mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-05-03 08:08:51 -05:00
2390c3ab31
Unifies Quarto metadata into shared base/format/volume fragments while carrying through chapter path, asset, and tooling updates to keep the repository consistent and easier to maintain.
1299 lines
46 KiB
JSON
1299 lines
46 KiB
JSON
{
|
||
"metadata": {
|
||
"generated_date": "2025-01-11T12:00:00",
|
||
"format_version": "2.0.0",
|
||
"source": "actual_chapter_content",
|
||
"total_chapters": 20
|
||
},
|
||
"chapters": {
|
||
"introduction": {
|
||
"concepts": [
|
||
"Machine Learning Systems Engineering",
|
||
"AI vs ML vs Deep Learning distinctions",
|
||
"Supervised, Unsupervised, Reinforcement Learning",
|
||
"ML Systems vs traditional software systems",
|
||
"Data-driven vs rule-based approaches",
|
||
"ML development lifecycle",
|
||
"AI pervasiveness in modern applications",
|
||
"Symbolic AI and rule-based systems",
|
||
"Expert Systems and knowledge representation",
|
||
"Statistical Learning paradigm shift",
|
||
"Shallow Learning (SVM, decision trees)",
|
||
"Deep Learning revolution",
|
||
"Neural network fundamentals",
|
||
"Perceptron history",
|
||
"Backpropagation introduction",
|
||
"AlexNet and ImageNet breakthrough",
|
||
"Foundation models concept",
|
||
"Model-Data-Infrastructure triangle",
|
||
"ML system lifecycle phases",
|
||
"Production deployment challenges"
|
||
],
|
||
"terminology": [
|
||
"Model", "Dataset", "Features", "Labels",
|
||
"Training", "Validation", "Testing",
|
||
"Overfitting", "Underfitting", "Generalization",
|
||
"Bias", "Variance", "Bias-variance tradeoff",
|
||
"Loss function", "Optimization", "Hyperparameters",
|
||
"Parameters", "Inference", "Serving",
|
||
"Perceptron", "Activation function",
|
||
"Gradient descent", "Epochs", "Iterations",
|
||
"Batch size", "Learning rate"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Accuracy",
|
||
"expression": "Accuracy = Correct Predictions / Total Predictions",
|
||
"variables": "Correct Predictions, Total Predictions",
|
||
"application": "Model evaluation"
|
||
},
|
||
{
|
||
"name": "Error Rate",
|
||
"expression": "Error = 1 - Accuracy",
|
||
"variables": "Accuracy",
|
||
"application": "Model evaluation"
|
||
},
|
||
{
|
||
"name": "Train/Test Split",
|
||
"expression": "70% train, 15% validation, 15% test (typical)",
|
||
"variables": "Dataset size",
|
||
"application": "Data partitioning"
|
||
}
|
||
]
|
||
},
|
||
"ml_systems": {
|
||
"concepts": [
|
||
"Cloud ML architecture",
|
||
"Edge ML deployment",
|
||
"Mobile ML constraints",
|
||
"TinyML and embedded systems",
|
||
"Deployment spectrum tradeoffs",
|
||
"Data center infrastructure",
|
||
"Power consumption constraints",
|
||
"Memory hierarchy in ML systems",
|
||
"Latency vs Throughput optimization",
|
||
"Network connectivity requirements",
|
||
"Resource-constrained deployment",
|
||
"System reliability and monitoring",
|
||
"MLOps practices",
|
||
"Model versioning and governance",
|
||
"A/B testing for ML",
|
||
"Shadow mode deployment",
|
||
"Production ML challenges",
|
||
"ML technical debt"
|
||
],
|
||
"terminology": [
|
||
"Inference", "Serving", "Deployment",
|
||
"Data pipeline", "Feature pipeline",
|
||
"Model registry", "Model versioning",
|
||
"A/B testing", "Shadow mode",
|
||
"SLA (Service Level Agreement)",
|
||
"MLOps", "Model governance",
|
||
"Edge device", "Embedded system",
|
||
"Data center", "Hyperscale",
|
||
"PUE (Power Usage Effectiveness)",
|
||
"Thermal limits", "Battery constraints"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Latency Percentiles",
|
||
"expression": "P50, P95, P99 calculations",
|
||
"variables": "Response time distribution",
|
||
"application": "Performance measurement"
|
||
},
|
||
{
|
||
"name": "Throughput",
|
||
"expression": "Throughput = Requests / Second",
|
||
"variables": "Number of requests, Time",
|
||
"application": "System performance"
|
||
},
|
||
{
|
||
"name": "System Availability",
|
||
"expression": "Availability = Uptime / Total Time × 100%",
|
||
"variables": "Uptime, Total Time",
|
||
"application": "Reliability measurement"
|
||
},
|
||
{
|
||
"name": "Power Efficiency",
|
||
"expression": "Performance / Watt",
|
||
"variables": "Computational performance, Power consumption",
