Few Shot Learning

I'm fascinated by the intersection of Machine Learning (ML) and Generative AI (Gen AI).

This document is written to share how we at Alvas.ai are working on combining the two technologies, to improve the output of what a Generative AI model can output.

First I’ll define a few key concepts, that are important to be aligned on. 

Understanding Machine Learning and Generative AI

• Machine Learning (ML): This technology is excellent at optimizing and identifying patterns to achieve specific goals, like increasing watch time on YouTube, maximizing return on ad spend on Facebook Ads, or improving product recommendations.

• Generative AI (Gen AI): This type of AI excels at creativity and generating content, which is crucial for marketing where creating engaging content is key.

Combining Creativity and Optimization

In marketing, we need to excel at both creating compelling content and optimizing it to maximize return on investment. Traditionally, marketers have created multiple versions of content and used split testing to find the best-performing variant.

Introducing Few Shot Learning

Few Shot Learning represents a new way to create high-performing marketing content. Here’s how it works:

• Zero Shot Learning: Typically, you might give an AI a single prompt and ask it to create something new. For example, you might prompt it like; "You're an excellent marketeer, write a promotional email for our new summer collection." The AI will produce a basic email, but the quality could be hit or miss because it only has one chance to get it right.

• Few Shot Learning: Instead, you show the AI multiple examples of good content before asking it to create something new. This way, the AI understands the style, tone, and quality you expect.

Here’s an example of promotional emails for an e-commerce business:

1. Prompt: "You're an excellent marketeer, write a promotional email for our new summer collection."

   • AI Output: "Good email we've sent to customers before."

2. Prompt: "Great, now write another promotional email about our upcoming holiday sale."

   • AI Output: "Another good email we've sent to customers."

3. Prompt: "Thanks, now write about our new arrivals in the electronics section."

   • AI Output: "Good email we've sent to customers."

4. Prompt: : "Now, please write a similar promotional email about our exclusive deals on home decor."

By providing these examples, the AI produces content that aligns with your brand and maintains a consistent tone.

Combining Few Shot Learning with Machine Learning

At Alvas, we use a combination of Machine Learning and Few Shot Learning to generate personalized emails. Here’s our process:

1. Segmenting Customers: We segment our customer base and identify which customers fit within each segment.

2. Finding Good Emails: Using Machine Learning, we find emails that have been sent to similar customers and have generated conversions. These are our "Good Emails."

3. Few Shot Learning: We then use these good emails as prompts for Few Shot Learning.

For example:

• Prompt: "Write an email to customer A about our winter sale."

  • AI Output: "Good email that made a conversion."

• Prompt: "Write an email to customer B about our new arrivals in fashion."

  • AI Output: "Good email that made a conversion."

• Prompt: "Write an email to customer C about our clearance event."

  • AI Output: "Good email that made a conversion."

• Prompt: "Thanks, now write an email to customer D about our loyalty rewards program."

Benefits

Using Few Shot Learning, we create content faster and better. Every new piece of content is based on data about what works and resonates with our customers, leading to improved results and more efficient marketing efforts.

For example, one of our clients, Shaping New Tomorrow, implemented Self Learning in their emails, and saw an uplift of more than 100% in Revenue Per Email.