AIF-C01 Fundamentals of AI and ML Fundamentals of AI and ML Practice Questions

Accuracy is the ratio of correct predictions to the total number of predictions, so it directly measures how many part images the model classified correctly, which is exactly what the team wants for this classification task.

Root mean squared error (RMSE) and mean absolute error (MAE) are regression metrics that quantify the difference between predicted and actual continuous values, so neither is suited to measuring classification correctness. Number of training epochs is a training configuration setting that controls how many passes the model makes over the data; it is not an evaluation metric at all.

Asynchronous inference is built for large payloads of up to 1 GB and long processing times of up to an hour; it queues incoming requests and processes them in the background, returning results when ready without requiring the caller to wait synchronously, which suits workloads that need near real-time handling of demanding requests.

Real-time inference targets small payloads and fast synchronous responses, and its payload and timeout limits would be exceeded by 1 GB inputs and hour-long processing. Serverless inference suits intermittent traffic with smaller payloads and likewise cannot handle such large inputs or long durations. Batch transform processes large datasets offline in bulk and is not designed to deliver near real-time latency.

Transfer learning lets a company take a model already trained on one task and adapt it to new, related tasks, which reuses the knowledge captured in the pre-trained model and saves significant time and data compared to training from scratch.

Increasing the number of epochs simply trains a model for more passes over the data and does not provide a way to reuse an existing pre-trained model. Decreasing the number of epochs reduces training passes and likewise has nothing to do with adapting pre-trained models for new tasks. Unsupervised learning finds patterns in unlabeled data but is a learning approach rather than a strategy for reusing and adapting an already trained model.

The probability here is a deterministic calculation of favorable outcomes divided by total outcomes, so a few lines of code applying basic arithmetic produce an exact answer with no data to gather, no model to train, and no endpoint to host, giving the least operational overhead.

The three machine learning options all stand up an unnecessary pipeline of data collection, training, tuning, and hosting for a problem that already has a known closed-form answer. Beyond that shared waste, supervised regression would only approximate a value that can be computed exactly; unsupervised learning is meant to uncover structure in unlabeled data, which is irrelevant when the formula is known; and reinforcement learning is designed for sequential decision-making guided by rewards in an environment, not for answering a static arithmetic question.