An interactive workshop where participants physically embody a neural network to learn how language models work. This educational game demonstrates neural network concepts through a simplified "human neural network" that predicts words, with participants playing neurons, connections, and various network components. Includes complete workshop materials and a Python simulation to test and visualize the learning process.
The name "Words Humans Yield" is a playful nod to WHY2025 (What Hackers Yearn), the international non-profit outdoor hacker camp/conference in The Netherlands where this workshop will be presented.
This simulation is to illustrate the process of the game. It's meant for facilitators, to understand the workings of the game better.
uv run game.py
This is the output of the simulation:
Strategic weights initialized for 'cat sits on' → 'mat' and 'child plays' → 'ball'
===== EXAMPLE PREDICTION =====
Input: 'the cat sits on the ___'
Prediction: 'mat'
--- NETWORK STATUS ---
INPUT LAYER:
Active words: 'the', 'on', 'cat', 'sits'
HIDDEN LAYER 1:
Neuron HL1-1: ACTIVE (1)
Neuron HL1-2: INACTIVE (0)
Neuron HL1-3: ACTIVE (1)
Neuron HL1-4: INACTIVE (0)
Neuron HL1-5: INACTIVE (0)
HIDDEN LAYER 2:
Neuron HL2-1: ACTIVE (1)
Neuron HL2-2: ACTIVE (1)
Neuron HL2-3: ACTIVE (1)
Neuron HL2-4: INACTIVE (0)
Neuron HL2-5: INACTIVE (0)
OUTPUT LAYER:
'mat': Score 3.2
'table': Score 0.4
'chair': Score -0.3
'floor': Score 0.8
'door': Score -0.4
'garden': Score -0.8
'book': Score 0.0
'ball': Score 3.0
'cabinet': Score 0.0
'lamp': Score 0.1
PREDICTION: 'mat' (Score: 3.2)
===== SIMULATION OF GAME ROUNDS =====
===== WORDS HUMANS YIELD GAME SIMULATION =====
Training rounds:
----- ROUND 1 -----
Training sentence: 'the cat sits on the ___' (Correct: 'mat')
Step 1: Input processors activate words
Active words: the, on, cat, sits
Step 2: Hidden Layer 1 calculates activations
HL1-1: Sum = 7.0 → ACTIVE (1)
HL1-2: Sum = -0.5 → INACTIVE (0)
HL1-3: Sum = 0.5 → ACTIVE (1)
HL1-4: Sum = -0.2 → INACTIVE (0)
HL1-5: Sum = -0.3 → INACTIVE (0)
Step 3: Hidden Layer 2 calculates activations
HL2-1: Sum = 3.1 → ACTIVE (1)
HL2-2: Sum = 0.4 → ACTIVE (1)
HL2-3: Sum = 0.2 → ACTIVE (1)
HL2-4: Sum = -0.3 → INACTIVE (0)
HL2-5: Sum = -0.1 → INACTIVE (0)
Step 4: Output processors calculate scores
'mat': Score 3.2
'table': Score 0.4
'chair': Score -0.3
'floor': Score 0.8
'door': Score -0.4
'garden': Score -0.8
'book': Score 0.0
'ball': Score 3.0
'cabinet': Score 0.0
'lamp': Score 0.1
Prediction: 'mat'
Correct answer: 'mat'
Prediction is CORRECT! No weight adjustments needed.
==================================================
----- ROUND 2 -----
Training sentence: 'the child plays with the ___' (Correct: 'ball')
Step 1: Input processors activate words
Active words: the, child, plays
Step 2: Hidden Layer 1 calculates activations
HL1-1: Sum = -0.1 → INACTIVE (0)
HL1-2: Sum = 4.8 → ACTIVE (1)
HL1-3: Sum = 0.9 → ACTIVE (1)
HL1-4: Sum = -0.8 → INACTIVE (0)
HL1-5: Sum = -1.0 → INACTIVE (0)
Step 3: Hidden Layer 2 calculates activations
HL2-1: Sum = -0.3 → INACTIVE (0)
HL2-2: Sum = 3.0 → ACTIVE (1)
HL2-3: Sum = -0.2 → INACTIVE (0)
HL2-4: Sum = 0.5 → ACTIVE (1)
HL2-5: Sum = 0.1 → ACTIVE (1)
Step 4: Output processors calculate scores
'mat': Score -0.2
'table': Score -0.8
'chair': Score 0.2
'floor': Score 0.8
'door': Score -0.6
'garden': Score -0.6
'book': Score -0.0
'ball': Score 2.1
'cabinet': Score -0.4
'lamp': Score -0.2
Prediction: 'ball'
Correct answer: 'ball'
Prediction is CORRECT! No weight adjustments needed.
