최신NVIDIA Generative AI Multimodal - NCA-GENM무료샘플문제

A. Using a larger training dataset with more diverse text descriptions.

B. Decreasing the capacity of the VAE's latent space.

C. Increasing the weight of the Kullback-Leibler (KL) divergence term in the VAE's loss function.

D. Decreasing the batch size during training.

E. Applying techniques like adversarial training to encourage the VAE to generate more realistic 3D models.

A. A and C

B. The artifacts are a result of an unstable adversarial training process. Apply gradient penalty during training.

C. The artifacts are likely due to aliasing during upsampling in the generator. Use filtered upsampling or anti-aliasing techniques to mitigate this.

D. Increase the learning rate to avoid local minima.

E. The artifacts are due to mode collapse. Use more diverse training data.

A. BLEU score between the generated recipe and a reference recipe.

B. Human evaluation of the generated recipe's clarity, coherence, and accuracy.

C. Calculating the cosine similarity between the word embeddings of the generated and reference recipes.

D. Precision and recall of the ingredients mentioned in the generated recipe compared to a ground truth ingredient list.

E. Inception Score of the input image.

A.

B.

C.

D.

E.

A. Mixed-precision training allows using lower precision for forward and backward passes but keeps weights and gradients in higher precision to maintain stability.

B. Tensor Cores are only useful for inference, not training.

C. Mixed-precision training guarantees the same convergence behavior as full-precision training.

D. Tensor Cores perform specialized matrix multiplications optimized for lower-precision data types, enabling faster computation and reduced memory footprint.

E. A and B.

A. Cross-modal attention reduces the computational complexity of the model by simplifying the feature extraction process.

B. Cross-modal attention is only effective when the data from both modalities is perfectly synchronized.

C. Cross-modal attention forces the CNN and RNN to learn identical feature representations.

D. Cross-modal attention allows the model to focus on the most relevant parts of one modality based on the information from the other modality.

E. Cross-modal attention ensures that the data from both modalities is perfectly aligned in time.

A. Translating the clinical notes into multiple languages and then back-translating to the original language.

B. Directly feeding the raw clinical notes into the model without any preprocessing.

C. Applying basic text cleaning (removing punctuation, converting to lowercase) and using a standard word embedding (e.g., Word2Vec).

D. Utilizing named entity recognition (NER) to identify medical entities (diseases, medications, etc.), and employing a medical-specific language model (e.g., BioBERT) for text embeddings.

E. Performing sentiment analysis on the clinical notes.

A. Data Augmentation

B. Generative Adversarial Networks (GANs)

C. Federated Learning

D. Transfer Learning

E. Reinforcement Learning

A. KL Divergence loss between the generated sound and the input text.

B. A combination of adversarial loss (GAN) to ensure realism, a perceptual loss to match high-level features, and a semantic consistency loss to align the generated image with the input text description.

C. Cosine similarity loss between the generated image and the input 3D models.

D. Cross-entropy loss for classifying different object categories in the scene.

E. Mean Squared Error (MSE) between the generated image and a reference image.

A. Global (Soft) Attention

B. Multi-Head Attention

C. Self-Attention

D. Bahdanau Attention (Additive Attention)

E. Local (Hard) Attention

A. Replace the missing heart rate data with the mean heart rate value calculated from the available data.

B. Use a time-series imputation technique, such as Kalman filtering or recurrent neural networks, to estimate the missing heart rate values based on the available data and temporal patterns.

C. Train a separate model specifically on data without the time-series component.

D. Remove the records of patients with missing heart rate data from the dataset.

E. Assign a fixed, arbitrary value (e.g., 0) to all missing heart rate data points.

A. Removing the self-attention mechanism.

B. Reducing the number of layers in the Transformer.

C. Using only audio as input.

D. Using a smaller embedding dimension.

E. Incorporating positional encodings and increasing the attention window size.

A. NVIDIA TensorRT

B. NVIDIA DALI

C. NVIDIA NCCL

D. NVIDIA NeMo

E. NVIDIA optiX

A. Separate models for each modality trained independently, and then ensembled together at the prediction stage.

B. A model that uses a Transformer encoder for each modality, followed by a shared Transformer decoder for prediction, enabling cross-modal attention at the decoder level.

C. A model that converts all data into a single text format and uses a large language model (LLM) for prediction.

D. A model that combines a Transformer for text, an LSTM for time-series, and a CNN for images, with a late fusion strategy using a weighted averaging of predictions.

E. A simple feed forward neural network with concatenated features from all modalities.