NCA Computer Vision & NLP Workloads 4 — Questions and Answers
Question 1: Which technique in computer vision training uses two networks — a teacher and a student — where the student learns to mimic the teacher's output distribution rather than hard labels?
- Data augmentation
- Knowledge distillation (Correct answer)
- Batch normalization
- Dropout regularization
Correct answer: Knowledge distillation
Knowledge distillation trains a compact student model to replicate the soft probability outputs of a larger teacher model, improving efficiency while retaining accuracy.
Question 2: What is the primary advantage of using NVIDIA's Triton Inference Server for serving both CV and NLP models in production?
- It only supports TensorFlow models
- It provides concurrent model execution and dynamic batching across multiple frameworks (Correct answer)
- It replaces TensorRT for model optimization
- It requires models to be written in CUDA C++
Correct answer: It provides concurrent model execution and dynamic batching across multiple frameworks
Triton supports multiple backends (TensorRT, PyTorch, TensorFlow, ONNX) simultaneously and enables dynamic batching to maximize GPU utilization across diverse workloads.
Question 3: For fine-tuning a large language model with limited GPU memory, which parameter-efficient method trains only a small set of adapter weights inserted between transformer layers?
- Full fine-tuning
- LoRA (Low-Rank Adaptation) (Correct answer)
- Quantization-aware training
- Pipeline parallelism
Correct answer: LoRA (Low-Rank Adaptation)
LoRA freezes the original model weights and injects trainable low-rank decomposition matrices into attention layers, drastically reducing the number of trainable parameters.
Question 4: In a computer vision inference pipeline on Jetson devices, what compilation step converts a trained model to an optimized format that runs natively on the Jetson's GPU and DLA?
- ONNX export
- TensorRT engine build (Correct answer)
- CoreML conversion
- OpenVINO compilation
Correct answer: TensorRT engine build
Building a TensorRT engine on the target Jetson device produces a platform-specific optimized binary that leverages the GPU and Deep Learning Accelerator (DLA) cores.
Question 5: Which metric is most commonly used to evaluate object detection model performance, measuring precision across multiple IoU thresholds?
- Top-1 Accuracy
- Mean Average Precision (mAP) (Correct answer)
- F1 Score
- BLEU Score
Correct answer: Mean Average Precision (mAP)
mAP averages the area under the precision-recall curve across all object classes and multiple IoU thresholds, providing a comprehensive detection quality metric.
Question 6: When building a Retrieval-Augmented Generation (RAG) system, which NVIDIA library accelerates the approximate nearest neighbor search used to retrieve relevant document embeddings at scale?
- cuBLAS
- FAISS with GPU support (Correct answer)
- cuSolver
- NCCL
Correct answer: FAISS with GPU support
FAISS (Facebook AI Similarity Search) with NVIDIA GPU support dramatically accelerates vector similarity search, enabling low-latency embedding retrieval for RAG pipelines.
Question 7: In transformer-based vision models (ViT), what does the patch embedding layer convert an input image into before feeding it to the attention layers?
- A sequence of pixel-level gradients
- A sequence of fixed-size patch tokens with positional embeddings (Correct answer)
- A feature pyramid with multiple scales
- A convolutional feature map
Correct answer: A sequence of fixed-size patch tokens with positional embeddings
ViT divides the image into fixed-size patches, linearly projects each patch to an embedding vector, and adds positional embeddings to form the input token sequence.
NCA Computer Vision & NLP Workloads 4
Which technique in computer vision training uses two networks — a teacher and a student — where the student learns to mimic the teacher's output distribution rather than hard labels?