Automated Machine Learning, commonly known as AutoML, is revolutionizing the field of machine learning by streamlining and automating the traditionally complex and time-intensive process of model development. This innovation empowers data scientists, analysts, and developers to build high-quality machine learning models with greater speed, efficiency, and scale, all while maintaining and even enhancing model performance. Originating from groundbreaking research within Microsoft Research, AutoML in Azure Machine Learning exemplifies the advancements in this domain.
Understanding Automated Machine Learning: How Does AutoML Operate?
The core function of AutoML lies in its ability to automate the iterative tasks inherent in machine learning model creation. During the training phase, AutoML systematically generates numerous parallel pipelines, each exploring diverse algorithms and parameter configurations. This service intelligently cycles through various machine learning algorithms combined with feature selection techniques. Each iteration yields a model accompanied by a training score. This score serves as a crucial indicator – the higher the score for the chosen optimization metric, the better the model is deemed to “fit” the provided data. The automated process continues until it meets predefined exit criteria specified at the outset of the experiment.
Azure Machine Learning provides a robust platform to design and execute automated ML training experiments. This process can be broken down into clear, manageable steps:
- Define Your Machine Learning Challenge: Clearly identify the type of machine learning problem you aim to solve. This could range from classification, forecasting, and regression to more specialized areas like computer vision or natural language processing (NLP).
- Choose Your Preferred Experience: Code-First or No-Code: Tailor your approach based on your technical preference. For those who favor a code-centric approach, the Azure Machine Learning SDKv2 and Azure Machine Learning CLIv2 are excellent tools. For users seeking a more visual, less code-intensive experience, the web interface within Azure Machine Learning studio at https://ml.azure.com offers an intuitive alternative.
- Data Integration: Specify the source of your labeled training data. Azure Machine Learning supports a wide array of methods to bring your data into the platform.
- Configure AutoML Parameters: Define the parameters for your automated machine learning process. This includes setting the number of iterations across different models, adjusting hyperparameter settings, implementing advanced preprocessing and featurization techniques, and selecting the metrics to guide the model selection process.
- Initiate Training: Submit your training job to begin the automated model development.
- Evaluate Results: Once training is complete, thoroughly review the outcomes to identify the best performing model and gain insights from the experiment.
This process is visually summarized in the diagram above. Furthermore, you have the capability to delve into the logged job information, which includes comprehensive metrics gathered throughout the training process. The culmination of a successful training job is a Python serialized object (a .pkl file). This file encapsulates the trained model and the data preprocessing steps applied, ready for deployment and use.
While AutoML automates model construction, it also provides transparency and control. Users can explore feature importance and understand which features are most influential in the generated models.
Applications Across Industries: When to Leverage AutoML
Automated ML is particularly beneficial when you need Azure Machine Learning to handle the complexities of model training and tuning, guided by your specified target metric. AutoML democratizes machine learning, making it accessible and empowering users across various expertise levels to develop end-to-end machine learning pipelines for diverse problem sets.
Professionals and developers across numerous sectors can utilize automated ML to:
- Implement ML Solutions Without Extensive Coding: Automate the development process, reducing the need for deep programming expertise in machine learning.
- Optimize Time and Resource Utilization: Significantly cut down on the time and computational resources typically required for model development.
- Adopt Data Science Best Practices: Leverage built-in best practices in data preprocessing, model selection, and evaluation.
- Enhance Agile Problem-Solving: Rapidly prototype and deploy machine learning solutions, fostering agility in addressing business challenges.
Classification Tasks
Classification, a fundamental type of supervised learning, involves training models to categorize new data based on patterns learned from labeled training data. Azure Machine Learning is equipped with specialized featurization techniques for classification tasks, including deep neural network text featurizers. Detailed information on featurization options can be found in Data featurization documentation. A list of algorithms supported by AutoML is available in Supported algorithms.
The primary objective of classification models is to accurately predict the category to which new data points belong. Common applications of classification include fraud detection systems, handwriting recognition software, and object detection in images.
Explore a practical example of classification using automated machine learning in this Python notebook example: Bank Marketing.
Regression Tasks
Regression tasks, similar to classification, are a cornerstone of supervised learning. Azure Machine Learning offers tailored featurization methods specifically designed for regression problems. Learn more about these in featurization options and explore the algorithms supported by AutoML in Supported algorithms.
Unlike classification, which predicts categorical outputs, regression models forecast numerical output values. These predictions are based on independent predictor variables. Regression aims to model the relationships between these variables, estimating how changes in one variable affect others. For instance, a regression model could predict vehicle prices based on features like fuel efficiency and safety ratings.
