HCIP-AI EI Developer flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

CANN

Huawei's compute architecture for neural networks; it sits between AI frameworks and Ascend chips, providing operators, graph compilation, runtime, and collective communication for training and inference.

Ascend 910

Huawei's Ascend training chip, designed for data-center deep learning training; it is paired with Ascend 310, which targets edge and inference workloads.

ModelArts

Huawei Cloud's one-stop AI development platform covering data labeling, training job management, hyperparameter tuning, model management, and deployment as online or batch services.

ExeML

ModelArts' automated machine learning feature that performs automatic data processing, model selection, and hyperparameter optimization, producing a deployable model without custom training code.

OBS

Huawei Cloud Object Storage Service; it stores datasets and model artifacts for ModelArts, and access requires IAM agency authorization and a matching region.

MindSpore Cell

The base class for neural networks in MindSpore; subclasses implement construct() to define the forward pass and are executed in PyNative or Graph mode.

Broadcasting rule

NumPy compares shapes from the right; dimensions are compatible when equal or when one is 1, and incompatible shapes raise a ValueError during the operation.

zero_grad

PyTorch optimizer method that clears accumulated gradients before the next backward pass; omitting it causes gradients to sum across iterations and destabilize training.

tf.function

TensorFlow decorator that traces a Python function into a graph for speed; a new trace is created whenever input shapes or dtypes change, which costs performance.

value_and_grad

MindSpore function that returns both the forward output and gradients of a function; it is the basis for building custom training steps.

CrossEntropyLoss

PyTorch loss combining log-softmax and negative log-likelihood, used for multi-class single-label classification where targets are class indices.

DataLoader num_workers

PyTorch setting for parallel data loading processes; setting it to zero avoids deadlocks on Windows and in notebooks while simplifying debugging.

Bias-variance diagnosis

High training and validation error indicate underfitting; low training error with high validation error indicates overfitting, guiding capacity and regularization choices.

Batch normalization

Layer that normalizes activations per channel over a batch and learns scale and shift parameters; it behaves differently in training and evaluation modes.

Adam optimizer

Optimizer that adapts a per-parameter learning rate using running estimates of first and second gradient moments, with bias correction for early steps.

CNN output size

Output width equals floor((W − K + 2P)/S) + 1, where W is input width, K kernel size, P padding, and S stride.

Self-attention formula

Attention(Q,K,V) = softmax(QKᵀ/√d_k)V; scaling by the square root of the key dimension keeps softmax gradients stable as dimension grows.

F1 score

Harmonic mean of precision and recall, computed as 2PR/(P+R); it is preferred over accuracy when classes are imbalanced.

K-fold cross-validation

Procedure that splits data into K folds, trains on K−1 and validates on the remaining fold, then averages results to reduce estimate variance.

Gradient clipping

Technique that caps gradient norm or value before the optimizer step, preventing exploding gradients in recurrent networks and deep transformers.

IoU

Intersection over Union: the area shared by two boxes divided by their combined area; it measures overlap for detection matching and evaluation.

Non-max suppression

Post-processing that sorts detection boxes by score, keeps the top box, and removes boxes whose IoU with it exceeds a threshold.

YOLO detection

Single-stage detector that predicts box offsets, objectness, and class scores in one forward pass, trading some localization precision for real-time speed.

Semantic segmentation

Task assigning a class label to every pixel; it does not separate individual instances, unlike instance segmentation which does.

Transfer learning

Reusing a pretrained backbone, replacing the final layer for the new class count, then fine-tuning with a smaller learning rate to adapt features.

mAP

Mean average precision: average precision per class computed from the precision-recall curve, then averaged over classes and often over IoU thresholds.

WordPiece tokenization

Subword scheme used by BERT that splits rare words into frequent pieces, balancing vocabulary size against out-of-vocabulary coverage.

BERT pretraining objectives

Masked language modeling predicts randomly masked tokens, and next-sentence prediction classifies whether two sentences are consecutive in the original text.

BIO tagging

Labeling scheme for named entity recognition where B marks an entity start, I marks continuation, and O marks tokens outside any entity.

Cosine similarity

Measure of angle between two vectors, computed as the dot product divided by the product of their norms; it ignores magnitude and ranges from −1 to 1.

Beam search

Decoding strategy that keeps the top-k partial sequences at each step, improving output quality over greedy decoding at higher computational cost.

Huawei Cloud NLP

EI service family offering text classification, sentiment analysis, entity extraction, and machine translation through REST APIs signed with AK/SK credentials.