Harsh RB

Deep Learning Blooms Taxonomy Questions

9 min read

UNIT III: Convolutional Networks

Bloom’s Taxonomy Level 5: Evaluation Questions

These questions require you to make judgments, critique methodologies, or select and defend the best approach based on criteria.

  1. Critique and Justify (Variants & Structured Outputs): A medical imaging team needs to build a model for real-time Semantic Segmentation of high-resolution surgical videos on an embedded device with limited computational power.

    • Critique the trade-offs between using Dilated Convolutions and Depthwise Separable Convolutions within the network architecture.

    • Propose and defend which variant is the superior choice for meeting both the accuracy (segmentation quality) and efficiency (real-time speed) requirements, explicitly explaining how the chosen variant exploits the Parameter Sharing motivation to achieve its efficiency gain.

  2. Evaluate and Defend (Pooling & Motivation): An engineer is experimenting with replacing Max Pooling with Average Pooling in a standard VGG-like CNN used for image classification.

    • Evaluate the expected impact of this change on the network’s overall resistance to minor shifts in the input data (Translation Invariance) and its ability to learn salient, distinct Hierarchical Features.

    • Defend which pooling technique is generally better suited for preserving the Infinitely Strong Prior of Spatial Locality across deeper layers of the network.

  3. Assess and Prioritize (Efficient Algorithms & Fundamentals): You are managing a large cloud platform that requires an algorithm for extremely fast, high-throughput convolution operations for feature extraction on large images with large kernels (e.g., ).

    • Assess the suitability and computational complexity of using the FFT-based convolution algorithm versus the Winograd algorithm for this task.

    • Defend which of the two efficient convolution algorithms should be prioritized for deployment, providing a justification based on its theoretical runtime complexity and its efficiency when dealing with large kernel sizes.

Bloom’s Taxonomy Level 4: Analysis Questions

These questions require you to break down concepts, compare and contrast methodologies, and analyze the relationships between different CNN components.

  1. Analyze and Relate (Transposed Convolutions & Structured Outputs): Analyze the exact function of a Transposed Convolution (Deconvolution) in a Semantic Segmentation network. Explain how it differs from a standard convolution with respect to its effect on the Feature Map size, and how this operation is essential for re-establishing the spatial resolution and localization accuracy lost due to Pooling layers.

  2. Compare and Contrast (Data Types & Prior): Contrast the design and application of a 1D CNN (used for analyzing a sensor sequence) with a 3D CNN (used for analyzing video volumes). Analyze how the core principle of Stationarity (as part of the Infinitely Strong Prior) is interpreted and applied differently by the kernel in these two models due to the structure of their respective input Data Types.

  3. Deconstruct and Explain (The Convolution Operation & Parameter Sharing): Break down the process of the Convolution Operation using a kernel, Stride, and Padding. Explain why and how Parameter Sharing is a direct consequence of this operation’s mechanism, and analyze the resulting computational saving compared to a fully connected layer with the same input and output dimensions.

  4. Analyze Mechanism (Depthwise Separable Convolutions): Analyze how the Depthwise Separable Convolution variant decomposes the standard convolution into two distinct steps. Explain the specific roles of the depthwise component and the pointwise component, and describe the relationship between this decomposition and the overall reduction in trainable Parameters compared to the full operation.

UNIT IV: Recurrent and Recursive Networks

Bloom’s Taxonomy Level 5: Evaluation Questions

These questions require making judgments, critiquing methodologies, and defending the optimal choice based on deep understanding.

  1. Critique and Justify (LSTM vs. GRU): A data scientist is developing a low-latency system for financial time-series forecasting, where the input sequences are often very long, making Gradient Vanishing a major concern.

    • Critique the trade-offs in computational overhead and performance between the LSTM and the GRU (Gated Recurrent Unit) architectures for this task.

    • Propose and defend the optimal choice of gated RNN, justifying your decision by explicitly relating the selected model’s gating mechanism (Input/Forget/Output or Reset/Update) to the requirements for high-speed training and effective capture of long-term dependencies.

  2. Evaluate and Defend (Architecture Selection): A team needs to build a model for complex, structured natural language inference based on abstract syntax Parse Trees rather than linear sentences.

    • Evaluate the fundamental limitations of a Bidirectional RNN or a standard Encoder-Decoder architecture for this specific task structure.

    • Defend the use of a Recursive Neural Network as the superior modeling choice, explaining how its inherent ability to process tree-structured inputs aligns with the problem’s underlying data dependency and information flow.

  3. Assess and Prioritize (Memory Networks vs. Gated RNNs): You are designing a QA system that must answer questions based on a large, external document (the “knowledge base”). The system needs to selectively retrieve relevant information from this document.

    • Assess the functional difference between an LSTM’s internal cell state memory and the Explicit Memory provided by architectures like Memory Networks or Differentiable Neural Computers.

    • Prioritize and justify which type of memory (internal Gated RNN or external Explicit Memory) is fundamentally required to successfully solve the task of referencing and reasoning over large, static external data sources.

Bloom’s Taxonomy Level 4: Analysis Questions

These questions require breaking down concepts, comparing and contrasting methodologies, and analyzing internal logic.

