Ever wonder what we actually do after the AI “reads” all those pictures? In the last part, we talked about Stateless and Stateful thinking. Now, we need to talk about the Intention; the reason why we are recognizing those pictures in the first place.
Even if two people look at the same video, they might have different goals. One person wants to know what is happening right now, while the other wants to know what will happen next. In the AI world, we split these intentions into two main paths: Forecasting and Classification.
1. Forecasting: The Fortune Teller
Forecasting is all about the future. The intent here is to take the “Time Series” backstory we discussed in Part I and extend it forward.
- Goal: “Based on the last 10 days of stock prices, what will the price be tomorrow?”
- Focus: Trends, seasonality, and cycles.
2. Classification: The Observer
Classification is all about the label. The intent here is to look at a sequence of events and put them into a category.
- Goal: “Based on this 5-second video, is the person walking, running, or falling?”
- Focus: Patterns, shapes, and specific “triggers” in the data.
As we have seen, the “Memory” of a model changes how it sees the world. But here is the secret to model selection: Forecasting requires a “story,” while Classification only needs “clues."
The Rule for Forecasting: You must use a Stateful process. Because you are predicting a future that hasn’t happened yet, the AI absolutely needs to remember the past to find the trend. Without memory, there is no prediction.
The Rule for Classification: This is more flexible. You can use both Stateful or Stateless processes depending on what you are looking for:
- Stateless Classification (The “Snapshot”): Use this when the answer is inside a single moment. Example: “Is there a fire in this video frame?" You don’t need to know what happened a minute ago to see a flame right now.
- Stateful Classification (The “Movement”): Use this when the answer is hidden in the motion. Example: “Is this person dancing?" If you only look at one “snapshot,” it might look like they are just standing on one leg. You need the memory of the previous frames to recognize the “dance.”
Whether you choose Forecasting or Classification, you have to decide on your Look-back Window. This is the amount of history the AI needs to make a decision.
- Small Window: Good for quick reactions (e.g., detecting a “trip and fall” in a video).
- Large Window: Good for slow trends (e.g., predicting stock market changes over a month).
If your window is too small, the AI misses the context. If it’s too large, the AI gets “distracted” by old, irrelevant information and slows down.
Summary Checklist for Model Selection
Before you choose, ask yourself these three questions:
- Is my goal Forecasting or Classification? (Future vs. Label)
- Does the “Now” need the “Before”? (Stateless vs. Stateful)
- How much history matters? (The Look-back Window)
graph TD
Start[AI Model Selection Blueprint] --> Q1{Question 1:
Goal?}
Q1 -->|Predict Future Value| Goal1[Forecasting]
Q1 -->|Assign Label/Category| Goal2[Classification]
Goal1 --> Q2
Goal2 --> Q2
Q2{Question 2:
Dependency?}
Q2 -->|Now is independent| S1[Stateless / Non-Memory]
Q2 -->|Now needs Before| S2[Stateful / Memory]
S1 --> Final[Final Architecture Defined]
S2 --> Q3{Question 3:
Look-back Window?}
Q3 -->|Short / Immediate| W1[Low Memory / CNN-1D]
Q3 -->|Long / Historic| W2[High Memory / LSTM / Transformer]
W1 --> Final
W2 --> Final
%% Styling for "Strong" finish
style Start fill:#1a1a1a,stroke:#facc15,stroke-width:2px,color:#fff
style Final fill:#b45309,stroke:#fff,stroke-width:2px,color:#fff
If you can answer these, you have the blueprint for your AI architecture.