Real-time analysis of your interaction patterns encoded through the VAE
SESSION STATISTICS
0.0s
Duration
0
Data Points
0.00
Total Distance
0
Direction Changes
INTERACTION TIMELINE
Position Velocity
POSITIONAL DWELL TIME DISTRIBUTION
DeepMoodyOriginalPlayfulEthereal
EXTRACTED BEHAVIORAL FEATURES (INPUT x)
x₀: Velocity Integral0.000
Accumulated movement speed over session
x₁: Dwell Entropy0.000
Information content of position distribution
x₂: Rhythm Score0.000
Periodicity detection via autocorrelation
x₃: Direction Changes0.000
Normalized reversal frequency
x₄: Oscillation Freq0.000
Dominant frequency component
x₅: Asymmetry Bias0.000
Left/right preference [-1, 1]
VAE ARCHITECTURE & FORWARD PASS
INPUT x
ℝ⁶
→
ENCODER fθ
6→12
FC, ReLU
→
μ, log σ²
ℝ⁶ each
→
z = μ + σ⊙ε
ε ~ 𝒩(0,I)
→
DECODER gφ
6→12→6
FC ReLU, FC σ
→
OUTPUT x̂
ℝ⁶ ∈ [0,1]
LATENT SPACE ANALYSIS
z₀0.000
→ Petal layers (3-8)
z₁0.000
→ Petals per layer (5-13)
z₂0.000
→ Petal length (0.5-0.9)
z₃0.000
→ Petal width (0.2-0.4)
z₄0.000
→ Curl amount (0.1-0.7)
z₅0.000
→ Center size (0.08-0.23)
β-VAE LOSS DECOMPOSITION
Reconstruction Loss
ℒrecon = (1/n)Σ‖x - x̂‖²
0.0000
KL Divergence
DKL = ½Σ(μ² + σ² - log σ² - 1)
0.0000
Total Loss (β=0.5)
ℒ = ℒrecon + βeff·DKL
0.0000
ELBO (Evidence Lower Bound)
log p(x) ≥ -ℒrecon - DKL
0.0000
Training Steps
0
Effective β (warmup)
0.00
Learning Rate
0.0100
Warmup Progress
0%
DECODED FLOWER PARAMETERS
Layers0
Petals/Layer0
Petal Length0.00
Petal Width0.00
Curl0.00
Base Hue0°
BEHAVIORAL ARCHETYPE ANALYSIS
This analysis simulates four distinct behavioral archetypes interacting with the encoder,
then visualizes their latent representations. If the VAE has learned meaningful structure,
different interaction styles should cluster separately in latent space.
⚡
Nervous
Jittery, rapid oscillations, high frequency movements
🌙
Contemplative
Slow, deliberate, dwells in regions, low velocity
🌌
Exploratory
Full range coverage, many direction changes, curious
🎯
Decisive
Fast, linear sweeps, minimal hesitation, goal-oriented