Understanding Model-Centric and Data-Centric AI
Model-centric AI is an approach to improving AI performance by enhancing the models themselves.
Where Should a Team Look First?
Imagine a team whose AI system is not performing as well as they want. It makes predictions, and some of them are not good enough. Where should they start looking for an improvement? There are many places they could look. One natural place is the model, which is the component that actually makes the predictions. The model-centric AI approach is built on exactly that idea: when you want an AI system to perform better, you can focus your effort on the model itself.
Keep this scenario in mind. As you read, you will trace what the model-centric approach aims at, which levers it uses, and how to decide whether a described change belongs to it. This article focuses on the model-centric approach alone so you can see its shape clearly before comparing it with other approaches.
Defining the Model-Centric Approach
Model-centric AI is an approach to improving AI performance by enhancing the models themselves.
The key word in that 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. The goal is better performance, and the model is the means of reaching that goal.
The source describes the approach as one of several ways to think about making AI systems better. It is a way of thinking, not a fixed recipe, and this article does not describe any other approach in detail.
The Three Levers
The model-centric approach works through three levers. Each lever is a different way of enhancing the model. Think of them as three separate handles a team can pull on the same object: the model itself.
The first lever is model architecture, which means designing new model architectures. The second is model size, which means increasing how large the model is. The third is training techniques, which means developing new ways of training the model.
The three levers are separate. A team can work on one of them, two of them, or all three. Recognizing which lever a change touches is the skill this article builds.
What the Approach Does and Does Not Promise
This matters when you evaluate a team's plan. A proposal that pulls a model-centric lever is a proposal to try to improve performance. Whether it actually improves performance must be judged separately, because this source gives no measured results for any lever.
Classifying a Described Change
The source gives a simple rule for classification: a change counts as model-centric when it targets the model itself, and you can identify which lever it pulls by matching the change to that lever's definition. The flowchart below turns that rule into a sequence of questions you can ask about any described change.
Classifying a team's change
A team reports three changes to its AI system. Classify each one as model-centric or not, and if it is model-centric, name the lever it pulls. Change A: the team redesigns how the model's internal structure is organized. Change B: the team cleans up mislabeled records in the data the system learns from, leaving the model unchanged. Change C: the team makes the model larger.
- 1Step 1: Ask whether the change targets the model itself
Change A redesigns the model's structure, so its target is the model. Change B leaves the model unchanged and edits data, so its target is something else. Change C makes the model larger, so its target is the model.
- 2Step 2: Match Change A to a lever definition
Redesigning the structure of the model matches designing new model architectures. Change A is model-centric and pulls the architecture lever.
- 3Step 3: Classify Change B
Change B does not target the model. It is not model-centric, even though it changes the system. Other parts of an AI system can be changed without the change being model-centric.
- 4Step 4: Match Change C to a lever definition
Making the model larger matches increasing model sizes. Change C is model-centric and pulls the size lever.
Change A is model-centric and pulls the architecture lever. Change B is not model-centric. Change C is model-centric and pulls the size lever. The team pulled two different levers in one plan, and it did so without any claim yet about measured results.
Mistakes Beginners Make When Classifying
Common mistake Treating any improvement to an AI system as model-centric
What it looks likeA team edits the records the system learns from and calls it model-centric because it improved the system.
Why it failsA change counts as model-centric only when it targets the model itself. Other parts of an AI system can be changed without making the change model-centric.
Do this insteadAsk first whether the target is the model. If the target is something else, the change is not model-centric.
Common mistake Assuming the three levers are one lever
What it looks likeA team says it pulled the model-centric lever when it only made the model larger, and treats architecture and training as the same thing.
Why it failsThe source describes the three levers as separate. Each one is a different way of enhancing the model.
Do this insteadName the specific lever: architecture, size, or training techniques. A change may pull one, two, or all three.
Common mistake Reading the approach as a guarantee of better performance
What it looks likeA plan pulls a lever, so the team concludes that performance will improve.
Why it failsThe source says the levers are methods the approach uses to try to improve performance and gives no guarantees or measured results.
Do this insteadDescribe the plan as an attempt to improve performance and judge its results separately.
Practice: Classify Three New Changes
A team proposes three changes. (1) It introduces a new training technique for its model. (2) It adds more examples to the collection of data it trains on, without touching the model. (3) It designs a new model architecture. For each change, say whether it is model-centric. If it is, name the lever it pulls.
Hints
- Start by asking whether the change targets the model itself.
- Change 2 targets the data, not the model.
- Match each model-targeting change to one of the three lever definitions: architecture, size, or training techniques.
What do you think happens?
Is a change that adds more examples to the training data, without touching the model, a model-centric change?
Reveal answer
No, because its target is the data rather than the model
A change counts as model-centric when it targets the model itself. Adding examples without touching the model targets something other than the model, so it is not model-centric.
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 is model-centric when it targets the model itself, and you identify its lever by matching the change to that lever's definition.
- The three levers are separate, and a team can pull one, two, or all three.
- The source describes these levers as methods that try to improve performance and gives no guarantees or measured results.