Logic Data Config Data Code / Logic ML Model (fixed in inference) Fully trained model, immutable (as API or library) [e.g. Service for image tagging] Fully trained model, retrainable (as API or library) [e.g. Service for image tagging] Predefined topology (as API or library) [e.g. predefined CNNs] Basic ML model (as library) [e.g. general NN logic] Degree of freedom Knowledge needed Effort needed

© Fraunhofer IESE 47 Microsoft AI & ML Technologies https://www.credera.com/wp-content/uploads/2018/04/The-Microsoft-AI-platform.png Fully trained model, immutable (as API or library) [e.g. Service for image tagging] Fully trained model, retrainable (as API or library) [e.g. Service for image tagging] Predefined topology (as API or library) [e.g. predefined CNNs] Basic ML model (as library) [e.g. general NN logic] Predefined topology (as API or library) [e.g. predefined CNNs]

© Fraunhofer IESE 48 AWS AI & ML Technologies https://www.quantiphi.com/machine_learning/

© Fraunhofer IESE 51 RELATIONSHIP TO QUALITY ATTRIBUTES

© Fraunhofer IESE 52 Quality Attributes in ML-based Systems (1/2) ◼ ML as a technology does inherently aim more at realizing functionality than at realizing quality attributes (in contrast to e.g. communication middleware, blockchain, …) ◼ However, ML can be used to support achieving some quality attributes (e.g. achieving certain aspects of security by for example detecting attack patterns with ML) ◼ The usage of ML has significant impact on quality attributes, and thus needs architectural treatment ◼ One key aspect: missing comprehensibility / explainability what is happening in the ML- component ◼ Safety, reliability: conflicting with safety standards, needs counter-measures ◼ UX: Explaining to the user what happens / integrating user into overall flow

© Fraunhofer IESE 53 Quality Attributes in ML-based Systems (2/2) ◼ Fulfill the respective quality attributes of the system, respecting the overall “scale” of the system ◼ Performance (latency, throughput), scalability, … ◼ Considering the runtime system, but also the devtime / learning system ◼ Provide an adequate execution environment ◼ Sufficient computing power ◼ Sufficient storage capacity ◼ Provide the right data with adequate frequency and latency ◼ Architect has to know the requirements / implications of the ML algorithm / model

© Fraunhofer IESE 57 Conclusion: What does it mean for me? What can I do? ◼ Keep an eye on the architectural big picture, even if there is ML in the system ;-) ◼ Understand the very nature of ML-based systems ◼ Learn from existing systems and their solution approaches ◼ Remember the essentials of software architecture ◼ Achieving quality attributes ◼ Dealing with uncertainty ◼ Organizing and distributing work ◼ Fill your toolbox with knowledge about patterns and technologies in the ML-area ◼ Start working with data scientists / data engineers and establish a common language

© Fraunhofer IESE 58 dominik.rost@iese.fraunhofer.de https://www.linkedin.com/in/domrost/ https://twitter.com/domrost https://speakerdeck.com/domrost Blog https://blog.iese.fraunhofer.de/category/architecture/ https://blog.iese.fraunhofer.de/category/digitale-oekosysteme/ ObjektSpektrum 01/2021 Architekturtrends: Machine-Learning-basierte Systeme: Eine Einführung für Architekten matthias.naab@iese.fraunhofer.de https://www.linkedin.com/in/matthiasnaab/

© Fraunhofer IESE 59 Dr. Dominik Rost Dr. Matthias Naab 14.10.2020 Software Architecture Gathering 2020 Machine-Learning-basierte Systeme: Eine Einführung für Architekten