Data-Centric AI Approach
Model-centric AI is an approach to improving AI performance by enhancing the models themselves.
Where an AI System Can Be Improved
Suppose a team wants its AI system to perform better. There are many places it could look for improvement. One natural place is the model, the component that actually makes the predictions. The model-centric AI approach is built on that idea: it treats the model as the thing to enhance in order to improve performance.
This article's source material defines the model-centric approach, and this article teaches that approach on its own. Seeing its shape clearly first makes it easier to compare it with other approaches later. Keep one question in mind as you read: when someone says they want to improve an AI system, which part are they actually changing?
Definition of Model-Centric AI
Model-centric AI is an approach to improving AI performance by enhancing the models themselves.
The key word in this definition is themselves. The improvement effort is aimed at the model. Other parts of an AI system can also be changed, but in the model-centric approach the target of the work is the model's own design, scale, and training methods. Performance is the goal, and the model is the means of reaching it.
Tracing Where the Model Sits
Before looking at the levers, trace the path an AI system follows. Input data goes into the system, the model processes it, and the model's output is the prediction. The model is the component doing the predicting, which is why it is a natural place to look when performance needs to improve.
What do you think happens?
If a team changes only the model's internal design and leaves the rest of the system as it is, which stage of the path has changed?
Reveal answer
The model stage
Model-centric changes act on the model itself, the component that makes the predictions. Other parts of the system are left as they are unless someone changes them separately.
What Stays Fixed and What Changes
The diagram below shows the contrast that matters for this approach. In the model-centric approach, the work targets the model's architecture, its size, and its training techniques. The rest of the AI system is not the target of that work. The model is what changes; the other parts are what the approach leaves alone unless they are changed for a different reason.
The Three Levers
Model-centric AI works through three levers. Each lever is a different way of enhancing the model. They are separate from one another, so a team can work on one lever, two of them, or all three. The diagram shows model-centric AI as the umbrella and the three levers beneath it.
Classifying a Described Change
The skill this article builds is taking a described change and deciding two things: does it target the model itself, and if so, which lever does it pull? The flowchart below lays out the questions to ask in order.
Classifying a described change
A team changes the model's internal structure so that it processes its inputs in a different arrangement. Is this change model-centric, and if so, which lever does it pull?
- 1Check the target
The change modifies the model itself, not a different part of the AI system. That satisfies the test that a change targets the model.
- 2Match to a lever
The change redesigns how the model is structured. Designing new model architectures is the first lever, so the change pulls the model architecture lever.
- 3Confirm the other levers
The change does not say it alters the model's size or its training methods, so it does not pull those levers.
The change is model-centric and pulls the model architecture lever.
Mistakes When Classifying Changes
Common mistake Calling any improvement to an AI system model-centric.
What it looks likeA team fixes a problem in how the inputs are prepared before the model sees them.
Why it failsA model-centric change must target the model itself. Work on a different part of the system does not meet that test.
Do this insteadAsk first whether the change targets the model. If it targets another part, it is not model-centric under this definition.
Common mistake Treating model size and model architecture as the same lever.
What it looks likeA change makes the model larger without changing how it is designed.
Why it failsSize and architecture are separate levers. Scaling a model is not the same as redesigning its structure.
Do this insteadName the lever by asking what the change actually alters: the design, the scale, or the training method.
Common mistake Assuming a model-centric change is guaranteed to help.
Why it failsThe source describes the three levers as methods the approach uses to try to improve performance. It gives no guarantees or measured results.
Do this insteadDescribe model-centric work as an attempt to improve performance, not as a proven outcome.
Practice Classifying Changes
A team changes the training method used to fit its model, leaving the model's structure and size untouched. Decide whether this is model-centric, and if it is, name the lever it pulls. Then explain why a change that only reorganizes the input data would be classified differently.
Hints
- Start with the test: does the change target the model itself?
- If the change alters how the model is trained, which of the three levers does that match?
- A change to the input data targets a different part of the system than the model.
Key Takeaways
- Model-centric AI is an approach to improving AI performance by enhancing the models themselves.
- The approach works through three levers: designing new model architectures, increasing model sizes, and developing new training techniques.
- A change counts as model-centric when it targets the model itself, and matching the change to a lever's definition identifies which lever it pulls.
- The three levers are separate, so a team can work on one, two, or all three.
- The source describes these levers as methods to try to improve performance and gives no guarantees or measured results.