The Case for Your AI Modeling Approach

Generative artificial intelligence (genAI) is a branch of AI focused on creating new, human-like content. Where traditional machine learning (ML) models learn from existing data and make predictions based on patterns, genAI models create new and realistic data that mimics the original data provided to them. Although genAI is not a 'cure-all,' it’s a great supplement to human creativity and intelligence. In this article, I’ll provide my key scenarios for using either genAI or traditional ML models alone, as well as scenarios and indicators for blending the two approaches by augmenting one with the other. (For purposes of this article, I will use the terms ‘data’ and ‘content’ interchangeably.)

Traditional ML models are usually based on supervised learning, meaning they require labeled data to train and evaluate their performance. For example, an ML model that can classify images of cats needs a dataset of images labeled as cat. Supervised learning models are limited by the quality and quantity of data provided to them but genAI models, which are usually based on unsupervised learning, are not. They don’t need labeled data for training and would be able to generate realistic images of cats without having a dataset of images labeled as cat.

Since traditional ML models are great at making predictions or decisions based on existing data, here are a few scenarios where I would recommend using them:

• Analysis: finding patterns, trends or insights (e.g., detecting faces in images)

• Classification: labeling data or groups of data into categories (e.g., classifying images of cats for animal identification purposes)

• Recommendations: suggest content for a specific purpose (e.g., movies based on interest or prior purchase) Although genAI models cannot make factual predictions, they can generate new and realistic data based on learnings. Here are a few scenarios where I recommend using genAI models:

• Summarization: creating abbreviated versions of data (e.g., paraphrasing complexly worded statutory laws for better understanding.

  ​Although genAI is not a 'cure-all,' it’s a great supplement to human creativity and intelligence  

• Deep Retrieval: searching for answers within, or asking questions of, a large body of data (e.g., asking questions about unstructured insurance claims, maybe interacting with a chatbot, etc.).

• Transformation: converting existing data to different formats (e.g., translating a document from English to Spanish, etc.)

• Augmentation: updating data to include additional points (e.g., incorporating personalization into college acceptance letters, etc.)

• Net-new Creation: creating new data based on a set of instructions (e.g., create a Python code snippet from pseudocode, etc.)

• Anomaly Detection: identifying abnormalities within data without prior knowledge of what’s considered uncharacteristic

Since using pre-trained genAI models alone may not always address business needs, here are a few scenarios where I suggest augmenting genAI models with custom ML models:

• Quality Control: genAI generates new data that mimics data provided to it, but that new data may not always be factual. Use models customized and trained with data unique to your business to determine if genAI-created data is factual, filtering out anything irrelevant.

• Domain Adaptation: genAI can generate data on any topic but may not be able to capture domain nuances. To ensure business or industry relevance, augment genAI models with custom models that fine-tune output to include the vocabulary, style or tone that makes sense for your business (e.g., legal terms, medical jargon, etc.).

As you devise your AI approach, I also recommend setting thresholds in the following areas to help you identify when it makes sense to augment (or even replace) a genAI-only approach with custom models:

• Accuracy: Identify the minimum level of accuracy genAIproduced data needs to be accepted. For images, this might be represented by the number of pixels – if pixel expectation falls below a defined threshold then consider using trained models that generate the right level of accuracy.

• Diversity: Identify a threshold for the minimum amount of variety or range the generated data needs to serve its purpose.

• Relevance: Identify the minimum level of relevance to which the generated data should be applicable. The relevance threshold can be a measure of semantic similarity or topic coherence, something that reflects the connection of the generated data to the input or context.

• Appropriateness: Identify the minimum level of appropriateness the generated data should have to be suitable. This threshold can be set as a measure of sentiment analysis or emotion recognition, something that reflects the suitability of the generated data for its purpose.

When accuracy, diversity, relevance or appropriateness fall below a defined threshold, consider augmenting (or replacing) your genAI models with custom models.

Finally, don’t forget to consider implementation costs (computational resources, processing time, configuration and deployment effort) and complexity (technical skills needed, ethical issues to consider, percentage of use cases addressed, customizations to ‘fit’ business purpose) when determining your AI approach. Generative AI is a fast-growing branch of AI, so I hope these tidbits help you on your journey!