Types of AI Workloads

What Are AI Workloads?

AI workloads are specific tasks or problem categories that artificial intelligence systems are designed to solve. On the AI-900 exam, you must differentiate between the three primary machine learning paradigms—supervised, unsupervised, and reinforcement learning—and recognize common AI workload types like computer vision, NLP, conversational AI, and anomaly detection. Each workload uses distinct algorithms, data requirements, and evaluation metrics.

Supervised Learning: Learning with Labels

Supervised learning is the most common workload. The algorithm is trained on a dataset where each example has an input (features) and a correct output (label). The goal is to learn a mapping from inputs to outputs so that it can predict labels for new, unseen data. - How it works: The model processes input features, makes a prediction, compares it to the true label using a loss function (e.g., mean squared error for regression, cross-entropy for classification), and adjusts its internal parameters via optimization (e.g., gradient descent) to minimize the loss. This process repeats over many epochs. - Key components: Features (independent variables), labels (target variable), training set, validation set, test set, loss function, optimizer. - Common algorithms: Linear regression, logistic regression, decision trees, random forests, support vector machines, neural networks. - Evaluation metrics: Accuracy, precision, recall, F1-score (classification); R-squared, mean absolute error (regression). - Azure services: Azure Machine Learning, Automated ML, Azure Cognitive Services (e.g., Custom Vision for classification).

Unsupervised Learning: Finding Hidden Patterns

Unsupervised learning works with unlabeled data. The algorithm must discover inherent structures, patterns, or groupings without any guidance. - How it works: The model identifies similarities or differences between data points based on their features. For clustering, it groups points so that intra-cluster similarity is high and inter-cluster similarity is low. For dimensionality reduction, it projects high-dimensional data into fewer dimensions while preserving variance. - Key components: No labels; only features. Distance metrics (e.g., Euclidean, cosine) for clustering. Number of clusters (k) must be specified or determined via heuristics. - Common algorithms: K-means clustering, hierarchical clustering, DBSCAN, principal component analysis (PCA), t-SNE. - Evaluation metrics: Silhouette score, inertia, Davies-Bouldin index (clustering). Explained variance (PCA). - Azure services: Azure Machine Learning (K-means module), Azure Cognitive Services (Anomaly Detector – uses unsupervised approach).

Reinforcement Learning: Learning from Rewards

Reinforcement learning involves an agent that interacts with an environment. The agent takes actions, receives rewards or penalties, and learns a policy that maximizes cumulative reward over time. - How it works: The agent observes the current state, selects an action using an exploration-exploitation strategy (e.g., epsilon-greedy), receives a reward and the next state, and updates its policy (e.g., via Q-learning or policy gradients). The process continues until the agent converges to an optimal policy. - Key components: Agent, environment, state, action, reward, policy, value function, Q-function, discount factor (gamma, typically 0.9–0.99), learning rate (alpha, e.g., 0.1). - Common algorithms: Q-learning, Deep Q-Networks (DQN), Policy Gradients, Proximal Policy Optimization (PPO). - Evaluation metrics: Cumulative reward, average reward per episode, convergence time. - Azure services: Azure Machine Learning (RL via Ray), Azure Cognitive Services (Personalizer – uses contextual bandits, a simplified RL).

Conversational AI: Understanding and Generating Dialogue

Conversational AI workloads power chatbots and virtual assistants. They combine natural language understanding (NLU) to interpret user intent and natural language generation (NLG) to produce responses. - How it works: The system receives text or speech, uses a language model to extract intent and entities (e.g., 'BookFlight' intent with 'destination' entity 'Paris'), retrieves or generates a response, and optionally sends it to a text-to-speech engine. - Key components: Intent recognition, entity extraction, dialogue management, response generation, language model. - Azure services: Azure Bot Service, Language Understanding (LUIS) – now part of Azure Cognitive Service for Language, QnA Maker (now part of Azure Cognitive Service for Language – custom question answering), Azure Speech Services. - Evaluation metrics: Intent accuracy, entity F1-score, user satisfaction, average conversation length.

Computer Vision: Interpreting Visual Data

Computer vision workloads process images and videos to extract information. Tasks include image classification, object detection, segmentation, optical character recognition (OCR), and facial recognition. - How it works: Convolutional neural networks (CNNs) extract hierarchical features from pixels. For classification, a CNN outputs probabilities for each class. For object detection, it also predicts bounding boxes. For segmentation, it assigns a class to every pixel. - Key components: Input image dimensions (e.g., 224x224x3 for ResNet), convolutional layers, pooling layers, fully connected layers, loss function (e.g., cross-entropy, IoU loss). - Azure services: Azure Computer Vision, Custom Vision, Face API, Form Recognizer (now Azure AI Document Intelligence). - Evaluation metrics: Accuracy, mean average precision (mAP) for detection, intersection over union (IoU) for segmentation.

