Huawei HCIA-AI flashcards
Huawei HCIA-AI · Build recall with a focused review session.
32 cards. Read the term and the definition here. The deck at the top of this page is optional.
- Huawei AI full stack
- Huawei's AI stack layers from bottom to top: Ascend chips, CANN chip enablement, MindSpore framework, ModelArts platform, and industry applications - spanning device, edge, and cloud.
- Ascend 310 vs 910
- Ascend 310 targets inference at edge and data center with high energy efficiency; Ascend 910 targets data-center training with high throughput and large on-chip compute.
- Atlas product family
- Huawei's Atlas line packages Ascend silicon: Atlas 200 modules, Atlas 300 cards, Atlas 500 edge stations, Atlas 800 servers, and Atlas 900 AI clusters.
- Inclusive AI strategy
- Huawei's stated aim to make AI affordable and usable across industries such as manufacturing, healthcare, and transport, not only within large cloud providers.
- ModelArts
- Huawei Cloud's AI development platform offering data labeling, training job management, hyperparameter tuning, and model deployment as a managed service.
- Strong vs weak AI
- Weak AI is task-specific and typically supervised; strong AI is general-purpose, often self-supervised and data-efficient. Huawei's strategy addresses both.
- Supervised vs unsupervised learning
- Supervised learning trains on labeled input-output pairs to predict labels; unsupervised learning finds structure such as clusters or components in unlabeled data.
- Bias-variance tradeoff
- Total expected error decomposes into bias squared plus variance plus irreducible noise. High bias underfits; high variance overfits; model complexity trades the two.
- k-fold cross-validation
- Split data into k folds; train on k−1 and validate on the remaining fold; rotate k times and average the validation scores to estimate generalization.
- Precision and recall
- Precision = TP/(TP+FP) measures correctness of positive predictions; recall = TP/(TP+FN) measures coverage of actual positives. F1 is their harmonic mean.
- Gradient descent update
- Parameter update rule θ ← θ − η·∇L(θ), where η is the learning rate and ∇L is the loss gradient. Variants include SGD, momentum, RMSProp, and Adam.
- L1 vs L2 regularization
- L1 (Lasso) adds the sum of absolute weights and drives some weights to exactly zero for sparsity; L2 (Ridge) adds squared weights and shrinks them smoothly.
- Overfitting detection
- Overfitting appears when training loss keeps falling while validation loss rises. Remedies include more data, regularization, simpler models, and early stopping.
- R-squared
- Coefficient of determination R² = 1 − SS_res/SS_tot for regression; 1 means perfect fit and 0 means the model performs no better than predicting the mean.
- Perceptron and activation
- A perceptron computes f(Σwᵢxᵢ + b). Stacking layers without a non-linear activation f collapses mathematically into a single linear transformation.
- ReLU activation
- ReLU(x) = max(0, x). It is cheap to compute and reduces vanishing gradients compared with sigmoid or tanh, making deep networks easier to train.
- Backpropagation
- Training algorithm that applies the chain rule from output layer back to input to compute gradients of the loss with respect to every parameter.
- Convolution weight sharing
- A convolutional filter reuses the same weights across spatial positions, giving local receptive fields and far fewer parameters than a fully connected layer.
- LSTM gates
- Long short-term memory networks add input, forget, and output gates to an RNN, controlling information flow and mitigating vanishing gradients over long sequences.
- Scaled dot-product attention
- Attention(Q,K,V) = softmax(QKᵀ/√d_k)V. Scaling by √d_k keeps dot products from growing too large and stabilizes softmax gradients.
- Cross-entropy loss
- Classification loss L = −Σ y·log(ŷ), commonly paired with a softmax output layer; it penalizes confident wrong predictions heavily.
- Batch normalization
- Normalizes layer activations per mini-batch to zero mean and unit variance, then applies learnable scale and shift; stabilizes and speeds up training.
- Transfer learning
- Reuse a model pre-trained on a large dataset as a starting backbone, then fine-tune on the target task to reduce required data and compute.
- TensorFlow 2 eager mode
- TensorFlow 2.x runs operations immediately by default (eager execution); wrapping code in tf.function compiles it into a graph for performance.
- Keras in TensorFlow
- TensorFlow's high-level neural network API offering Sequential, Functional, and subclassing styles for building and training models concisely.
- MindSpore execution modes
- Huawei's open-source framework supports PyNative eager mode and Graph mode, with automatic differentiation and unified device-edge-cloud deployment.
- MindSpore train.Model
- High-level MindSpore API that binds a network, loss function, and optimizer, providing train, eval, and predict methods over a dataset.
- MindSpore automatic parallelism
- MindSpore splits computation across devices using data, model, or hybrid parallelism strategies, targeting large-scale distributed training.
- MindSpore Lite
- Lightweight MindSpore runtime for deploying trained models on edge and device hardware with reduced footprint and latency.
- CANN
- Compute Architecture for Neural Networks: Huawei's chip-enablement layer connecting frameworks such as MindSpore and TensorFlow to Ascend hardware.
- AscendCL
- Ascend computing language runtime API providing device, context, and stream management, memory allocation, and model execution for Ascend applications.
- TBE custom operators
- Tensor Boost Engine develops custom operators for Ascend, compiling them to run efficiently on the Ascend AI Core alongside built-in operator libraries.