Machine Learning and Neural Network Fundamentals
Free NVIDIA Generative AI & LLMs (NCA-GENL) practice — 6 questions on Machine Learning and Neural Network Fundamentals, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is training a deep neural network to detect fraudulent transactions. After 50 epochs, the training loss continues to decrease steadily to 0.02, but the validation loss has plateaued at 0.18 for the last 20 epochs while showing slight upward drift. The model achieves 94% accuracy on training data but only 78% on validation data. Which action would MOST effectively address this issue?
The symptoms clearly indicate overfitting: training loss decreasing while validation loss plateaus/increases, and a 16-point accuracy gap between training and validation. Dropout (randomly dropping units during training) and L2 regularization (penalizing large weights) are the standard techniques to reduce overfitting by preventing the model from memorizing training data.
An AI research team is comparing activation functions for a 12-layer convolutional neural network used for medical image classification. During training, they observe that layers 7-12 have gradients approaching zero (10^-7 to 10^-8 range), causing these layers to learn extremely slowly. Which activation function replacement would BEST resolve this vanishing gradient problem?
ReLU and LeakyReLU solve vanishing gradients because their derivatives are either 1 (for positive inputs) or a small constant (LeakyReLU for negative inputs), allowing gradients to flow backward through many layers without exponentially decreasing. ReLU has become the standard for deep networks specifically because it eliminates the vanishing gradient problem that plagues sigmoid and tanh in deep architectures.
A manufacturing company deploys a predictive maintenance model using a random forest ensemble with 200 decision trees. Each tree is trained on the full dataset of 50,000 samples with all 35 features. During inference, predictions are highly accurate but too slow for real-time monitoring (850ms per prediction vs required 100ms). The model shows training accuracy of 98.2% and test accuracy of 97.8%. Which optimization would provide the BEST speed improvement while maintaining model performance?
Random forests often exhibit diminishing returns after a certain number of trees—accuracy plateaus while inference time scales linearly with tree count. Reducing from 200 to 40-50 trees (a 75% reduction) would reduce inference time proportionally while likely maintaining near-identical accuracy since the model is already showing minimal overfitting (0.4% gap) and random forests are robust to tree count reduction after the initial plateau. This directly addresses the speed bottleneck.
A data scientist is training a binary classification neural network to identify defective products from sensor data. The dataset contains 95,000 non-defective samples and 5,000 defective samples (95:5 imbalance). After training with binary cross-entropy loss, the model achieves 95.2% accuracy but predicts 'non-defective' for almost all inputs, missing 94% of actual defects. Which combination of techniques would MOST effectively improve defect detection?
This addresses class imbalance at both data and loss levels. SMOTE creates synthetic examples of the minority class to balance training. Focal loss or class-weighted loss penalizes misclassification of the minority class more heavily (e.g., weight=19 for defective, weight=1 for non-defective), forcing the model to prioritize learning defect patterns rather than just predicting the majority class. This combination is the standard solution for severe class imbalance in neural networks.
An autonomous vehicle company is training a convolutional neural network to detect pedestrians. The model uses three convolutional layers with kernel sizes of 5×5, 3×3, and 3×3 respectively, each followed by 2×2 max pooling. Input images are 256×256×3. After the final pooling layer, the spatial dimensions are 30×30 with 128 feature maps. Which statement BEST explains the function of the max pooling layers in this architecture?
This accurately describes max pooling's dual purpose: (1) downsampling spatial dimensions (256→128→64→32 after three 2×2 poolings with stride 2), reducing computation and memory by 75% per layer, and (2) providing translation invariance—if a feature moves slightly within a pooled region, the maximum is still captured. 'Retaining most prominent features' is correct because max pooling keeps the strongest activation in each region, which typically represents detected features.
A healthcare AI team is fine-tuning a pre-trained transformer model (250M parameters) for medical report summarization. They have 12,000 labeled medical reports and limited compute budget. After 5 epochs of full model fine-tuning, validation loss is 1.84 and training loss is 0.92, but generated summaries are incoherent. Training takes 40 hours per epoch on their hardware. Which approach would MOST efficiently improve results within their constraints?
LoRA and adapter-based methods are specifically designed for this scenario: limited data and compute for large pre-trained models. LoRA trains low-rank matrices (typically reducing trainable parameters from 250M to 1-2M) that are applied to attention layers, achieving 95-98% of full fine-tuning performance while being 3-10x faster and using far less memory. This addresses the compute constraint and reduces overfitting (the 0.92 training loss vs 1.84 validation loss gap) by drastically limiting trainable parameters. This is the current standard approach (2024-2026) for fine-tuning large language models with limited resources.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$54.99 $17.99 with code FREETEST33 — valid through August 23.
Get my $17.99 deal →