A featured contribution from Leadership Perspectives, a curated forum for enterprise technology leaders, nominated by our subscribers and vetted by the CIOApplications Editorial Board.
Unum
Choosing the Right AI Model for Your Application
OpenAI launched ChatGPT in late 2022 to 1 million users in its first 5 days, then acquired 100 million users by January 2023 and 180 billion by March 2024. But how can you harness the power of ChatGPT’s underlying technology for your application?
ChatGPT is based on generative machine learning models pre-trained on an insanely massive body of data (think everything on the internet through 2023, depending on the ChatGPT version). Where traditional machine learning models use data provided to them to make predictions, generative models can create new and realistic content using the data on which it was trained. Artificial intelligence (AI) that uses generative models (let’s call this generative AI or “GenAI”) may not be an AI cure-all, but it can certainly enhance the human creativity aspects of your application. In this article, I’ll provide key use cases to consider when deciding between traditional models or GenAI models for your application, as well as indicators for when to augment your GenAI modelling approach with traditional models.
Traditional models use labelled data (labels provide context and meaning), to train themselves to predict outcomes and recognize patterns. For example, traditional models need a dataset of images labelled as dogs to classify images of dogs. This is called supervised learning. These are the key use cases where I’d recommend using traditional models since they’re great at using their training to make accurate predictions.
● Analysis: Identifying trends or insights (e.g., detecting faces of dogs in images)
● Classification: Grouping similar data (e.g., categorizing images of dogs for pet adoptions)
● Recommendations: Suggesting content for a specific purpose (e.g., dogs to consider adopting based on prior searches)
● Summarization: Creating an abbreviated version of data (e.g., paraphrasing dense or complexly worded statutory language for easier understanding)
● Deep retrieval: Searching for answers within data (e.g., asking a chatbot specific questions about statutory language).
● Transformation: Converting data to different formats (e.g., translating statutory language from English to Spanish)
● Augmentation: Updating data with additional points (e.g., personalizing college acceptance letters)
● Net-new creation: Creating new data using instructions (e.g., creating a college acceptance letter using bullet points)
● Anomaly detection: Identifying abnormal data (e.g., credit card fraud detection)
Even with this breadth of options, I’m not convinced that using GenAI models alone will address your application needs, so here are some considerations for supplementing GenAI models with traditional models trained on your unique data (let’s call these “custom models”).
Note: With the popularity of ChatGPT, people sometimes subconsciously use the term GenAI to refer to the OpenAI models supporting ChatGPT. If using OpenAI APIs for your application, consider prompt engineering as a first step before creating custom models.
GenAI models can unlock new possibilities for your application, but don't underestimate the power of combining them with traditional models for optimal results
● Quality control: GenAI models may sometimes create fictitious responses (hallucinations), so use custom models to filter out anything inaccurate.
● Domain adaptation: GenAI models can generate responses on any topic, but they may not always capture domain nuances. Use custom models to fine-tune GenAI model outputs to a style that makes sense for your application.
Because it may not always be obvious when it’s time to incorporate custom models, consider defining thresholds for the following metrics, to help you identify when to act.
● Accuracy: Identify the minimum level of accuracy needed for GenAI responses and incorporate custom models when responses fall below that level.
● Diversity: Identify the variety or range needed to serve the application’s purpose and leverage custom models to recommend values that enhance that range when GenAI responses cannot.
● Relevance: Identify the minimum level of relevance GenAI responses need to be applicable. This could be a measure of semantic similarity or topic coherence (i.e., something that reflects a connection between GenAI responses and its inputs).
● Appropriateness: Consider measuring sentiment analysis or emotion recognition (i.e., something that reflects the suitability of the GenAI response to its purpose).
Finally, I know I’ve spent a lot of time focusing on use cases as a guide for determining your AI approach but, don’t forget to also consider complexity (skills, ethical issues, use cases addressed, customizations to “fit” purpose) and cost (implementation and run, compute resources, OPEX or revenue impacts, the price you pay for model errors).
GenAI is not new but it’s a fast-growing branch of AI, so we’re all still learning how and where to apply it. I hope you can incorporate what I’ve shared here into checklists or KPIs that help you map out your AI transformation journey.