|
||
"application": "Energy efficiency"
|
||
}
|
||
]
|
||
},
|
||
"nn_computation": {
|
||
"concepts": [
|
||
"Neural network architecture",
|
||
"Biological vs artificial neurons",
|
||
"Perceptron model",
|
||
"Multi-layer perceptrons (MLP)",
|
||
"Forward propagation",
|
||
"Backward propagation algorithm",
|
||
"Activation functions (ReLU, Sigmoid, Tanh)",
|
||
"Gradient descent optimization",
|
||
"Stochastic gradient descent (SGD)",
|
||
"Deep vs shallow networks",
|
||
"Representation learning",
|
||
"Feature hierarchies",
|
||
"Vanishing/Exploding gradients",
|
||
"Weight initialization",
|
||
"Loss function types",
|
||
"Training vs inference phases",
|
||
"Pre-processing and post-processing",
|
||
"Network topology design"
|
||
],
|
||
"terminology": [
|
||
"Neurons", "Layers", "Weights", "Biases",
|
||
"Forward pass", "Backward pass",
|
||
"Learning rate", "Batch size", "Epochs",
|
||
"Gradients", "Chain rule",
|
||
"Hidden layers", "Input layer", "Output layer",
|
||
"Activation", "Non-linearity",
|
||
"Loss", "Cost function",
|
||
"Optimizer", "Momentum",
|
||
"Mini-batch", "Iteration"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Neuron Output",
|
||
"expression": "y = f(Σ(wi×xi) + b)",
|
||
"variables": "weights (w), inputs (x), bias (b), activation (f)",
|
||
"application": "Neural computation"
|
||
},
|
||
{
|
||
"name": "ReLU Activation",
|
||
"expression": "f(x) = max(0, x)",
|
||
"variables": "input x",
|
||
"application": "Non-linear activation"
|
||
},
|
||
{
|
||
"name": "Sigmoid Activation",
|
||
"expression": "σ(x) = 1/(1+e^(-x))",
|
||
"variables": "input x",
|
||
"application": "Probability output"
|
||
},
|
||
{
|
||
"name": "Parameter Count",
|
||
"expression": "params = (input_size + 1) × output_size",
|
||
"variables": "input_size, output_size",
|
||
"application": "Memory estimation"
|
||
},
|
||
{
|
||
"name": "Memory for Activations",
|
||
"expression": "batch_size × layer_size × precision_bytes",
|
||
"variables": "batch_size, layer_size, precision",
|
||
"application": "Memory planning"
|
||
}
|
||
]
|
||
},
|
||
"data_engineering": {
|
||
"concepts": [
|
||
"Data collection strategies",
|
||
"Data quality and validation",
|
||
"Data preprocessing pipelines",
|
||
"Feature engineering techniques",
|
||
"Data augmentation methods",
|
||
"Data versioning and lineage",
|
||
"Distributed data processing",
|
||
"Streaming vs batch processing",
|
||
"ETL pipeline design",
|
||
"Data lake vs data warehouse",
|
||
"Schema validation",
|
||
"Data drift detection",
|
||
"Feature stores",
|
||
"Data provenance tracking",
|
||
"Sampling strategies",
|
||
"Class imbalance handling"
|
||
],
|
||
"terminology": [
|
||
"ETL (Extract, Transform, Load)",
|
||
"Data lake", "Data warehouse",
|
||
"Schema validation", "Data drift",
|
||
"Feature store", "Feature catalog",
|
||
"Data provenance", "Lineage tracking",
|
||
"Sampling", "Stratification",
|
||
"Data augmentation", "Synthetic data",
|
||
"Batch processing", "Stream processing",
|
||
"Data pipeline", "Data quality"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Dataset Size",
|
||
"expression": "Samples × Features × Bytes_per_feature",
|
||
"variables": "Number of samples, features, data type size",
|
||
"application": "Storage requirements"
|
||
},
|
||
{
|
||
"name": "Sampling Rate",
|
||
"expression": "Sample_size / Population_size",
|
||
"variables": "Sample size, Total population",
|
||
"application": "Data sampling"
|
||
},
|
||
{
|
||
"name": "Class Imbalance Ratio",
|
||
"expression": "Minority_class / Majority_class",
|
||
"variables": "Class counts",
|
||
"application": "Balance assessment"
|
||
},
|
||
{
|
||
"name": "Data Throughput",
|
||
"expression": "GB/s processing rate",
|
||
"variables": "Data volume, Processing time",
|
||
"application": "Pipeline performance"
|
||
}
|
||
]
|
||
},
|
||
"nn_architectures": {
|
||
"concepts": [
|
||
"Convolutional Neural Networks (CNNs)",
|
||
"Recurrent Neural Networks (RNNs)",
|
||
"Long Short-Term Memory (LSTM)",
|
||
"Gated Recurrent Units (GRU)",
|
||
"Transformer architecture",
|
||
"Self-attention mechanism",
|
||
"Multi-head attention",
|
||
"Encoder-decoder architectures",
|
||
"Residual connections (ResNet)",
|
||
