Free H13-311 practice test: 30 questions with answers

XOR outputs 1 only when the two inputs differ, so no single straight line separates the classes, and a single perceptron is limited to linearly separable problems. A learning rate problem would not block all binary functions, perceptrons accept numeric inputs including binary ones, and perceptrons typically use a step activation rather than softmax.

The sigmoid function squashes any real number into the interval between 0 and 1, so it suits binary probability outputs. Tanh produces values between -1 and 1, ReLU has no upper limit on its positive side, and a linear activation performs no squashing at all.

Large oscillations in loss after each update are a classic sign that gradient descent steps are too big, so lowering the learning rate is the standard first fix. Raising the rate would worsen the oscillation, removing hidden layers changes model capacity rather than step size, and more epochs do not address unstable updates.

Convolutional layers use learnable kernels that slide over local regions to produce feature maps, which suits edge and texture detection. Pooling layers have no learnable filters and only downsample, recurrent layers process sequences over time, and embedding layers map tokens to vectors rather than scanning images.

LSTM cells use input, forget and output gates to regulate memory, which helps preserve information over long sequences and reduces vanishing gradients. A single perceptron, a max-pooling CNN and K-means are not designed to carry memory across sequence steps, so none of them addresses this problem.

Labeled historical outcomes paired with a target to predict define supervised learning, here a classification task. Clustering ignores the existing labels, reinforcement learning needs an agent learning from rewards over time, and dimensionality reduction does not predict an outcome.

Unsupervised learning works on unlabeled data to find structure, and grouping customers into segments by behaviour is a clustering task. The other three options all learn from labeled examples to predict a known target, which makes them supervised learning.

A large gap between training and test accuracy signals overfitting, which pruning or a depth limit reduces. Underfitting would show low training accuracy and a shallower tree would make it worse, shuffled labels would destroy learning, and decision trees are not trained with gradients.

K-means is an unsupervised clustering method that requires the number of clusters k to be chosen beforehand, which fits an unlabeled segmentation task. Linear and logistic regression both need labeled targets, and a support vector machine is a supervised classifier, so none of them fits this task.

Precision is true positives divided by all predicted positives: 80 / (80 + 20) = 0.80. The 0.67 value is recall (80 / 120), while 0.40 and 0.20 do not measure precision, and 0.20 is only the false positive share, the complement of precision.

PyTorch builds its graph dynamically while the code executes, so developers can inspect intermediate values and debug with ordinary Python tools. Static graphs are a different execution style, PyTorch supports GPU acceleration, and models do not need to be converted before training.

PyTorch accumulates gradients across backward calls unless they are cleared, so optimizer.zero_grad() must run before each backward pass. Calling model.eval() changes layers such as dropout, saving a checkpoint does not alter gradients, and a larger batch size does not stop accumulation.

torch.nn.Module provides parameter registration, nesting of sub-modules, and the call mechanism that runs forward. Tensor stores data, Optimizer applies updates and Dataset organises samples, so none of them is the right base class for a model.

A typical workflow prepares data, trains and evaluates the model, deploys it, and then monitors and improves it over time. The other sequences deploy before a trained model exists, evaluate only on training data, or start from monitoring and place training after the application is live.

Ascend is Huawei's AI processor series, used in accelerators for training and inference. Kunpeng is a server CPU family, Balong handles cellular baseband processing, and HarmonyOS is an operating system, so none of them is an AI processor line.

An AI business process starts by framing the problem, the decision it supports and the success metrics, which then guide data and model choices. Deployment, architecture selection and hardware purchases all come before the requirements are known, which risks wasted effort.

Prompt engineering shapes the model's behaviour through the input text, such as instructions, context and examples, without changing its weights. Retraining, compression and data labeling are separate techniques that change the model or its data pipeline rather than the prompt.

Retrieval-augmented generation fetches up-to-date documents at query time and supplies them in the prompt, so no retraining is required. A higher temperature only increases randomness, training from scratch is costly, and removing the tokenizer would break the model rather than update its knowledge.

Lower temperature sharpens the output distribution, and hallucination is the term for confident but incorrect generated text. Tokens are often sub-word pieces rather than whole words, a larger context window still cannot use information it was never given, and prompt engineering changes only the input, not the stored weights.

Operations and post-deployment evaluation measure live results against business KPIs and feed the findings into retraining or redesign. Problem definition, data labeling and storage purchases belong to earlier stages and do not confirm business value after go-live.

Artificial intelligence is the umbrella field of building systems that perform intelligent tasks, machine learning is a subset that learns from data, and deep learning is a subset of machine learning built on multi-layer neural networks. The other options reverse the hierarchy or wrongly narrow the scope of AI and machine learning to robotics or image recognition.

Natural language processing covers analyzing and interpreting text, which is exactly what intent extraction from typed questions requires. Computer vision works on images, knowledge graphs store relationships rather than interpret sentences, and speech synthesis produces audio from text rather than understanding it.

The system excels at one task and cannot transfer that ability elsewhere, which matches narrow AI. General AI would handle any task a human can and super AI would exceed humans across all domains, so neither fits, and a foundation model describes a reusable pretrained model rather than a level of intelligence.

DeepSeek's releases demonstrated that competitive models could be developed with lower training costs than many expected, which encouraged interest in efficiency and open weights across the industry. The other statements are false, since they claim supervised learning has ended, GPUs have become unnecessary or reasoning research has stopped.

Experts disagree about whether and when general AI matching humans across all cognitive tasks will arrive, so this remains an open debate about the future. The other three statements describe established facts about data storage, neural network training and convolutional filters.

AI4Science covers applying AI to advance scientific research, and protein structure prediction is a leading example. The other areas concern user preference prediction, household automation and spoken commands, none of which matches scientific structure prediction.

SLAM, short for simultaneous localization and mapping, builds a map while tracking the robot's position using sensors. Collaborative filtering is a recommendation technique, speech synthesis produces audio, and max pooling is a CNN operation, so none of them describes mapping and localization together.

Text-to-speech synthesis turns the reply text into audio, which is the final step of the voice pipeline. Speech recognition produces the transcript, language understanding extracts intent, and wake-word detection only starts the interaction, so none of them speaks the reply.

Planning and control take the perceived scene and decide how the vehicle should move, then command steering, throttle and braking. Data labeling, model compression and recommendation ranking support perception or other services and do not choose the vehicle's motion.

Learning household routines from sensor data to automate heating for comfort and energy savings is a smart home application. Vehicles, scientific research and product recommendation involve different sensing and decision tasks, so they do not match this scenario.