How to Create a Dataset for Machine Learning | #AI101
The Three Pillars of Dataset Creation
EverydAI Host outlines the three fundamental stages of creating a dataset: data collection, data cleaning, and data labeling. This structured approach is essential before any model training can commence, with each step presenting unique challenges and requiring dedicated effort.
Data Collection Strategies
The host details various data collection methods, from leveraging existing public or private datasets (like Google's YouTube-8M or hospital data agreements) to generating entirely new data through personal collection, crowdsourcing (e.g., Amazon Mechanical Turk), or citizen science initiatives. The choice depends heavily on the problem and data type.
Enhancing Datasets: Augmentation and Synthesis
When data is insufficient or needs refinement, techniques like data augmentation (e.g., rotating/flipping images) and data synthesis using Generative Adversarial Networks (GANs) can expand existing datasets. This allows for more data without new collection efforts, though careful application is needed to maintain representativeness.
The Nuances of Data Cleaning
Data cleaning involves filtering, cropping, reorganizing, and altering data to remove unwanted elements and improve usability. The critical caveat is that cleaning should not compromise the dataset's representativeness of the target population, preventing over-claiming.
The Necessity of Data Labeling
For supervised learning, data labeling is indispensable, assigning meaningful tags or categories to data points. This process can be highly subjective and labor-intensive, often necessitating crowdsourcing or gamified approaches like 'Stall Catchers' to expedite completion and engage participants.
Iterative Refinement of Datasets
Training a model often reveals dataset issues that impact outcomes, necessitating a return to the data collection, cleaning, or labeling stages. This iterative process is crucial for building better AI systems, emphasizing that dataset creation is not a one-off task.