"Skip connections",
|
||
"Batch normalization",
|
||
"Dropout regularization",
|
||
"Transfer learning",
|
||
"Fine-tuning strategies",
|
||
"Neural Architecture Search (NAS)",
|
||
"Vision transformers (ViT)"
|
||
],
|
||
"terminology": [
|
||
"Convolution", "Pooling", "Stride", "Padding",
|
||
"Receptive field", "Feature maps",
|
||
"Sequence modeling", "Hidden state",
|
||
"Self-attention", "Cross-attention",
|
||
"Query", "Key", "Value",
|
||
"Positional encoding", "Embeddings",
|
||
"Skip connections", "Bottleneck layers",
|
||
"Kernel", "Filter", "Channel"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "CNN Parameter Count",
|
||
"expression": "(kernel_h × kernel_w × in_channels + 1) × out_channels",
|
||
"variables": "Kernel dimensions, channel counts",
|
||
"application": "CNN memory estimation"
|
||
},
|
||
{
|
||
"name": "CNN Output Size",
|
||
"expression": "(input - kernel + 2×padding) / stride + 1",
|
||
"variables": "Input size, kernel, padding, stride",
|
||
"application": "Layer output dimensions"
|
||
},
|
||
{
|
||
"name": "Transformer Complexity",
|
||
"expression": "O(n²×d)",
|
||
"variables": "sequence_length (n), dimension (d)",
|
||
"application": "Computational complexity"
|
||
},
|
||
{
|
||
"name": "Attention Scores",
|
||
"expression": "softmax(QK^T/√d)",
|
||
"variables": "Query (Q), Key (K), dimension (d)",
|
||
"application": "Attention computation"
|
||
},
|
||
{
|
||
"name": "FLOPs per Layer",
|
||
"expression": "2 × input_size × output_size × batch_size",
|
||
"variables": "Layer dimensions, batch size",
|
||
"application": "Compute requirements"
|
||
}
|
||
]
|
||
},
|
||
"frameworks": {
|
||
"concepts": [
|
||
"Deep learning frameworks (TensorFlow, PyTorch, JAX)",
|
||
"Computational graphs",
|
||
"Static vs dynamic graphs",
|
||
"Automatic differentiation",
|
||
"Distributed training frameworks",
|
||
"Model serialization formats",
|
||
"ONNX interoperability",
|
||
"Framework-specific optimizations",
|
||
"Custom operators and kernels",
|
||
"JIT compilation",
|
||
"XLA optimization",
|
||
"Mixed precision training",
|
||
"Gradient checkpointing",
|
||
"CUDA programming",
|
||
"Kernel fusion"
|
||
],
|
||
"terminology": [
|
||
"Tensors", "Operations", "Graphs",
|
||
"Eager execution", "Graph execution",
|
||
"JIT compilation", "XLA",
|
||
"Checkpointing", "Serialization",
|
||
"CUDA", "cuDNN", "Kernel fusion",
|
||
"Mixed precision", "FP16", "FP32",
|
||
"Autograd", "Backward pass",
|
||
"SavedModel", "ONNX"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Memory Usage",
|
||
"expression": "Model_size + Batch_size × Activation_memory",
|
||
"variables": "Model parameters, batch size, activations",
|
||
"application": "GPU memory planning"
|
||
},
|
||
{
|
||
"name": "Gradient Memory",
|
||
"expression": "2 × Model_size (for Adam optimizer)",
|
||
"variables": "Model parameter count",
|
||
"application": "Optimizer memory"
|
||
},
|
||
{
|
||
"name": "Communication Overhead",
|
||
"expression": "2 × Model_size × (n-1)/n",
|
||
"variables": "Model size, number of devices (n)",
|
||
"application": "Distributed training"
|
||
}
|
||
]
|
||
},
|
||
"training": {
|
||
"concepts": [
|
||
"Training loops and procedures",
|
||
"Optimization algorithms (SGD, Adam, RMSprop)",
|
||
"Learning rate scheduling",
|
||
"Batch normalization",
|
||
"Dropout regularization",
|
||
"Early stopping criteria",
|
||
"Distributed training strategies",
|
||
"Data parallelism",
|
||
"Model parallelism",
|
||
"Pipeline parallelism",
|
||
"Gradient accumulation",
|
||
"Gradient clipping",
|
||
"Weight decay",
|
||
"Convergence monitoring",
|
||
"Training system evolution"
|
||
],
|
||
"terminology": [
|
||
"Mini-batch", "Epoch", "Iteration",
|
||
"Momentum", "Adaptive learning rates",
|
||
"Weight decay", "L1/L2 regularization",
|
||
"Gradient clipping", "Gradient scaling",
|
||
"Data parallelism", "Model parallelism",
|
||
"Synchronous training", "Asynchronous training",
|
||
"All-reduce", "Parameter server",
|
||
"Learning rate", "Batch size"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "SGD Update",
|
||
"expression": "w = w - lr × gradient",
|
||
"variables": "weights (w), learning rate (lr), gradient",
|
||
"application": "Parameter update"
|
||
},
|
||
{
|
||
