Transformer Blueprint Poster
System Model: Decoder-Encoder REF // VER-4.0.26
Transformer Blueprint
How Artificial Intelligence processes language, learns context, and predicts words.
ATTENTION ENGINE INPUT → OUTPUT LEVEL: HIGH SCHOOL CS
Architecture Schematic
Output Target Next Word Prediction Softmax → "Cool" (Probability: 92%)
Decoder Stack
Output Generation Autoregressive
Combines encoder context with previously generated words to choose the next best token.
Cross-Attention Feed Forward
Context Bridge
Encoder Stack
Contextualization Self-Attention
Builds rich understanding of inputs by calculating relations between all words simultaneously.
Multi-Head Attention FFN Layer
Input Embeddings
[0.18, -0.42, 0.91, ...]
+ Positional Code
Converts static token IDs into rich dimensional coordinate vectors.
Tokenizer
"AI" → 4212 "is" → 318
Slices text strings into manageable sub-word token pieces.
DATA PATH: BOTTOM-UP PIPELINE
1. Translating Language to Numbers
Computers can't read words directly, so Transformers translate text in two foundational steps:
A. Tokenization
Slices inputs into "tokens" (letters, word fragments, or full words) and matches them to numerical indices in a massive dictionary.
B. Vector Embeddings
Words with similar semantic meanings are grouped near each other in a multi-dimensional coordinate map (e.g., "royal" + "man" ≈ "king").
Critical Mechanism
2. Self-Attention: Context is King
Unlike older algorithms that read sequentially left-to-right, Transformers read all words simultaneously. Attention maps connections to establish the correct context.
Context Resolver: "The bank of the river"
bank → river (85%)
bank → money (10%)
The model links "bank" intensely to "river," adjusting the representation of "bank" to mean "sloped land" rather than "financial institution."
3. Crucial Architectures Explained
Positional Encoding
Since Transformers process all words at once, they require positional tags added to embeddings to know the exact placement order of each word in the text.
Multi-Head Attention
Attention is run multiple times in parallel ("heads"). Each head focuses on a different semantic angle (e.g., pronouns, active verbs, tense).
4. Autoregressive Output
Decoders output language one token at a time. After a word is predicted, it is fed back into the decoder block as part of the input sequence to help compute the word after that. This continues until a special [EOS] (End of String) token is chosen.
COMP-SCI CORNER // ARTIFICIAL INTELLIGENCE & DEEP LEARNING Series
COGNITIVE COMPUTING SYSTEMS LABORATORY © 2026
Transformer Map Graphic Organizer
Student Activity Map
Transformer Journey
WORKSHEET // COMP-SCI-101 ATTENTION & EMBEDDINGS
NAME:
DATE:
PERIOD:
1
Trace the Pipeline
Refer to your blueprint poster for reference.
Below is a simplified pipeline of an input word traveling through a Transformer. Complete the missing definitions using your own words.
Stage 1 Tokenizer
My Definition:
Stage 2 Embedding
My Definition:
Stage 3 Self-Attention
My Definition:
Stage 4 Softmax Output
My Definition:
2
Embedding Vector Space Exploration
Transformers represent word meanings as points (vectors) in space. Words with related meanings map closer together. Look at the coordinates of three simplified semantic vectors on the grid below:
Semantic Map (X, Y)
4
2
Computer [2, 4]
Robot [3, 4]
Banana [4, 1]
0 1 2 3 4
Question A: Distance Check
Calculate the distance between Computer [2, 4] and Robot [3, 4] vs Computer [2, 4] and Banana [4, 1]. Why are the coordinates for Computer and Robot closer?
Question B: High Dimensions
In real models, embeddings have hundreds or thousands of coordinates (dimensions), not just 2 (X, Y). What details or categories of word meaning might these extra dimensions measure?
3
Self-Attention Connection Challenge
Look at the target word "mouse" in two different sentences. Trace context links by completing the activities.
Sentence A
"The mouse was clicked with a finger."
1. Which word contextually defines "mouse" here?
2. Describe the meaning of "mouse" in this context:
Sentence B
"The mouse squeaked and ate the cheese."
1. Which words contextually define "mouse" here?
2. Describe the meaning of "mouse" in this context:
COMP-SCI STUDENT LAB SHEET // ATTENTION MODEL WORKROOM
COGNITIVE COMPUTING LABS © 2026
