Skip to main content
Glite Dictionary
🧀 Partly hard 🤖 Moderately difficult
semihard

semihard

8.3
A training example that is not easy but not the hardest to tell from a correct one
  • adjective
  • /ˈsɛmiˌhɑrd/
  • Specialized
translation icon : semiduro
  • During training, the model sampled semihard negatives to help the embedding separate similar classes.

Examples

  • Researchers found that mining semihard examples improved convergence without introducing noisy false negatives in batches.

  • The triplet loss benefits when each batch contains at least one semihard pair alongside easy and hard pairs.

  • The semihard examples greatly improved the model's accuracy.

  • We need to include more semihard examples in our training set.

  • During training, the model sampled semihard negatives, making it easier to distinguish between positive and negative examples.

  • Researchers found that using semihard examples during model training improved accuracy without adding unnecessary noise.

  • Including semihard examples in each training batch significantly enhanced the model's ability to differentiate classes.

Antonyms

easy
∞ 1.0

Not hard to do, understand, or achieve

How Distinguishable

distinct
168 3.5
semihard
  • Specialized
8.3

Surface Forms

semihard positive
semiharder comparative
semihardest superlative

Morphology

semihard = hard (semi-transparent) = semi + hard

The compositional meaning is clear, but the word is used as specialized machine-learning jargon rather than ordinary language.

Etymology

The word semihard in machine learning uses the same parts semi- 'partly' and hard, and it describes an example that is not easy to separate from a correct one but not impossible either. So a semihard example is moderately difficult and helps the model learn to tell similar cases apart.