Layer 5: BiLSTM | Temporal Memory Cathedral

🧠 Memory Cell Architecture

🚪 FORGET

📥 INPUT

💾 CELL

📤 OUTPUT

Hidden State (hₜ) 210-dim

Cell State (Cₜ) 210-dim

↔️ Bidirectional Streams

➡️

Forward

t=0 → t=23

⬅️

Backward

t=23 → t=0

Gate Equations

fₜ = σ(Wf·[hₜ₋₁, xₜ] + bf)
iₜ = σ(Wi·[hₜ₋₁, xₜ] + bi)
oₜ = σ(Wo·[hₜ₋₁, xₜ] + bo)

⚡ OutputMode = 'last'

Only the final hidden state (t=24) is output.
24 timesteps → 1 vector
[64×24] → [420] (210 fwd + 210 bwd)

🔗 Data Flow

Dropout1 [64×24] → BiLSTM → [420] → Dropout2 → Dense [64]

🎛️ Controls

Animation Speed 1.0×

Helix Rotation 0.5

Particle Density 100%

📊 Statistics

210

Forward

210

Backward

420

Combined

Cell & Hidden Update

Cₜ = fₜ ⊙ Cₜ₋₁ + iₜ ⊙ tanh(Wc·[hₜ₋₁,xₜ])
hₜ = oₜ ⊙ tanh(Cₜ)

BayesOpt Tuned

BiLSTM units optimized via Bayesian search in range [100, 250].
Current: 210 units per direction.

الذاكرة الزمنية: شبكة BiLSTM