For a hands-on example of regression with automated machine learning, see these Python notebooks focusing on predictions: Hardware Performance.
Time-Series Forecasting
Accurate forecasting is vital for business planning, whether for revenue projections, inventory management, sales forecasts, or understanding customer demand. Automated ML provides powerful tools to combine various forecasting techniques and methodologies, delivering reliable, high-quality time-series forecasts. The range of supported algorithms is detailed in Supported algorithms.
In AutoML, time-series forecasting is approached as a multivariate regression problem. Historical time-series data points are transformed into additional dimensions for the regression model, alongside other relevant predictors. This method offers a significant advantage over traditional time-series approaches by naturally integrating multiple contextual variables and their interrelationships during the training phase. AutoML can learn a unified, yet potentially branched, model applicable across all items in the dataset and prediction horizons, maximizing data utilization for parameter estimation and enhancing generalization to unseen data series.
Advanced forecasting configurations include:
- Holiday Effects: Incorporation of holiday data to account for predictable demand fluctuations.
- Specialized Time-Series Learners: Including models like Auto-ARIMA, Prophet, and ForecastTCN, tailored for time-series data.
- Group Forecasting: Support for handling and forecasting multiple related time series concurrently.
- Rolling-Origin Cross-Validation: Robust validation techniques to ensure model reliability over time.
- Configurable Lags: Customization of lag features to capture temporal dependencies effectively.
- Rolling Window Aggregate Features: Creation of aggregated features over time windows to smooth noise and highlight trends.
Explore a practical example of forecasting using automated machine learning in this Python notebook: Energy Demand.
Computer Vision Applications
Automated ML extends its capabilities to computer vision tasks, enabling the development of models trained on image data for applications such as image classification and object detection.
Key features for computer vision include:
- Seamless Integration with Azure Machine Learning Data Labeling: Direct compatibility with Azure Machine Learning’s data labeling tools for efficient data preparation.
- Image Model Generation from Labeled Data: Utilize labeled image datasets to train robust computer vision models.
- Performance Optimization: Fine-tune model performance by selecting appropriate algorithms and optimizing hyperparameters.
- Deployment Flexibility: Deploy trained models as web services within Azure Machine Learning or download them for local deployment.
- Scalable Operations: Leverage Azure Machine Learning MLOps and ML Pipelines for scalable model operationalization.
Model creation for vision tasks within AutoML is facilitated through the Azure Machine Learning Python SDK. Experimentation jobs, models, and outputs are accessible via the Azure Machine Learning studio UI.
Learn more about setting up AutoML training for computer vision models in this guide: AutoML for Computer Vision Models.
Image source: Stanford CS231n Lecture Slides
Automated ML supports several computer vision tasks, including:
| Task | Description |
|---|---|
| Multi-class Image Classification | Classifying images into one of several exclusive categories (e.g., classifying images as either ‘cat’, ‘dog’, or ‘duck’). |
| Multi-label Image Classification | Assigning one or more labels from a predefined set to each image (e.g., an image may be labeled as both ‘cat’ and ‘dog’). |
| Object Detection | Identifying and localizing objects within images by drawing bounding boxes around each detected object (e.g., finding all dogs and cats in an image and outlining them). |
| Instance Segmentation | Identifying objects at the pixel level, delineating each object with a polygon mask (e.g., precisely outlining each instance of an object in an image). |
Natural Language Processing (NLP) Capabilities
Automated ML extends its reach to natural language processing (NLP), enabling the creation of models from text data for tasks like text classification and named entity recognition. NLP model training in AutoML is supported via the Azure Machine Learning Python SDK, with experiment management and results accessible through the Azure Machine Learning studio UI.
Key NLP capabilities include:
- End-to-End Deep Learning NLP Training: Utilizing state-of-the-art pre-trained BERT models for advanced NLP tasks.
- Integration with Azure Machine Learning Data Labeling for Text: Seamlessly works with Azure Machine Learning’s text data labeling tools.
- NLP Model Generation from Labeled Text Data: Train custom NLP models using your labeled text datasets.
- Multi-lingual Support: Supports processing and understanding of 104 languages.
- Distributed Training: Leverages Horovod for efficient distributed training, speeding up model development.
Learn how to configure AutoML training for NLP models in this guide: AutoML for NLP Models.