"name": "Adam Update",
|
||
"expression": "m = β1×m + (1-β1)×g; v = β2×v + (1-β2)×g²",
|
||
"variables": "moments (m,v), betas, gradient (g)",
|
||
"application": "Adaptive optimization"
|
||
},
|
||
{
|
||
"name": "Learning Rate Decay",
|
||
"expression": "lr = lr_0 × decay_rate^(epoch/decay_steps)",
|
||
"variables": "Initial lr, decay rate, epochs",
|
||
"application": "LR scheduling"
|
||
},
|
||
{
|
||
"name": "Effective Batch Size",
|
||
"expression": "batch_size × gradient_accumulation_steps × num_gpus",
|
||
"variables": "Batch size, accumulation, GPU count",
|
||
"application": "Distributed training"
|
||
},
|
||
{
|
||
"name": "Training Time",
|
||
"expression": "(dataset_size / batch_size) × epochs × time_per_step",
|
||
"variables": "Dataset size, batch size, epochs, step time",
|
||
"application": "Time estimation"
|
||
},
|
||
{
|
||
"name": "GPU Utilization",
|
||
"expression": "Active_time / Total_time × 100%",
|
||
"variables": "Active computation time, Total time",
|
||
"application": "Efficiency metric"
|
||
}
|
||
]
|
||
},
|
||
"efficient_ai": {
|
||
"concepts": [
|
||
"Model compression techniques",
|
||
"Pruning (structured and unstructured)",
|
||
"Quantization (post-training and QAT)",
|
||
"Knowledge distillation",
|
||
"Teacher-student models",
|
||
"Low-rank decomposition",
|
||
"Neural Architecture Search for efficiency",
|
||
"Efficient attention mechanisms",
|
||
"Mobile and edge optimizations",
|
||
"Lottery ticket hypothesis",
|
||
"Magnitude pruning",
|
||
"Gradual pruning",
|
||
"INT8/INT4 quantization",
|
||
"Mixed-precision inference",
|
||
"Model compression pipelines"
|
||
],
|
||
"terminology": [
|
||
"Sparsity", "Compression ratio",
|
||
"Teacher model", "Student model",
|
||
"Lottery ticket", "Winning ticket",
|
||
"Magnitude pruning", "Gradual pruning",
|
||
"Quantization", "Dequantization",
|
||
"INT8", "INT4", "Binary networks",
|
||
"Factorization", "SVD decomposition",
|
||
"Distillation loss", "Knowledge transfer"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Compression Ratio",
|
||
"expression": "Original_size / Compressed_size",
|
||
"variables": "Original model size, Compressed size",
|
||
"application": "Compression efficiency"
|
||
},
|
||
{
|
||
"name": "Sparsity",
|
||
"expression": "Pruned_params / Total_params × 100%",
|
||
"variables": "Pruned parameters, Total parameters",
|
||
"application": "Pruning measurement"
|
||
},
|
||
{
|
||
"name": "Quantization Speedup",
|
||
"expression": "Original_bits / Quantized_bits",
|
||
"variables": "Original precision, Quantized precision",
|
||
"application": "Speed improvement"
|
||
},
|
||
{
|
||
"name": "Distillation Loss",
|
||
"expression": "α×CE(student, labels) + (1-α)×KL(student, teacher)",
|
||
"variables": "Cross-entropy (CE), KL divergence, alpha weight",
|
||
"application": "Knowledge distillation"
|
||
},
|
||
{
|
||
"name": "Memory Savings",
|
||
"expression": "(1 - new_bits/original_bits) × 100%",
|
||
"variables": "Bit widths",
|
||
"application": "Memory reduction"
|
||
}
|
||
]
|
||
},
|
||
"optimizations": {
|
||
"concepts": [
|
||
"Compiler optimizations for ML",
|
||
"Graph optimization techniques",
|
||
"Operator fusion",
|
||
"Kernel optimization",
|
||
"Memory optimization strategies",
|
||
"Cache optimization",
|
||
"Vectorization and parallelization",
|
||
"Hardware-specific optimizations",
|
||
"Profiling and bottleneck analysis",
|
||
"Roofline model",
|
||
"Memory pooling",
|
||
"Tensor layout optimization",
|
||
"Loop tiling and blocking",
|
||
"SIMD instructions",
|
||
"Computational intensity"
|
||
],
|
||
"terminology": [
|
||
"Kernel fusion", "Graph rewriting",
|
||
"Memory pooling", "Memory reuse",
|
||
"Tensor layout", "NCHW", "NHWC",
|
||
"Loop tiling", "Blocking",
|
||
"SIMD", "Vector instructions",
|
||
"Cache locality", "Prefetching",
|
||
"Arithmetic intensity", "Bandwidth",
|
||
"FLOPs", "Memory bandwidth",
|
||
"Roofline", "Bottleneck"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Arithmetic Intensity",
|
||
"expression": "FLOPs / Bytes_accessed",
|
||
"variables": "Floating-point operations, Memory bytes",
|
||
"application": "Roofline analysis"
|
||
},
|
||
{
|
||
"name": "Cache Hit Rate",
|
||
"expression": "Cache_hits / Total_accesses × 100%",
