What Is an AI Model?

An artificial intelligence (AI model) is an extensive deep learning model trained on a broad dataset and capable of performing various general tasks, such as conversing in natural language or generating text and images. Deep learning models use neural network architecture to perform tasks like classification and filtering on datasets similar to what they have been trained on. AI models take deep learning one step further — the scale of their training is so vast that they can produce compelling outputs even with unpredictable inputs. Artificial intelligence models can collect and review large data points and apply their learning to achieve predefined goals using inference and transference.

Under the ISO/IEC standard, the definition of an artificial intelligence system is an “engineered system that generates outputs such as content, forecasts, recommendations or decisions for a given set of human-defined objectives.” The corresponding model is a “physical, mathematical or otherwise logical representation of a system, entity, phenomenon, process or data.” Given this broad definition, there are many types of artificial intelligence models. However, the term AI model usually refers to generative, multi-modal AI models we explain below.

Non-generative AI model

Non-generative AI models are primarily concerned with data analysis and perform classification tasks in specialized fields, such as financial fraud detection and medical imagery analysis. They were early-stage models in the AI evolution that began with machine learning and deep learning. However, non-generative AI is not the focus of this particular knowledge article.

Generative AI model

Generative AI is designed to create new content beyond what is included in its training data. Examples include OpenAI’s ChatGPT and Mistral AI, which take almost any text-based or image-based input (prompt) and generate outputs for tasks of almost any type. This includes creating recipes, event planning, fact-checking information, and composing emails. These models are also called large language models (LLMs) because of their advanced natural language processing capabilities.

Multi-modal AI model

Multi-modal AI means the AI has additional capabilities to work with both text and images. It can perform complementary tasks, such as image generation or describing images with captions. Multi-modal AI models are a type of generative AI.

Since machine learning (ML) is a subset of AI, it stands to reason that a machine learning model is a type of AI model. However, in the context of generative AI, ML models are more basic types of AI models.

Machine learning model

Machine learning models learn and use a mathematical relationship to predict an output from given inputs. For instance, given historical house sales data in a suburb, the model could predict the sale price of a given house on the market based on the mathematical function it has learned from the historical data. Common ML models include linear regression, logistic regression, and decision trees.

Deep learning model

Deep learning model is another term for ML models that are slightly more advanced—like recurrent neural networks, generative adversarial networks, and convolutional neural networks. Deep learning models better fit unstructured input data, like video sets or large amounts of text data. In contrast, basic machine learning models are often used with structured data like financial tables and GPS coordinates.

Foundation model

Researchers coined the term foundation model to describe generative AI models since these models were trained on broad data and capable of performing various general tasks. When learning how to create technical, practical applications with AI, students start by creating basic machine learning applications to understand ML model fundamentals. They can then move on to foundation models.

AI models can be used for an extraordinary range of applications. Every industry or task can be enhanced in some way by an AI application. Examples of the applications of AI models include:

• Customer service chatbots that complete self-service tasks within enterprise systems

• Predictive maintenance for industrial systems and plants

• Competitive research to determine a business strategy

• New music creation and arrangements

• Advanced video editing and effects

• Medical advice based on patient histories, symptoms, and diagnostics

Organizations build AI applications with artificial intelligence models using three primary approaches. We give an overview below.

Training an AI model from scratch

Building a foundation model from scratch is expensive and requires enormous resources. In addition to significant computational power and infrastructure, collecting, cleaning, and preparing training data can take months. Substantial time and resources must be dedicated to experimenting with different model architectures, optimization algorithms, and training strategies. After training, further investments are needed to scale the model for practical applications.

However, organizations may prefer this approach for long-term cost benefits, monetization, and leadership in the AI field.

Adapting existing foundation models

Instead of training AI models from scratch, organizations can customize and deploy existing models for their specific use cases. Fine-tuning a model involves adapting it to the organization’s dataset to perform better for a particular field or task. Retrieval augmented generation (RAG) is another approach to increase the model’s knowledge without retraining the model. With RAG, prompting the AI model first queries the organization’s knowledge base, then passes the relevant information or data and the prompt itself to the LLM to develop an output. To summarize,

• Fine-tuning modifies the model itself by continuing the training process with new data.

• RAG adds a layer of information retrieval to the model’s generative process without altering the underlying model.

RAG and fine-tuning approaches use different technologies and must be evaluated based on organizational use cases.

Prompt engineering

In this approach, the foundation model is not re-trained, fine-tuned, or deployed. Instead, organizations simply communicate prompts with it via the model’s API. For instance, consider an image editing application. To introduce AI functionality, the app adds an image generation feature. It works as follows:

• The user enters text input to generate an image.

• The app adds to the prompt and sends it to the external foundation model via APIs.

• The model returns an image to the app.

• The app displays the image to the user.

Prompt engineering is developing the best-performing prompt for the most appropriate output.

The business use case must be worthwhile for AI model implementation and customization to deliver investment returns for an organization. For instance, medium-value tasks that receive high traffic may provide value, but AI investment may not be worth it for low-value tasks. Enterprises must evaluate all proposed AI model implementations before proceeding with development.

Other challenges are given below.

Training data availability

Organizations use various third-party applications from different vendors. Often, enterprise data is locked in these applications, only available for extraction via API or data export, or sometimes not at all. Depending on the proposed AI model, there may be difficult or nuanced integration activities involved in the implementation.

Data security

Access permissions and authorization are challenging when developing internal AI applications. The training data for a given AI application must only include data that the AI application user is authorized to access. Training the AI model on a broad range of internal data increases security risks. For customer-facing applications, this may be even more complex to configure.

Automation and workflows

AI applications require automated management workflows to support the AI’s capabilities. This means building extra functionality around the AI model, either with regular programming code or no-code tools. Adapting an organization’s existing processes to meet these requirements is challenging.

Machine learning and artificial intelligence on AWS include hundreds of services to build and scale AI applications for every type of use case. You can implement basic AI functionality using APIs without building or retraining models. For example, you can use:

• Amazon Fraud Detector to detect online fraud and enhance detection models

• Amazon Monitron to detect infrastructural issues before they occur

• Amazon Rekognition to automate, streamline, and scale image recognition and video analysis

For those wishing to retrain existing AI models, Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models from leading AI companies like AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon via a single API. You also get a broad set of capabilities to build AI applications with security and privacy. Using Amazon Bedrock, you can:

• Experiment with and evaluate top FMs for your use case.

• Customize FMs privately with your data using techniques such as fine-tuning and RAG.

• Build agents that execute tasks using your enterprise systems and data sources.

Since Amazon Bedrock is serverless, you don't have to manage any infrastructure. You can securely integrate and deploy AI models using the AWS services you are already familiar with.

For customers wishing to build and train new AI models from scratch, AWS Trainium is the second-generation accelerator that AWS purpose-built for training 100B+ parameter models. Each Amazon Elastic Compute Cloud (EC2) Trn1 instance deploys up to 16 AWS Trainium accelerators to deliver a high-performance, low-cost solution for AI model training in the cloud.

Get started with AI models and developing innovative business solutions on AWS by creating a free account today.