Huawei HCIA-AI key terms

The field of building systems that perform tasks normally requiring human intelligence, such as perception, reasoning, and language use. It spans rule-based methods and learned models.

A branch of AI where programs improve at a task by learning patterns from data rather than following hand-written rules. Training adjusts model parameters to reduce error on examples.

Machine learning built on neural networks with many stacked layers, letting the model learn hierarchical feature representations directly from raw inputs such as pixels, audio, or text.

Training with labeled examples, where each input is paired with the correct answer. The model learns a mapping from input to label and is judged on unseen data.

Learning from unlabeled data to discover structure, such as groups or compressed representations. There is no target answer, so evaluation relies on the usefulness of the discovered patterns.

An agent learns by taking actions in an environment and receiving rewards or penalties, gradually finding a policy that maximizes cumulative reward over time.

A supervised task predicting a discrete category, such as spam or not spam. The model outputs a class label or a probability per class.

A supervised task predicting a continuous numeric value, such as price or temperature. Models are typically scored with error measures like mean squared error.

An unsupervised technique that groups similar samples together based on distance or similarity, without predefined labels. Common uses include customer segmentation and data exploration.

A data split where training data fits model parameters, validation data tunes settings and selects models, and test data gives a final unbiased performance estimate.

When a model memorizes training data, including noise, and performs poorly on new data. Signs include low training error but high validation error.

When a model is too simple or undertrained to capture the underlying pattern, so it performs poorly on both training and validation data. Adding capacity or better features usually helps.

Techniques that discourage overly complex models to reduce overfitting, such as adding a penalty on large weights (L1/L2) or randomly dropping units during training.

A resampling method that splits data into folds, training on some folds and validating on the held-out fold repeatedly, then averaging results for a more reliable estimate.

A mathematical measure of how far a model's predictions are from the target values. Training minimizes this quantity by adjusting parameters.

An optimization method that repeatedly moves parameters in the direction that reduces the loss, using the gradient computed from data batches.

A hyperparameter controlling the step size of each parameter update during training. Too large causes instability; too small makes training slow or stuck.

A setting chosen before training rather than learned from data, such as learning rate, batch size, or number of layers. It is tuned using validation performance.

Precision is the share of predicted positives that are correct; recall is the share of actual positives that were found. They trade off, so F1 combines them.

A table comparing predicted versus actual classes, showing true positives, true negatives, false positives, and false negatives. Precision, recall, and accuracy can be derived from it.

A model of layered units, each computing a weighted sum of inputs plus a bias, then applying an activation function. Stacking layers lets the network learn complex mappings.

A nonlinear function applied to a neuron's weighted sum, such as ReLU or sigmoid. Nonlinearity is what allows stacked layers to represent complex patterns.

The algorithm that computes gradients of the loss with respect to every weight by applying the chain rule backward through the network layers.

A network using convolution filters that slide over input data to detect local patterns, widely used for images. Weight sharing makes it efficient for grid-like data.

A network with connections that carry state across time steps, suited to sequences such as text or sensor streams. Variants like LSTM reduce vanishing-gradient problems.

A neural architecture built on self-attention that weighs relationships between all positions in a sequence in parallel. It underpins most modern large language models.

A regularization technique that randomly deactivates a fraction of units during each training step, preventing the network from relying too heavily on specific neurons.

The number of samples processed before one parameter update. Larger batches give steadier gradients but need more memory; smaller batches add noise that can aid generalization.

One complete pass through the entire training dataset. Training usually runs for many epochs, with validation checked periodically to detect overfitting.

An open-source deep learning framework from Google that represents computation as dataflow graphs and tensors, with high-level APIs such as Keras for building and training models.

An open-source deep learning framework known for dynamic computation graphs and Pythonic, debugging-friendly code. Widely used in research and increasingly in production.

Huawei's open-source deep learning framework, designed for cloud, edge, and device scenarios. It supports automatic differentiation and can compile models for Ascend hardware.

Huawei's family of AI processors and accelerators, including the Ascend 310 for inference and Ascend 910 for training, built on the Da Vinci architecture.

Huawei's portfolio of AI computing products, including modules, cards, edge devices, and servers that combine Ascend processors with software for training and inference.

Huawei's Compute Architecture for Neural Networks, a software stack between AI frameworks and Ascend hardware that provides operators, graph compilation, and runtime scheduling.

Huawei's NPU architecture for Ascend chips, featuring specialized compute units such as cube units for matrix multiplication and vector units for elementwise operations.

A very large neural network, usually transformer-based, trained on broad text to predict tokens. It can perform generation, summarization, and question answering after fine-tuning or prompting.

Further training of a pretrained model on a smaller, task-specific dataset so it adapts to a particular domain or objective while retaining general knowledge.