|
||
"variables": "Cache hits, Total memory accesses",
|
||
"application": "Cache efficiency"
|
||
},
|
||
{
|
||
"name": "Memory Bandwidth Utilization",
|
||
"expression": "Actual_bandwidth / Peak_bandwidth × 100%",
|
||
"variables": "Measured bandwidth, Theoretical peak",
|
||
"application": "Memory efficiency"
|
||
},
|
||
{
|
||
"name": "Kernel Efficiency",
|
||
"expression": "Actual_FLOPs / Peak_FLOPs × 100%",
|
||
"variables": "Measured FLOPs, Theoretical peak",
|
||
"application": "Compute efficiency"
|
||
},
|
||
{
|
||
"name": "Speedup from Fusion",
|
||
"expression": "Time_unfused / Time_fused",
|
||
"variables": "Execution times",
|
||
"application": "Optimization impact"
|
||
}
|
||
]
|
||
},
|
||
"hw_acceleration": {
|
||
"concepts": [
|
||
"GPU architecture and programming",
|
||
"TPU and specialized accelerators",
|
||
"FPGA for ML inference",
|
||
"ASIC design for AI",
|
||
"Hardware-software co-design",
|
||
"Memory hierarchy and bandwidth",
|
||
"Tensor cores and matrix units",
|
||
"Power efficiency considerations",
|
||
"Heterogeneous computing",
|
||
"CUDA programming model",
|
||
"Warp execution model",
|
||
"Shared memory optimization",
|
||
"Memory coalescing",
|
||
"Systolic arrays",
|
||
"Neural Processing Units (NPUs)"
|
||
],
|
||
"terminology": [
|
||
"CUDA cores", "Tensor cores",
|
||
"Warp", "Block", "Grid",
|
||
"Shared memory", "Global memory",
|
||
"Memory coalescing", "Bank conflicts",
|
||
"Systolic arrays", "Matrix units",
|
||
"FLOPS", "TOPS", "Performance/Watt",
|
||
"SM (Streaming Multiprocessor)",
|
||
"Occupancy", "Throughput"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "GPU Utilization",
|
||
"expression": "Active_SMs / Total_SMs × 100%",
|
||
"variables": "Active streaming multiprocessors, Total SMs",
|
||
"application": "GPU efficiency"
|
||
},
|
||
{
|
||
"name": "Memory Bandwidth",
|
||
"expression": "Bus_width × Memory_clock × 2",
|
||
"variables": "Bus width, Clock frequency",
|
||
"application": "Bandwidth calculation"
|
||
},
|
||
{
|
||
"name": "Theoretical Peak FLOPS",
|
||
"expression": "Cores × Clock × Ops_per_cycle",
|
||
"variables": "Core count, Clock speed, Operations",
|
||
"application": "Peak performance"
|
||
},
|
||
{
|
||
"name": "Power Efficiency",
|
||
"expression": "FLOPS / Watts",
|
||
"variables": "Computational throughput, Power consumption",
|
||
"application": "Efficiency metric"
|
||
},
|
||
{
|
||
"name": "Speedup",
|
||
"expression": "CPU_time / GPU_time",
|
||
"variables": "Execution times",
|
||
"application": "Acceleration factor"
|
||
}
|
||
]
|
||
},
|
||
"benchmarking": {
|
||
"concepts": [
|
||
"Performance metrics and KPIs",
|
||
"Benchmarking methodologies",
|
||
"Profiling tools and techniques",
|
||
"A/B testing for ML systems",
|
||
"Statistical significance testing",
|
||
"Performance regression detection",
|
||
"Scalability analysis",
|
||
"Cost-performance tradeoffs",
|
||
"Latency measurement (P50, P95, P99)",
|
||
"Throughput measurement",
|
||
"Time-to-first-token (TTFT)",
|
||
"Cold start vs warm start",
|
||
"Strong vs weak scaling",
|
||
"MLPerf benchmarks",
|
||
"Performance variance analysis"
|
||
],
|
||
"terminology": [
|
||
"Latency", "P50", "P95", "P99",
|
||
"Throughput", "QPS (Queries Per Second)",
|
||
"TTFT (Time-to-first-token)",
|
||
"Cold start", "Warm start",
|
||
"Performance variance", "Jitter",
|
||
"Strong scaling", "Weak scaling",
|
||
"Benchmark", "Baseline",
|
||
"Profiler", "Trace"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Percentile Calculation",
|
||
"expression": "rank = ceil(p × n / 100)",
|
||
"variables": "Percentile (p), Sample count (n)",
|
||
"application": "Latency percentiles"
|
||
},
|
||
{
|
||
"name": "Throughput",
|
||
"expression": "Requests / Time_period",
|
||
"variables": "Request count, Time duration",
|
||
"application": "System throughput"
|
||
},
|
||
{
|
||
"name": "Latency Breakdown",
|
||
"expression": "Total = Preprocessing + Inference + Postprocessing",
|
||
"variables": "Component latencies",
|
||
"application": "Performance analysis"
|
||
},
|
||
{
|
||
"name": "Scaling Efficiency",
|
||
"expression": "Speedup / Number_of_processors",
|
||
"variables": "Speedup factor, Processor count",
|
||
