NCA-GENM 無料問題集「NVIDIA Generative AI Multimodal」

（A）L2 regularization

（B）Early Stopping

（C）L1 regularization

（D）Batch Normalization

（E）Dropout

（A）F 1-Score

（B）Pixel Accuracy.

（C）ROUGE (Recall-Oriented Understudy for Gisting Evaluation).

（D）BLEU (Bilingual Evaluation Understudy).

（E）Inception Score.

（A）Randomly shuffle data during training.

（B）Ensure the training dataset contains diverse representation of all demographic groups and viewpoints.

（C）Use a pre-trained language model that has been debiased.

（D）Increase the number of training epochs.

（E）Evaluate the model's summaries on different demographic groups to identify and mitigate any disparities in performance.

（A）Implement a perceptual loss function that compares high-level features of generated and real images.

（B）Increase the batch size used for training.

（C）Use a simpler generator architecture.

（D）Train the model for fewer epochs.

（E）Decrease the learning rate of the discriminator.

（A）Use reinforcement learning to train the model to maximize a reward function that reflects the desired dialogue characteristics (e.g., coherence, emotional appropriateness).

（B）Train each mode separately to achieve the best result and them merge at the end.

（C）All of the above. Except D

（D）Implement a hierarchical dialogue generation architecture that first plans the overall dialogue structure and then generates individual utterances.

（E）Incorporate attention mechanisms that allow the model to selectively focus on the most relevant aspects of the game state and character personality profile.

（A）Using just the Python backend with the models on CPU.

（B）A Python backend where both models are loaded into memory and inference is performed sequentially.

（C）A single model repository containing both models as TorchScript, served by a single Triton instance using the PyTorch backend.

（D）A single model repository with two model instances (CLIP as ONNX, StyleGAN as TensorRT) served by a single Triton instance, leveraging concurrent execution.

（E）Two separate model repositories, one for CLIP (as ONNX) and one for StyleGAN (as TensorRT), served by two Triton instances on different GPUs.

（A）Cross-entropy loss between the generated image and the text description.

（B）Increasing the batch size during training to improve gradient estimation.

（C）L1 loss between the generated image and the target image.

（D）Mean Squared Error (MSE) loss between the generated image and a downscaled version of the target image.

（E）Perceptual loss, which compares the feature representations of the generated and target images in a pre-trained CNN.

（A）Fine-tuning a pre-trained transformer model (e.g., BERT, RoBERTa) on a dataset of sarcastic and non-sarcastic reviews.

（B）Using a rule-based system that identifies keywords associated with positive or negative sentiment.

（C）Bag-of-words model with TF-IDF weighting.

（D）Sentiment lexicon-based approach.

（E）Calculating the average word embedding for each review.

（A）Weight clustering to reduce model size, pruning low-importance connections, and using a larger learning rate during fine-tuning.

（B）FP16 mixed-precision training, layer fusion to combine multiple operations into one, and increasing the batch size to improve GPU utilization.

（C）Converting all layers to FP16, removing skip connections from the IJ-Net architecture, and using a smaller input image resolution.

（D）Quantization Aware Training (QAT) to INT8, Knowledge Distillation from the FP32 model to a smaller student model, and channel pruning to reduce the number of filters.

（E）Applying standard post-training quantization to INT8, replacing convolutional layers with fully connected layers, and using a smaller batch size.

（A）Using TF-IDF for text, pixel values for images, and one-hot encoding for user profile information.

（B）Using a combination of TF-IDF for text, pixel values for images, and numerical features for user profile information. Then apply PCA for dimensionality reduction.

（C）Using character-level n-grams for text, edge detection for images, and boole an features for user profile information.

（D）Using bag-of-words for text, histogram of oriented gradients (HOG) for images, and simple numerical features (e.g., number of followers) for user profiles.

（E）Using word embeddings (e.g., Word2Vec, GloVe) for text, pre-trained CNN features (e.g., from ResNet, Inception) for images, and embedding user profiles using a graph embedding technique.

（A）The text encoding is overly complex and should be simplified to reduce computational overhead.

（B）The text encoding only supports ASCII characters and does not account for word embeddings or sequence length variations. Use a tokenizer like BERT or SentencePiece to generate embeddings and pad sequences to a fixed length

（C）The text encoding is suitable for small datasets but will not scale to larger datasets.

（D）It adequately addresses the complexities inherent in natural language, making it suitable for a variety of multimodal models.

（E）The text encoding is efficient but incompatible with common deep learning architectures.

（A）Combine B and C.

（B）Increase the size of the training dataset for the object detection model.

（C）Incorporate visual attention mechanisms that allow the description generation model to focus on the most salient regions of the image.

（D）Use a more powerful transformer-based model (e.g., GPT-3) to generate the image descriptions from the extracted object, attribute, and relationship information.

（E）Manually rewrite a subset of descriptions to be more in line with the requirements.

（A）Decrease the resolution of the video frames.

（B）Using only pretrained model weights.

（C）Implement a contrastive learning objective that encourages similar embeddings for corresponding video frames and text descriptions.

（D）Use a larger vocabulary for the text encoder.

（E）Increase the frame rate of the training videos.