Expand description
Encoder trait for JEPA.
Implements RFC-002 (Encoder Module).
An encoder maps raw input (images, video, or already-embedded tokens)
into a Representation in embedding space.
In a JEPA training loop two encoder instances exist:
| Role | Gradients? | Weight update |
|---|---|---|
| Context encoder (θ) | Yes | Backpropagation |
| Target encoder (ξ) | No | EMA of θ (see crate::ema::Ema) |
Both share the same architecture and implement this trait. The asymmetric update (EMA on the target) is what prevents collapse.
See crate::collapse for the regularizers that complement EMA.
Traits§
- Encoder
- Trait for JEPA encoders.