"application": "Parallel efficiency"
|
||
},
|
||
{
|
||
"name": "Cost per Inference",
|
||
"expression": "(Hardware_cost + Energy_cost) / Total_inferences",
|
||
"variables": "Costs, Inference count",
|
||
"application": "Economic efficiency"
|
||
}
|
||
]
|
||
},
|
||
"ml_ops": {
|
||
"concepts": [
|
||
"CI/CD for ML systems",
|
||
"Model versioning and governance",
|
||
"Experiment tracking and management",
|
||
"Model monitoring and observability",
|
||
"Automated retraining pipelines",
|
||
"Feature drift and concept drift",
|
||
"Model rollback strategies",
|
||
"Infrastructure as Code for ML",
|
||
"Model registry design",
|
||
"A/B testing frameworks",
|
||
"Blue-green deployment",
|
||
"Canary releases",
|
||
"Shadow mode evaluation",
|
||
"Champion-challenger patterns",
|
||
"MLOps maturity models"
|
||
],
|
||
"terminology": [
|
||
"Model registry", "Artifact store",
|
||
"Pipeline orchestration",
|
||
"Data drift", "Concept drift",
|
||
"Model decay", "Retraining triggers",
|
||
"Blue-green deployment", "Canary release",
|
||
"Shadow mode", "Champion-challenger",
|
||
"CI/CD", "GitOps",
|
||
"Observability", "Monitoring"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Drift Detection (KL Divergence)",
|
||
"expression": "KL(P||Q) = Σ P(x) log(P(x)/Q(x))",
|
||
"variables": "Distributions P and Q",
|
||
"application": "Distribution shift detection"
|
||
},
|
||
{
|
||
"name": "Model Performance Decay",
|
||
"expression": "Performance(t) = Performance(0) × e^(-λt)",
|
||
"variables": "Initial performance, decay rate (λ), time",
|
||
"application": "Decay modeling"
|
||
},
|
||
{
|
||
"name": "Retraining Frequency",
|
||
"expression": "1 / Mean_time_to_decay_threshold",
|
||
"variables": "Time to performance threshold",
|
||
"application": "Retraining schedule"
|
||
},
|
||
{
|
||
"name": "A/B Test Sample Size",
|
||
"expression": "n = 2σ²(Z_α + Z_β)² / δ²",
|
||
"variables": "Variance, Z-scores, effect size",
|
||
"application": "Experiment design"
|
||
}
|
||
]
|
||
},
|
||
"ondevice_learning": {
|
||
"concepts": [
|
||
"Federated learning",
|
||
"Edge training techniques",
|
||
"Incremental learning",
|
||
"Few-shot adaptation",
|
||
"Privacy-preserving training",
|
||
"Resource-constrained optimization",
|
||
"Online learning algorithms",
|
||
"Personalization strategies",
|
||
"Federated averaging",
|
||
"Client selection strategies",
|
||
"Communication-efficient training",
|
||
"Differential privacy in FL",
|
||
"Local vs global model updates",
|
||
"Non-IID data handling",
|
||
"Asynchronous federated learning"
|
||
],
|
||
"terminology": [
|
||
"Federated averaging", "FedAvg",
|
||
"Differential privacy",
|
||
"Client selection", "Aggregation",
|
||
"Communication rounds",
|
||
"Local epochs", "Global rounds",
|
||
"Model personalization",
|
||
"Non-IID", "Data heterogeneity",
|
||
"Privacy budget", "Epsilon (ε)"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Communication Cost",
|
||
"expression": "Model_size × Clients × Rounds",
|
||
"variables": "Model size, Client count, Round count",
|
||
"application": "FL communication overhead"
|
||
},
|
||
{
|
||
"name": "Privacy Budget",
|
||
"expression": "ε_total = √(2k) × ε_round",
|
||
"variables": "Round privacy (ε_round), rounds (k)",
|
||
"application": "Differential privacy"
|
||
},
|
||
{
|
||
"name": "Convergence with Partial Participation",
|
||
"expression": "O(1/√(nK))",
|
||
"variables": "Clients (n), rounds (K)",
|
||
"application": "FL convergence rate"
|
||
},
|
||
{
|
||
"name": "Bandwidth Requirements",
|
||
"expression": "Updates/second × Update_size",
|
||
"variables": "Update frequency, Model update size",
|
||
"application": "Network planning"
|
||
}
|
||
]
|
||
},
|
||
"robust_ai": {
|
||
"concepts": [
|
||
"Adversarial robustness",
|
||
"Adversarial examples and attacks",
|
||
"Out-of-distribution detection",
|
||
"Uncertainty quantification",
|
||
"Model calibration",
|
||
"Robust optimization",
|
||
"Defense mechanisms",
|
||
"Certified robustness",
|
||
"Ensemble uncertainty",
|
||
"Calibration error measurement",
|
||
"Robust accuracy metrics",
|
||
"Attack budgets",
|
||
"Defense radius",
|
||
"Safety guarantees",
|
||
"Robustness verification"
|
||
],
|
||
"terminology": [
|
||
"Adversarial examples", "Perturbations",
