Verfasst von Sebastian F. Genter
Z
zero-shot learning
A machine learning approach where models make predictions for classes or tasks they haven't explicitly been trained on. This capability emerges when:
- Models learn generalized representations during training
- Task descriptions are provided at inference time
- Semantic relationships between concepts are captured
Practical applications include:
- Classifying new product categories without retraining
- Adapting to novel language tasks
- Recognizing unseen object combinations
Implementation methods often involve:
- Large pretrained language models
- Semantic embedding spaces
- Auxiliary information (e.g., class attributes)
zero-shot prompting
A technique for guiding large language models where:
- No examples are provided in the prompt
- The task is described purely through instructions
- The model relies on its pretrained knowledge
Example structure:
Translate English to Russian:
"Good morning" →
Possible correct responses might include:
- "Доброе утро"
- "Утро доброе"
- "С добрым утром"
Key advantages:
- Minimal prompt engineering required
- Flexible task specification
- No need for task-specific training data
Performance factors:
- Model size and pretraining quality
- Clarity of instructions
- Task similarity to pretraining objectives
Z-score normalization
A standardization technique that transforms features to have:
- Mean of 0
- Standard deviation of 1
Calculation formula: Where:
- = original value
- = feature mean
- = feature standard deviation
Benefits for machine learning:
- Puts all features on comparable scales
- Improves optimization stability
- Reduces sensitivity to outlier values
Common use cases:
- Preparing data for distance-based algorithms
- Normalizing inputs to neural networks
- Preprocessing for principal component analysis
Implementation considerations:
- Should be calculated on training data only
- Parameters (, ) must be stored for inference
- May not be ideal for sparse data