Machine Learning Fundamentals Quiz

The term "deep learning" refers to the use of artificial neural networks, which are complex structures inspired by the human brain, for learning tasks.

Supervised learning is a type of machine learning where the algorithm is trained on labeled data, meaning each data point has a corresponding output or target value. This allows the algorithm to learn the relationship between inputs and outputs and predict outcomes for new data.

The availability of vast amounts of data generated by digital activities on devices like smartphones and computers has fueled the growth of machine learning.

The ability to train large and complex neural networks, with increasing computational power, has significantly contributed to the progress of deep learning.

The performance of older machine learning algorithms like logistic regression often plateaus, meaning they stop improving even with more data. This happens because they might not be designed to effectively handle the complexity of large datasets.

Even with more data, older algorithms can sometimes fall short of the performance achieved by deep learning models, especially when dealing with complex patterns.

Larger neural networks, with more hidden layers and parameters, can often achieve better performance on complex tasks by capturing intricate patterns and relationships.

The performance improvements achieved by increasing the size of neural networks are often accompanied by a need for larger datasets to train them effectively.

Your training data should represent the real-world scenarios you want your model to perform well on.

Your 'dev set' and 'test set' should be drawn from the same distribution as the real-world data you want to model.

If your 'dev set' and 'test set' have different distributions, your model might perform well on one but not the other.

When you're focused on improving dev set performance, ensure you are not simply overfitting to the specific data in the dev set.

The test set might be inherently more challenging than the dev set, making it difficult to improve model performance.

The dev set and test set might have different characteristics that lead to different model performance, even if the test set isn't inherently harder.

When dev and test sets come from different distributions, it's hard to tell if improvements on the dev set actually translate to improved performance on the test set.

Focus on improving your model's performance on the real-world data distribution, not just on a specific dev set.

The performance of machine learning models can vary significantly depending on the size of the training dataset. In general, neural networks excel with massive datasets, while traditional algorithms may perform well with smaller datasets, particularly if features are carefully engineered.

When a machine learning model performs well on the training data but poorly on unseen data, it indicates that the model has not generalized well. This means it has learned specific patterns from the training data, but struggles to apply these patterns to new, real-world scenarios.

Hand-engineered features are manually crafted representations of data that highlight important characteristics relevant to the task. These features require domain expertise and can significantly improve the performance of traditional algorithms.

A development set is a subset of data used to evaluate and tune hyperparameters, including the model's architecture and learning rate. It helps to find the optimal configuration for the model before deploying it on the test set.

A test set is a completely unseen dataset used to measure the final performance of a machine learning model. It provides an objective evaluation of the model's ability to generalize to new, unseen data.

In machine learning, a common practice is to split data into training and test sets using a random 70%/30% split. This approach assumes a homogeneous data distribution, which might not hold true in real-world scenarios.

When training data and test data differ significantly in their underlying distribution, the model may perform poorly on the test set due to its inability to generalize to the real-world data distribution. This is known as data mismatch or distribution shift.

The primary metric used to judge the performance of your model. It is the one you strive to optimize by adjusting your system's parameters.

Criteria that are set to specific, acceptable values, ensuring a certain level of performance. They help constrain optimization efforts.

The process of iteratively designing, implementing, testing, and refining a machine learning system to achieve the desired performance.

The measure used to evaluate the performance of your machine learning model during training and on the dev set, providing a concrete gauge of its effectiveness.

The frequency with which a system incorrectly activates when the target trigger is not present, often applied to voice recognition systems.

The frequency with which a system fails to detect the target trigger when it is present, often applied to voice recognition systems.

A process of selecting a specific performance metric to prioritize and optimize while keeping other metrics within acceptable ranges.

The testing data you use for your machine learning model should closely resemble the real-world data you expect to encounter.

The metric used to evaluate your model should accurately reflect the true goal of your project. Choosing an inappropriate metric can lead to misleading results and decisions.

When your chosen metric doesn't fully capture the desired outcome of your project, it's important to adjust the metric to better align with your project's goals.

During the development process, it's common to make adjustments to the development and test sets, as well as the evaluation metric, to ensure they are relevant and useful throughout the lifecycle of your project.

An initial development and testing setup, even if not perfect, helps you get started and iterate quickly during the early stages of your project.

The practice of assessing your model on the final testing data only after you've completed all your development and tuning.

Constantly evaluating on the same test data during development can lead to overfitting to the test set, making the results less reliable for future performance.

The desired performance level you hope your machine learning algorithm will reach, for example human-level accuracy or the performance needed for a satisfactory user experience.

A visual representation that shows how the performance of your machine learning model changes as the size of the training data increases. It's like a graph that tracks your model's progress.

The error rate calculated on the training data. It typically decreases as the amount of training data increases.

The error rate calculated on a separate dataset that wasn't used for training. It usually decreases as the amount of training data grows.

The error rate on a completely independent dataset that is not used for training or validation. Serves as a final evaluation of the model's ability to generalize.

When the model sees the same training examples multiple times during training, it can 'memorize' these specific instances, leading to low training error but poor generalization to new data.