|
||
"Certified robustness",
|
||
"Ensemble uncertainty",
|
||
"Calibration error",
|
||
"Robust accuracy",
|
||
"Attack budget", "Defense radius",
|
||
"FGSM", "PGD attack",
|
||
"Adversarial training"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Perturbation Budget",
|
||
"expression": "||δ||_p ≤ ε",
|
||
"variables": "Perturbation (δ), norm (p), budget (ε)",
|
||
"application": "Attack constraints"
|
||
},
|
||
{
|
||
"name": "Calibration Error",
|
||
"expression": "ECE = Σ |acc(bin) - conf(bin)| × |bin|/n",
|
||
"variables": "Accuracy, confidence per bin",
|
||
"application": "Calibration measurement"
|
||
},
|
||
{
|
||
"name": "Uncertainty (Entropy)",
|
||
"expression": "H = -Σ p(x) log p(x)",
|
||
"variables": "Probability distribution",
|
||
"application": "Uncertainty quantification"
|
||
},
|
||
{
|
||
"name": "Robust Loss",
|
||
"expression": "max_δ L(x+δ, y) s.t. ||δ|| ≤ ε",
|
||
"variables": "Loss, perturbation, constraint",
|
||
"application": "Adversarial training"
|
||
}
|
||
]
|
||
},
|
||
"security_privacy": {
|
||
"concepts": [
|
||
"Differential privacy mechanisms",
|
||
"Secure multi-party computation",
|
||
"Homomorphic encryption",
|
||
"Model inversion attacks",
|
||
"Membership inference attacks",
|
||
"Data poisoning attacks",
|
||
"Backdoor attacks",
|
||
"Secure aggregation",
|
||
"Privacy-utility tradeoffs",
|
||
"Noise mechanisms (Laplace, Gaussian)",
|
||
"Secure enclaves and TEE",
|
||
"Model stealing attacks",
|
||
"Privacy amplification",
|
||
"Gradient privacy",
|
||
"Private training algorithms"
|
||
],
|
||
"terminology": [
|
||
"Privacy budget", "Epsilon (ε)", "Delta (δ)",
|
||
"Laplace mechanism", "Gaussian mechanism",
|
||
"Secure enclaves", "TEE",
|
||
"Model stealing", "Extraction attacks",
|
||
"Backdoor", "Poisoning",
|
||
"Privacy amplification",
|
||
"Secure aggregation",
|
||
"Homomorphic encryption"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Differential Privacy Definition",
|
||
"expression": "Pr[M(D) ∈ S] ≤ e^ε × Pr[M(D') ∈ S] + δ",
|
||
"variables": "Mechanism M, datasets D/D', privacy params",
|
||
"application": "Privacy guarantee"
|
||
},
|
||
{
|
||
"name": "Laplace Noise Scale",
|
||
"expression": "scale = Δf/ε",
|
||
"variables": "Sensitivity (Δf), privacy budget (ε)",
|
||
"application": "Noise addition"
|
||
},
|
||
{
|
||
"name": "Privacy Composition",
|
||
"expression": "ε_total = Σ ε_i",
|
||
"variables": "Individual privacy budgets",
|
||
"application": "Sequential composition"
|
||
},
|
||
{
|
||
"name": "Attack Success Rate",
|
||
"expression": "ASR = Successful_attacks / Total_attempts",
|
||
"variables": "Attack outcomes",
|
||
"application": "Security evaluation"
|
||
}
|
||
]
|
||
},
|
||
"responsible_ai": {
|
||
"concepts": [
|
||
"Bias detection and mitigation",
|
||
"Fairness metrics and constraints",
|
||
"Demographic parity",
|
||
"Equal opportunity",
|
||
"Explainability and interpretability",
|
||
"LIME and SHAP explanations",
|
||
"Model transparency",
|
||
"Ethical AI frameworks",
|
||
"Regulatory compliance",
|
||
"Human-in-the-loop systems",
|
||
"Accountability mechanisms",
|
||
"Model cards",
|
||
"Datasheets for datasets",
|
||
"Algorithmic auditing",
|
||
"Fairness-aware learning"
|
||
],
|
||
"terminology": [
|
||
"Demographic parity", "Equal opportunity",
|
||
"Disparate impact", "Disparate treatment",
|
||
"LIME", "SHAP", "Feature importance",
|
||
"Model cards", "Datasheets",
|
||
"Algorithmic audit",
|
||
"Fairness constraint",
|
||
"Protected attributes",
|
||
"Bias mitigation"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Statistical Parity Difference",
|
||
"expression": "P(Y=1|A=0) - P(Y=1|A=1)",
|
||
"variables": "Outcome Y, Protected attribute A",
|
||
"application": "Fairness measurement"
|
||
},
|
||
{
|
||
"name": "Disparate Impact Ratio",
|
||
"expression": "P(Y=1|A=0) / P(Y=1|A=1)",
|
||
"variables": "Outcome probabilities by group",
|
||
"application": "Bias detection"
|
||
},
|
||
{
|
||
"name": "Equal Opportunity Difference",
|
||
"expression": "TPR(A=0) - TPR(A=1)",
|
||
"variables": "True positive rates by group",
|
||
"application": "Fairness metric"
|
||
},
|
||
{
|
||
"name": "SHAP Values",
|
||
"expression": "φ_i = Σ |S|!(p-|S|-1)!/p! × [f(S∪{i}) - f(S)]",
