one fable-low hijack cell
pair attention / gradient
envelope en_arrow
ordering AB
effort low
English with arrow '→' — '{X} → {Y}'
verdictpicked gradient (second-offered)
· classifierB
· tiernone
response (200 chars)
Could you clarify what you're after? A few common interpretations:
**1. Gradients *through* attention (backprop)**
Attention is fully differentiable, so gradients flow through the softmax:
- Output:
neighbors