Artificial Intelligence (AI) and machine learning are terms that are often used interchangeably. While AI is about creating machines capable of mimicking human-like decision-making, machine learning is the method by which this is achieved. In this article, we will explore the four distinct branches of machine learning: Traditional ML, Deep Learning, Fine-tuned ML, and Prompt-engineering.
1. Traditional ML
Also known as the backbone of machine learning, Traditional ML models are quick to build and deploy, backed by well-established scientific methods. These models have been renamed to fall under the ‘data scientist’ job role over the years. Tools like Scikit-Learn and methods such as logistic regression, linear regression, clustering, and TF-IDF are commonly used. Applications include supply chain forecasting, recommender systems, and more.
2. Deep Learning ML
The last decade has seen a surge in the use of deep learning, characterized by artificial neural networks trained on large datasets. This field is more of an art than a science, often requiring trial and error to find the right hyperparameters. These models demand significant computational power, usually GPUs, and substantial financial investment. Tools like TensorFlow and PyTorch, along with cloud solutions from AWS and Azure, are commonly used. However, these models often specialize in narrow tasks, leading researchers to seek more generalizable solutions.
3. Fine-tuned ML
Recent advancements in transformer models have offered a solution to the limitations of deep learning. Utilizing the attention mechanism, these models are more robust and less brittle than their predecessors. Fine-tuned ML refers to pre-trained deep learning models that can be further trained on specific datasets. Notably, Meta’s LLama2 is available for public use, while OpenAI has kept the underlying model of their ChatGPT platform private. Developers have had to rely on the technique of prompt engineering to work with top models like GPT-4.
Initially considered a subset of Fine-tuned ML, prompt-engineering has rapidly evolved into a category of its own. This approach involves controlling a machine learning model’s behavior through contextual cues. Advanced techniques go beyond simple instructions and use vector databases like Pinecone to store an “AI mental model” of an organization’s data. Frameworks like Langchain offer various techniques for this purpose. The field is even advancing towards chaining different AI agents together for tasks, exemplified by Azure’s AutoGen.
The machine learning ecosystem is evolving at a rapid pace, with both big and small players contributing to innovation. Companies like OpenAI and Microsoft are pushing for an approach where businesses can plug into pre-built platforms, eliminating the need for in-house model training. However, the demand for all types of machine learning is on the rise. Organizations will need to evaluate their goals, available resources, and desired differentiators when deciding which type of machine learning to adopt.