Data Handling: Training, Validation, and Testing
In automated ML, you provide training data to build your machine learning models. You also have the flexibility to specify the type of model validation to be employed. Model validation is integral to the AutoML training process. Validation data is used to fine-tune model hyperparameters for each algorithm, aiming to find the configuration that best fits the training data. However, using the same validation data across iterations can introduce bias, as the model may become overly optimized to this specific dataset.
To mitigate this potential bias and ensure the robustness of the final model, AutoML supports the use of test data. By providing test data as part of your AutoML experiment setup, the final recommended model undergoes a rigorous evaluation at the conclusion of the experiment. This step is crucial for confirming that the model generalizes well to unseen data and that the performance metrics are reliable.
It’s important to note that the use of test data for model evaluation is currently a preview feature, offering an experimental way to enhance model validation.
Learn how to configure AutoML experiments to incorporate test data with the SDK or through the Azure Machine Learning studio interface.
Feature Engineering: Enhancing Model Learning
Feature engineering is a critical process in machine learning that involves using domain knowledge to create features that improve the learning process of ML algorithms. Azure Machine Learning incorporates scaling and normalization techniques as part of its feature engineering process. These techniques, along with other feature engineering steps, are collectively referred to as featurization.
In automated machine learning experiments, featurization is applied automatically. However, it can also be customized to suit the specifics of your data. Understanding the included featurization techniques and how AutoML addresses common challenges like overfitting and imbalanced data is crucial for effective model development.
It is important to note that the featurization steps applied by automated machine learning (such as feature normalization, handling missing data, and converting text to numerical formats) become embedded within the underlying model. This means that when the model is used for predictions, the same featurization steps are automatically applied to the input data, ensuring consistency and accuracy.
Customizing Featurization
For more advanced control, additional feature engineering techniques, including encoding and transforms, are available for customization, allowing you to tailor the data preprocessing to your specific needs.
These settings can be adjusted to refine the featurization process.
Ensemble Models: Combining Strengths
Automated machine learning inherently supports ensemble models, a feature enabled by default to enhance predictive performance. Ensemble learning combines multiple models to achieve superior results compared to relying on single models. In AutoML, ensemble iterations are typically the final stages of a training job. Two primary ensemble methods are used: voting and stacking.
- Voting: This method predicts outcomes based on a weighted average of predicted class probabilities (for classification) or predicted regression targets (for regression).
- Stacking: Stacking involves combining diverse models and training a meta-model on the outputs of these individual models. Currently, LogisticRegression is the default meta-model for classification tasks, while ElasticNet is used for regression and forecasting.
The Caruana ensemble selection algorithm, along with sorted ensemble initialization, is employed to determine which models are included in the ensemble. This algorithm starts by selecting up to five top-performing models and ensures they are within a 5% performance threshold of the best score to avoid starting with a suboptimal ensemble. In each iteration, a new model is added, and the ensemble’s score is recalculated. If the new model improves the ensemble score, it is incorporated into the ensemble.
Default ensemble settings in automated machine learning can be modified via the AutoML package.
AutoML and ONNX: Enhancing Interoperability
Azure Machine Learning enables you to build Python-based models using automated ML and convert them to the ONNX (Open Neural Network Exchange) format. ONNX facilitates greater interoperability, allowing models to be run across various platforms and devices. ONNX accelerates ML model deployment and broadens compatibility.
Instructions on converting to ONNX format are available in this Jupyter notebook example. A list of ONNX-compatible algorithms is provided in the supported algorithms documentation.
The ONNX runtime also supports C#, enabling the use of AutoML-generated models directly in C# applications without needing recoding or facing latency issues associated with REST endpoints. Detailed guides are available on using AutoML ONNX models in .NET applications with ML.NET and inferencing ONNX models with the ONNX runtime C# API.
Further Exploration: Next Steps with AutoML
Numerous resources are available to help you get started and deepen your understanding of AutoML.
Tutorials and How-To Guides
Tutorials offer comprehensive, step-by-step introductions to various AutoML scenarios, ideal for beginners. How-to articles provide detailed insights into specific functionalities and advanced features of automated ML, catering to users looking for more in-depth knowledge.
Jupyter Notebook Examples
For practical, code-focused learning, explore the extensive collection of Jupyter notebook samples in the GitHub repository for automated machine learning samples. These notebooks cover a wide range of use cases and provide valuable hands-on experience.
Python SDK Reference
For developers and advanced users, the AutoML Job class reference documentation offers detailed specifications of SDK design patterns and class details, essential for programmatic interaction with AutoML.
Automated machine learning capabilities are also integrated into other Microsoft solutions, including ML.NET, HDInsight, Power BI, and SQL Server, extending its accessibility and utility across different platforms and applications.
