Vanishing & Exploding Gradients: Multiplying Step Factors Back Through Time
This lesson is part of the course
In this lesson
You compute the factor u·(1 − h²) that a gradient is multiplied by at every step back, chain ten steps with PRODUCT for a small and a large recurrent weight and chart how the signal vanishes or explodes.
Buy the course once for $9.99 to unlock every lesson and lab, the final exam and the certificate – lifetime access, no subscription.
Lessons in this course
- Sequences as Rows: Time Steps, Scaled Inputs, Words as Numbers & Why Order Matters Locked
- The Simple Recurrent Cell: Weighted Input, Recurrent Weight, Bias & tanh, Step by Step Locked
- Unrolling Through Time & Many-to-One: The Final Hidden State as a Summary and a Toy Sentiment Score Locked
- Vanishing & Exploding Gradients: Multiplying Step Factors Back Through Time Locked
- LSTM Gates: Forget, Input & Output Gates, the Candidate Value and the Cell State over Six Steps Locked
- GRU Cells: Update & Reset Gates, the New Hidden State and a Comparison with the LSTM Locked
- Parameter Counts & Cost: Weights per Gate, Simple vs GRU vs LSTM and a Size Table Locked
- Capstone: Next-Value Prediction with Given Trained Weights, MAE vs a Naive Forecast & a Chart Locked