Natural Language Processing (NLP): Understanding Text

NLP workloads analyze and generate human language. Key tasks include sentiment analysis, key phrase extraction, language detection, text classification, summarization, and translation. - How it works: Text is tokenized into words or subwords, converted to embeddings (e.g., word2vec, BERT embeddings), and processed by a model (e.g., transformer) to perform the task. For translation, an encoder-decoder architecture is common. - Key components: Tokenizer, embedding dimension (e.g., 768 for BERT), attention mechanism, transformer layers (e.g., 12 for BERT-base). - Azure services: Azure Cognitive Service for Language (includes sentiment analysis, key phrase extraction, language detection, custom text classification), Azure Translator, Azure Speech Services. - Evaluation metrics: Accuracy, F1-score, BLEU score (translation), ROUGE score (summarization).

Anomaly Detection: Identifying Outliers

Anomaly detection identifies data points that deviate significantly from the norm. It is used for fraud detection, equipment monitoring, and security. - How it works: The model learns the normal pattern of data (e.g., using statistical methods or autoencoders). New data points are scored based on their deviation; points with scores above a threshold are flagged as anomalies. - Key components: Training data (mostly normal), anomaly score, threshold (e.g., 95th percentile). - Common algorithms: One-class SVM, isolation forest, autoencoders, ARIMA (time series). - Azure services: Azure Anomaly Detector (part of Cognitive Services), Azure Machine Learning (built-in anomaly detection modules). - Evaluation metrics: Precision, recall, F1-score at various thresholds, area under ROC curve.

How These Workloads Relate to Azure

Azure offers a unified platform for these workloads through Azure Machine Learning (for custom model training) and Azure Cognitive Services (pre-built AI models). The exam expects you to know which Azure service maps to which workload. For example, if a scenario describes identifying objects in images, you should select Azure Computer Vision. If it describes training a model to classify customer feedback into positive/negative, you should select Azure Cognitive Service for Language (sentiment analysis) or Custom Text Classification.

Exam Relevance

AI-900 tests your ability to identify the correct workload type for a given business problem. Common questions present a scenario and ask which Azure service or ML paradigm to use. You must also distinguish between supervised, unsupervised, and reinforcement learning based on data labeling and goal. For example, if data is labeled and you want to predict a category, it's supervised classification. If data is unlabeled and you want to group customers, it's unsupervised clustering.

Enterprise Scenario 1: Retail Customer Segmentation

A large retail chain wants to group its customers based on purchasing behavior to tailor marketing campaigns. They have transaction data with features like total spend, frequency, and product categories purchased, but no customer labels. This is an unsupervised learning problem. The data science team uses Azure Machine Learning with K-means clustering. They set k=5 after elbow analysis. The model identifies five segments: budget shoppers, luxury buyers, frequent small purchasers, seasonal buyers, and inactive customers. The marketing team then designs targeted promotions. Common pitfall: choosing too many clusters leads to overfitting; too few misses important distinctions. The team uses silhouette score (target >0.5) to validate. Performance consideration: clustering on millions of customers requires scaling to distributed compute (e.g., Azure Spark ML). Misconfiguration: using Euclidean distance on unscaled features (e.g., income vs. number of visits) skews results.

Enterprise Scenario 2: Manufacturing Defect Detection

A factory produces circuit boards and needs to automatically detect soldering defects from camera images. They have thousands of labeled images: 'defective' and 'non-defective'. This is a supervised image classification problem. They use Azure Custom Vision to train a model with transfer learning (ResNet backbone). After training with 500 images per class, the model achieves 98% accuracy. It is deployed as a container on the factory edge device via Azure IoT Edge. Common issue: imbalanced classes (only 5% defective) – they use oversampling. Performance: inference must complete within 100ms per image; they optimize by reducing image resolution from 1024x1024 to 320x320. Misconfiguration: not retraining periodically as new defect types appear.

Enterprise Scenario 3: Chatbot for Customer Support

A telecom company wants a chatbot to handle common inquiries like bill payments and plan changes. They use Azure Bot Service with Language Understanding (LUIS) for intent recognition. They define intents: 'CheckBalance', 'PayBill', 'ChangePlan', and entities: 'amount', 'date'. They train with 50 example utterances per intent. The chatbot uses QnA Maker for FAQs. When a user says 'I want to pay my bill', LUIS identifies intent 'PayBill' with entity 'bill'. The bot then prompts for payment details. Common pitfalls: overlapping intents (e.g., 'PayBill' and 'CheckBalance' both triggered by 'bill') – they use phrase lists and add negative examples. Performance: latency must be under 2 seconds; they cache common responses. Misconfiguration: not handling out-of-scope queries – they add a fallback intent.