Practice Exam

B) Generative AI. Generative AI models create new content (text, images, code, audio) that did not exist before. Examples include GPT (text), DALL-E (images), and GitHub Copilot (code). Predictive AI forecasts outcomes from data but does not create new content.

C) Transformer. The Transformer architecture, introduced in the 2017 paper "Attention Is All You Need," uses self-attention mechanisms to process all tokens in parallel. It is the foundation of GPT, BERT, Claude, and virtually all modern LLMs.

C) Unsupervised clustering. Finding groups in unlabeled data without predefined categories is clustering, a form of unsupervised learning. Common algorithms include K-Means, DBSCAN, and hierarchical clustering.

B) Word embedding. Word embeddings (Word2Vec, GloVe, BERT embeddings) represent words as dense vectors in a continuous space. Similar words have similar vectors. "King" and "queen" are close together; "king" and "bicycle" are far apart. Tokenization splits text into tokens; one-hot encoding creates sparse binary vectors.

C) Object detection. Object detection identifies multiple objects within an image and draws bounding boxes around each one with a class label. Image classification assigns a single label to the entire image. Image segmentation classifies each pixel.

B) Narrow AI performs specific tasks; general AI can perform any intellectual task. All current AI systems are narrow AI (also called weak AI) — they excel at specific tasks but cannot generalize. Artificial General Intelligence (AGI or strong AI) would match human-level intelligence across all domains, but it does not yet exist.

B) Predicting the next token based on the preceding context. GPT (Generative Pre-trained Transformer) models are autoregressive — they generate text one token at a time by predicting the most likely next token given all previous tokens. They do not store or retrieve facts from a database.

C) Test set. The test set is held out and never seen during training or hyperparameter tuning. It provides an unbiased final evaluation of the model's performance on unseen data. The validation set is used during training for tuning.

B) Low precision (false positives). Blocking legitimate emails means the model is incorrectly predicting "spam" when it should not be — these are false positives. Precision measures how many positive predictions (spam) were actually correct. Low precision means too many false positives.

B) Scales features to a common range (e.g., 0 to 1). Normalization rescales feature values so they fall within a consistent range. This prevents features with large magnitudes (like salary: 50,000-200,000) from dominating features with small magnitudes (like age: 20-70) during model training.

B) Data augmentation. Data augmentation creates additional training examples by applying transformations (rotation, flipping, cropping, color changes) to existing data. It helps prevent overfitting and improves model generalization, especially when training data is limited.

B) Learning rate. The learning rate determines the step size when updating weights during gradient descent. Too high: training is unstable, may not converge. Too low: training is extremely slow. Finding the right learning rate is one of the most important tuning decisions.

B) Evaluate model performance more reliably by training on multiple data splits. In k-fold cross-validation, data is split into k parts. The model trains on k-1 parts and validates on the remaining part, rotating through all k folds. This gives a more robust performance estimate than a single train-test split.

C) Edge inference on the vehicle's onboard computer. A 50ms latency requirement eliminates cloud inference (network round-trip adds significant latency). Edge deployment runs the model locally on the device, ensuring immediate response times critical for safety applications.

B) Concept drift. When the real-world relationship between inputs and outputs changes over time, the model's learned patterns become stale. This is concept drift. The solution is to monitor model performance continuously and retrain with recent data.

B) Containers (Docker). Containers package the model, runtime, libraries, and configuration together, ensuring the model runs identically in development, testing, and production. This eliminates "it works on my machine" problems common in ML deployment.

B) A/B testing (canary deployment). Directing a small percentage of traffic to the new model allows you to compare real-world performance before full rollout. If the new model performs worse, you only affect 10% of users and can quickly rollback.

B) Human-in-the-loop. The AI assists the human expert (radiologist) by flagging potential cases, but the final decision remains with the human. This is the recommended approach for high-stakes decisions in healthcare, legal, and financial domains.

B) Predictive maintenance. Predictive maintenance uses sensor data (vibration, temperature, pressure) and ML models to predict when equipment will fail, allowing maintenance to be scheduled before a breakdown occurs. This reduces downtime and repair costs compared to reactive or scheduled maintenance.

B) Selection bias in training data. The training dataset likely over-represented lighter-skinned individuals, causing the model to perform poorly on underrepresented groups. This is a well-documented bias in facial recognition systems and requires diversifying the training data.

B) Right to erasure (right to be forgotten). Under GDPR Article 17, individuals can request that their personal data be deleted. This poses challenges for AI/ML systems because training data may need to be removed and models may need to be retrained.

B) Model Card. Model Cards are standardized documentation for AI models, introduced by Google in 2019. They describe what the model does, how it was trained, its performance across different demographics, known limitations, and ethical considerations. They are a key governance tool.

B) Automation bias. Automation bias is the tendency to over-rely on automated systems and defer to AI decisions without questioning them. This is dangerous in high-stakes scenarios. Mitigation includes training users to critically evaluate AI outputs and implementing human oversight.

B) Data poisoning. Data poisoning targets the training phase by injecting malicious or manipulated data to corrupt the model's behavior. Unlike adversarial attacks (which target the inference phase with crafted inputs), data poisoning compromises the model itself during training.

C) Explainability. Explainability means AI decisions can be understood and explained to affected stakeholders. Techniques like SHAP values, LIME, and feature importance scores provide explainability. It is especially critical for regulated industries (healthcare, finance, hiring) where decisions must be justified.