The Data-Centric AI Playbook

Here are a few options, all similar in length and capturing the core meaning: * **Engineering Quality Datasets: A Guide for Data-Scarce Scenarios.** * **Building Robust Datasets: Techniques for Limited, Imperfect Data.** * **A Web Guide: Creating High-Quality Datasets from Challenging Data.** * **Overcoming Data Challenges: A Guide to Building Datasets.** * **Data Scarcity Solutions: A Web Guide to Dataset Engineering.**

The Paradigm Shift to Data-Centric AI

For years, AI advanced by refining models. Now, the key is data engineering. This change stems from limitations in the model-first approach.

0

% of AI Projects Fail

due to issues with data quality, not model flaws.

0

% of Research was Model-Centric

historically, leading to a focus on code over data.

2023

Projected Year of Data Exhaustion

This line rewritten with a similar size: **High-quality public text demands data engineering expertise.**

Toolkit: Programmatic Labeling with Weak Supervision

Lacking labels but with lots of data? Weak supervision lets you build a training set using "Labeling Functions" (LFs), which are heuristic rules derived from domain expertise.

The Weak Supervision Pipeline

1

Write Labeling Functions (LFs)

Here are a few options for rewriting the line while maintaining a similar size and meaning: * **Implement label voting or abstention using encoded heuristics (keywords, patterns, LLM prompts).** * **Use encoded heuristics (searches, patterns, prompts) for programmatic label voting/abstention.** * **Programmatically vote or abstain on labels through encoded heuristics (keyword, pattern, LLM methods).** * **Leverage encoded heuristics (keywords, patterns, prompts) to programmatically assign labels or abstain.**

2

Train Generative Label Model

Here are a few options, all similar in length and capturing the core idea: * This model gauges LF performance via agreement/disagreement, without requiring ground truth. * The model learns LF accuracy and relationships from their consensus and conflicts, without labels. * By analyzing LF agreement and disagreement, this model estimates performance; no truth data is used. * This approach assesses LFs by observing their agreement and disagreement patterns, foregoing ground truth.

3

Generate Probabilistic Labels

Here are a few options, all of similar length, capturing the essence of the original: * The output provides a complete training set with probabilistic ('soft') labels reflecting the model's certainty. * This output is a comprehensive training set with 'soft' labels (e.g., 90% Class A) to show model confidence. * Generated output: a full training set with soft labels indicating the model's aggregate confidence. * Result: A training set with 'soft' labels (e.g., 90% Class A), representing the model's confidence.

4

Train Discriminative End Model

Here are a few options, aiming for a similar size and meaning: * A strong end model (like a Transformer) uses probabilistic labels to surpass LF heuristic limitations. * By learning from probabilistic labels, a powerful model (e.g., Transformer) extends beyond LF-based rules. * Unlike simple LFs, a robust end model (like a Transformer) exploits probabilistic labels for broader generalization. * Leveraging probabilistic labels, a high-performing model (e.g., Transformer) achieves generalization beyond LF heuristics.

Toolkit: Efficient Data Selection with Active Learning

To boost model performance on a tight budget, Active Learning strategically prioritizes data points for human labeling.

The Query Strategy Explorer

Here are a few options, all aiming for similar length and clarity: * Active learners' "brains" are query strategies. Explore a strategy below to grasp its data-labeling core. * The query strategy is the "brain" of active learning. Pick a strategy to learn its data selection logic. * An active learner's core is its query strategy. Choose one below and understand its data-labeling process. * Query strategies form the active learner's core. Select a strategy to learn how it decides what to label.

Toolkit: Creating New Data with Generative Methods

For dataset augmentation, handling outliers, and boosting data volume, generative techniques are key. Optimal Transport furnishes a principled, geometric approach.

Principled Augmentation with Optimal Transport

Naive Interpolation (e.g., Mixup)

Averaging data risks generating artificial samples outside the realistic data distribution, which hurts model learning.

Wasserstein Barycenters (OT)

OT generates plausible, high-quality samples by identifying a geometry-aware "average" that aligns with the data.

Build Your Strategy: A Unified Framework

Here's a rewritten line of similar size and tone: **Synergy unlocks the greatest potential. Consult this guide to tailor your data strategy.**

Recommended Strategy:

Your recommended workflow will appear here...