  1. Deconstruct and Analyze (LSTM Gating): Analyze the distinct roles of the Forget Gate and the Input Gate within the LSTM architecture. Explain how the synergistic function of these two gates specifically combats the problem of Gradient Vanishing and allows the network to maintain relevant information across large temporal distances.

  2. Compare and Contrast (RNNs vs. ESNs): Differentiate between the training procedures of a Vanilla RNN and an Echo State Network (ESN). Analyze how the ESN’s strategy of using Fixed Recurrent Weights and only training the Readout layer inherently provides an effective but specialized solution for avoiding the instability of gradient-based training across long sequences.

  3. Analyze Unfolding and Bidirectionality: Analyze the process of Unfolding the Computational Graph for a sequential task. Explain how a Bidirectional RNN utilizes this unrolled structure to integrate information from both the forward and backward contexts, and discuss a scenario (like Named Entity Recognition) where this integration is critical for accurate output.

  4. Relate Architecture to Challenge (Deep RNNs): Analyze the motivation for moving from a single-layer RNN to Deep Recurrent Networks (Stacked RNNs). Explain why the addition of Residual Connections becomes particularly critical in deep RNNs, relating this to the potential challenges of optimizing deep sequential models.

  5. Differentiate and Apply (Sequence-to-Sequence): Differentiate between the roles of the Encoder and the Decoder in a standard Sequence-to-Sequence architecture. Analyze how this architecture models the transformation for a task like Machine Translation, specifically explaining how the internal hidden state acts as the fixed-size “context vector” that bridges the variable-length input and output sequences.

UNIT V: Practical Methodology & Applications

Bloom’s Taxonomy Level 5: Evaluation Questions

These questions require making judgments, critiquing methodologies, and defending the optimal choice based on performance criteria and resource constraints.

  1. Critique and Justify (Metrics & Debugging): A self-driving car company is training an Object Detection model (e.g., YOLO) to prioritize avoiding pedestrians over incorrectly classifying stationary objects.

    • Critique the sufficiency of using only Precision and Recall as performance metrics for this safety-critical task.

    • Justify which additional metric (F1-Score, AUC-ROC, or MAP) should be prioritized for model selection, and defend a specific Debugging Strategy (e.g., Gradient Checks vs. Activation Distributions) that would be essential to verify the stability of the loss function early in the training process.

  2. Evaluate and Defend (System Design & Baselines): You are tasked with developing a system for Multi-Digit Number Recognition from street view images. Initial trials with a Logistic Regression baseline show poor performance.

    • Evaluate the necessity of immediately switching to an End-to-End Deep Learning System (like a customized ResNet) versus attempting to fine-tune a Random Forest or Linear SVM first.

    • Defend your final choice of model architecture and methodology, explicitly explaining how the system’s inherent ability to perform hierarchical feature extraction makes it fundamentally superior for solving the complex Computer Vision task compared to the chosen default baseline.

  3. Assess and Optimize (Hyperparameters & Scaling): A large language model team is struggling to train a massive Transformer (e.g., GPT) model on a cluster of GPUs. The initial Grid Search for the learning rate was too slow, and the training is failing due to memory issues.

    • Assess the utility of switching the hyperparameter optimization strategy from Grid Search to Bayesian Optimization or Random Search to maximize parameter finding efficiency.

    • Prioritize and justify a combination of Distributed Training techniques (Model Parallelism vs. Data Parallelism) required to successfully scale the training process, linking the chosen scaling method back to the fundamental memory and computational constraints of the large NLP model.

Bloom’s Taxonomy Level 4: Analysis Questions

These questions require breaking down concepts, comparing and contrasting methodologies, and analyzing internal relationships.

  1. Differentiate and Analyze (Parallelism): Differentiate between Data Parallelism and Model Parallelism in Large-Scale Deep Learning. Analyze which approach is essential for training models like BERT where the full model parameters cannot fit onto a single GPU, and which is suitable for simply speeding up the training of a smaller model on a massive dataset.

  2. Relate Metrics to Task (Classification vs. Regression): Contrast the primary goal of the AUC-ROC metric in Classification tasks with the goal of the R² metric in Regression tasks. Analyze how the selection of the most appropriate metric inherently guides the model’s optimization process (e.g., cross-entropy minimization vs. least squares minimization).

  3. Analyze Data Saturation and Gathering: Analyze the utility of Learning Curves for determining whether to Gather More Data. Explain how the shape of the learning curve (specifically, the gap between the training curve and the validation/test curve) indicates whether the model is experiencing high bias or high variance, and how this diagnosis informs the decision to collect more data.

  4. Compare and Contrast (Deep Learning Applications): Contrast the output structure and loss function requirements for an End-to-End Speech Recognition model (DeepSpeech) that uses CTC Loss versus an Image Segmentation model (U-Net). Analyze how the temporal alignment challenge in speech dictates the need for CTC, which is fundamentally absent in the pixel-wise prediction of segmentation.

  5. Deconstruct and Explain (Ranking and Default Baselines): Analyze the calculation of the Normalized Discounted Cumulative Gain (NDCG) metric in Ranking problems. Explain why a Default Baseline Model like Logistic Regression is often insufficient for achieving high NDCG scores, relating this to the metric’s inherent focus on the positional relevance of predictions.

Graph View

Backlinks