==================================================
----- ROUND 3 -----
Training sentence: 'the dog lies on the ___' (Correct: 'floor')
Step 1: Input processors activate words
Active words: the, on, dog, lies
Step 2: Hidden Layer 1 calculates activations
HL1-1: Sum = 1.2 → ACTIVE (1)
HL1-2: Sum = -0.8 → INACTIVE (0)
HL1-3: Sum = 1.0 → ACTIVE (1)
HL1-4: Sum = -0.1 → INACTIVE (0)
HL1-5: Sum = -0.5 → INACTIVE (0)
Step 3: Hidden Layer 2 calculates activations
HL2-1: Sum = 3.1 → ACTIVE (1)
HL2-2: Sum = 0.4 → ACTIVE (1)
HL2-3: Sum = 0.2 → ACTIVE (1)
HL2-4: Sum = -0.3 → INACTIVE (0)
HL2-5: Sum = -0.1 → INACTIVE (0)
Step 4: Output processors calculate scores
'mat': Score 3.2
'table': Score 0.4
'chair': Score -0.3
'floor': Score 0.8
'door': Score -0.4
'garden': Score -0.8
'book': Score 0.0
'ball': Score 3.0
'cabinet': Score 0.0
'lamp': Score 0.1
Prediction: 'mat'
Correct answer: 'floor'
Prediction is INCORRECT. Backpropagation trainers adjust weights:
HL2-1 → floor: 0.1 → 1.1
HL2-1 → mat: 3.0 → 2.0
HL2-2 → floor: 0.5 → 1.5
HL2-2 → mat: 0.1 → -0.9
HL2-3 → floor: 0.2 → 1.2
HL2-3 → mat: 0.1 → -0.9
After weight adjustments:
New prediction would be: 'floor'
==================================================
----- ROUND 4 -----
Training sentence: 'the plant stands on the ___' (Correct: 'table')
Step 1: Input processors activate words
Active words: the, on, plant, stands
Step 2: Hidden Layer 1 calculates activations
HL1-1: Sum = 1.9 → ACTIVE (1)
HL1-2: Sum = -0.5 → INACTIVE (0)
HL1-3: Sum = 0.4 → ACTIVE (1)
HL1-4: Sum = -0.6 → INACTIVE (0)
HL1-5: Sum = 0.2 → ACTIVE (1)
Step 3: Hidden Layer 2 calculates activations
HL2-1: Sum = 3.3 → ACTIVE (1)
HL2-2: Sum = 0.0 → ACTIVE (1)
HL2-3: Sum = 0.7 → ACTIVE (1)
HL2-4: Sum = -0.6 → INACTIVE (0)
HL2-5: Sum = -0.2 → INACTIVE (0)
Step 4: Output processors calculate scores
'mat': Score 0.2
'table': Score 0.4
'chair': Score -0.3
'floor': Score 3.8
'door': Score -0.4
'garden': Score -0.8
'book': Score 0.0
'ball': Score 3.0
'cabinet': Score 0.0
'lamp': Score 0.1
Prediction: 'floor'
Correct answer: 'table'
Prediction is INCORRECT. Backpropagation trainers adjust weights:
HL2-1 → table: 0.5 → 1.5
HL2-1 → floor: 1.1 → 0.1
HL2-2 → table: -0.3 → 0.7
HL2-2 → floor: 1.5 → 0.5
HL2-3 → table: 0.3 → 1.3
HL2-3 → floor: 1.2 → 0.2
After weight adjustments:
New prediction would be: 'table'
==================================================
===== TESTING THE NETWORK =====
Now we test the network on new sentences it hasn't seen before:
Input: 'the small cat lies on the ___'
Prediction: 'table' (Correct: 'mat') ✗ INCORRECT
Relevant weights for 'cat+sits+on→mat' pattern:
Weight 'cat'→HL1-1: 3.0
Weight HL1-1→HL2-1: 3.0
Weight HL2-1→'mat': 2.0
Input: 'the child stands on the ___'
Prediction: 'table' (Correct: 'chair') ✗ INCORRECT
Input: 'the red car stands next to the ___'
Prediction: 'ball' (Correct: 'door') ✗ INCORRECT
Input: 'the dog plays with the ___'
Prediction: 'ball' (Correct: 'ball') ✓ CORRECT