CVPR-2019 Unsupervised Embedding Learning via Invariant and Spreading Instance Feature
Paper: Unsupervised Embedding Learning via Invariant and Spreading Instance Feature
InvaSpread: Instance-discrimination surrogate task + NCE loss + end-to-end learning
- Positive and negative pairs are drawn from a minibatch processed by the same encoder. Without TPU-scale batching, training stays end-to-end and needs no external data structures, which improves speed; in practice, performance improves further when more negatives are available.
- The paper explains the mathematical rationale behind the instance-wise objective—why it encourages spread-out and invariant representations.
- It stresses that learning deeper feature embeddings yields better generalization.
<HR align=left color=#987cb9 SIZE=1>