Google

A training dataset is a large partition of data used to teach and optimize a machine learning model. Unlike a test dataset which is held out strictly for evaluation, the training dataset is directly used during the learning process to update model parameters. For example, the Fashion-MNIST dataset contains a training set of $$60,000$$ images.

Training Dataset

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

Nonparametric models are highly flexible statistical descriptions that do not rely on strong, simplifying assumptions about the underlying data generation process. A defining common theme of nonparametric approaches is that their level of complexity grows as the amount of available training data increases. While parametric models are necessary when data is scarce to prevent overfitting, the abundance of data in modern machine learning allows for the use of fully nonparametric models that can better fit complex realities.

Nonparametric Models

The Fashion-MNIST dataset is a popular benchmark dataset containing grayscale images from $$10$$ different categories of clothing and accessories. It is structurally similar to the original MNIST dataset but offers a more complex alternative to handwritten digits. The dataset is divided into a training set of $$60,000$$ images ($$6,000$$ per category) and a test dataset of $$10,000$$ images ($$1,000$$ per category). In practice, these images are often upscaled to a resolution of $$32 \times 32$$ pixels for use in deep learning frameworks.

Fashion-MNIST Dataset

A realizable dataset is one in which the classes are inherently and truly separable, containing no label noise. For example, a dataset composed of unambiguous images for distinguishing cats from dogs is considered realizable. When training on such cleanly separated data, regularization techniques like early stopping typically do not lead to significant improvements in generalization.

Definition of Realizable Dataset

Which statement best describes the role of a training dataset?

Is the training dataset held out strictly for evaluation?

The learning algorithm runs on the _____ dataset.

Match each dataset concept with its correct description.

Order the steps for creating and using the cat training set.

Question:
In a concise analytical response, explain how the training dataset differs from the test dataset and why their uses should not be confused.

Sample answer:
The learning algorithm runs on the training dataset and uses its examples to update model parameters. The test dataset, by contrast, is held out strictly for evaluation. Using these partitions for their distinct purposes separates the data used for learning from the data reserved to assess the model.

Key points:
- The learning algorithm runs on the training dataset.
- Training examples are used to update model parameters.
- The test dataset is held out strictly for evaluation.
- Training and evaluation data have distinct roles.

Rubric:
A strong response clearly identifies the training set as learning data, states that it updates model parameters, identifies the test set as held-out evaluation data, and explains the significance of the distinction.

Explain why training and test datasets have different roles.

Case context:
A team downloads cat and non-cat pictures from different websites. It labels them as positive and negative examples but refuses to run its learning algorithm on them because it believes every dataset must remain untouched until evaluation.

Question:
What has the team misunderstood, and how should it use these pictures?

Sample answer:
The team has confused a training dataset with a test dataset. It should assemble the labeled cat and non-cat pictures into a large training set and run the learning algorithm on that set so the model parameters can be updated. Data reserved as a test set should remain held out for evaluation.

Key points:
- The downloaded pictures can form a training set.
- Cat pictures are positive examples.
- Non-cat pictures are negative examples.
- The learning algorithm should run on the training set.
- A test dataset is held out for evaluation.

Rubric:
The response should diagnose the training-versus-test confusion and recommend using the labeled pictures for learning while keeping test data reserved for evaluation.

Diagnose the mistake in a team's use of cat-image data.

Question:
Answer in one to three sentences: How does a machine learning system use a training dataset?

Sample answer:
The system runs its learning algorithm on the training dataset. The examples in that dataset are used to update and optimize the model's parameters.

Key points:
- The learning algorithm runs on the training dataset.
- Training data supports parameter updates and optimization.

Rubric:
The answer should mention both running the learning algorithm on the training data and updating or optimizing model parameters.

Learn Before

Related