AI-ordlista
Den kompletta ordlistan över AI
Per-Batch Calibration
Process of adjusting scale and zero-point quantization parameters for each data batch, optimizing representation accuracy in real-time.
Dynamic Scale
Variable quantization parameter that adapts the conversion factor between floating-point and integer numbers according to the statistical distribution of the current batch.
Dynamic Zero-Point
Offset value dynamically adjusted to ensure that zero in floating-point is represented exactly after quantization, varying per batch.
Dynamic Min-Max
Calibration algorithm that determines quantization bounds by calculating min and max values of each activation batch by batch.
Dynamic Entropy
Calibration method based on minimizing KL divergence entropy, dynamically recalculated to optimize the quantized distribution.
Dynamic KL-Divergence
Calibration technique measuring divergence between original and quantized distributions, dynamically adjusted to minimize information loss.
Dynamic Post-Training Quantization
Quantization approach applied after model training, using runtime statistics to determine optimal quantization parameters.
Adaptive Calibration
Mechanism for continuous adjustment of quantization parameters based on the evolving characteristics of input data in production.
Dynamic Quantization Factor
Variable multiplicative coefficient applied during activation conversion, dynamically optimized based on the amplitude of batch values.
Dynamic Precision
Ability of a quantization system to maintain optimal precision by continuously adapting parameters to variations in data distribution.
Dynamic Error Reconstruction
Technique evaluating and minimizing quantization error in real-time, adjusting parameters to optimize reconstruction fidelity.
Dynamic Bit Optimization
Strategy adapting the number of quantization bits per layer or per batch based on sensitivity and activation characteristics.
Dynamic Hybrid Quantization
Combination of static and dynamic approaches where certain weights are quantized statically while activations use batch-by-batch dynamic calibration.
Active-Passive Calibration
Hybrid method alternating between active calibration (frequent recalculation) and passive (caching) to optimize performance and accuracy.
Dynamic Thresholding
Technique adapting quantization saturation and truncation thresholds according to the evolving statistical characteristics of activations.