|
||
"variables": "Feature subsets, model output",
|
||
"application": "Feature importance"
|
||
}
|
||
]
|
||
},
|
||
"sustainable_ai": {
|
||
"concepts": [
|
||
"Carbon footprint of ML",
|
||
"Energy-efficient training",
|
||
"Green AI initiatives",
|
||
"Compute-optimal scaling",
|
||
"Sustainable hardware",
|
||
"Lifecycle assessment",
|
||
"Carbon-aware computing",
|
||
"Efficient infrastructure",
|
||
"Power Usage Effectiveness (PUE)",
|
||
"Renewable energy usage",
|
||
"Embodied carbon",
|
||
"Carbon credits and offsetting",
|
||
"Energy-aware scheduling",
|
||
"Model efficiency metrics",
|
||
"Sustainable deployment strategies"
|
||
],
|
||
"terminology": [
|
||
"Carbon intensity", "PUE",
|
||
"Energy per inference",
|
||
"Compute-optimal",
|
||
"Carbon credits", "Offsetting",
|
||
"Renewable energy",
|
||
"Embodied carbon",
|
||
"Green AI", "Red AI",
|
||
"Sustainability metrics"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Carbon Footprint",
|
||
"expression": "Energy_consumed × Carbon_intensity",
|
||
"variables": "Energy (kWh), Carbon intensity (gCO2/kWh)",
|
||
"application": "Emissions calculation"
|
||
},
|
||
{
|
||
"name": "PUE (Power Usage Effectiveness)",
|
||
"expression": "Total_facility_power / IT_equipment_power",
|
||
"variables": "Power measurements",
|
||
"application": "Data center efficiency"
|
||
},
|
||
{
|
||
"name": "Energy Efficiency",
|
||
"expression": "FLOPs / Joule",
|
||
"variables": "Computational work, Energy consumed",
|
||
"application": "Hardware efficiency"
|
||
},
|
||
{
|
||
"name": "Training Cost",
|
||
"expression": "kWh × $/kWh + Carbon_cost",
|
||
"variables": "Energy, electricity price, carbon price",
|
||
"application": "Economic impact"
|
||
},
|
||
{
|
||
"name": "Total Emissions",
|
||
"expression": "Power × Time × Carbon_intensity + Embodied_carbon",
|
||
"variables": "Power, duration, carbon factors",
|
||
"application": "Lifecycle assessment"
|
||
}
|
||
]
|
||
},
|
||
"ai_good": {
|
||
"concepts": [
|
||
"Social impact applications",
|
||
"Healthcare AI systems",
|
||
"Climate modeling with ML",
|
||
"Accessibility technologies",
|
||
"Educational AI tools",
|
||
"Humanitarian applications",
|
||
"Scientific discovery acceleration",
|
||
"Global challenges and ML solutions",
|
||
"Impact assessment methodologies",
|
||
"Beneficiary population identification",
|
||
"Resource-limited deployment",
|
||
"Community engagement",
|
||
"Participatory design",
|
||
"Technology transfer",
|
||
"Scalability for social good"
|
||
],
|
||
"terminology": [
|
||
"Impact assessment",
|
||
"Beneficiary populations",
|
||
"Resource-limited settings",
|
||
"Community engagement",
|
||
"Participatory design",
|
||
"Technology transfer",
|
||
"Social impact",
|
||
"Digital divide",
|
||
"Inclusive AI"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Impact Metrics",
|
||
"expression": "Lives_improved × Improvement_magnitude",
|
||
"variables": "Population reached, Improvement per person",
|
||
"application": "Social impact measurement"
|
||
},
|
||
{
|
||
"name": "Cost-Benefit Analysis",
|
||
"expression": "Social_benefit / Implementation_cost",
|
||
"variables": "Benefit value, Cost",
|
||
"application": "Project evaluation"
|
||
},
|
||
{
|
||
"name": "Accessibility Score",
|
||
"expression": "Accessible_features / Total_features",
|
||
"variables": "Feature counts",
|
||
"application": "Inclusivity measurement"
|
||
},
|
||
{
|
||
"name": "Scalability Projection",
|
||
"expression": "Impact × (Target_population / Current_population)",
|
||
"variables": "Current impact, Population sizes",
|
||
"application": "Scale planning"
|
||
}
|
||
]
|
||
},
|
||
"ml_workflow": {
|
||
"concepts": [
|
||
"End-to-end ML project lifecycle",
|
||
"Team collaboration and roles",
|
||
"Project management for ML",
|
||
"Technical debt management",
|
||
"Documentation best practices",
|
||
"Code review for ML",
|
||
"Reproducibility strategies",
|
||
"Knowledge transfer",
|
||
"Version control for ML",
|
||
"Experiment management",
|
||
"Pipeline orchestration",
|
||
"Testing strategies for ML",
|
||
"Continuous integration for ML",
|
||
"Model handoff processes",
|
||
"Production readiness checklists"
|
||
],
|
||
"terminology": [
|
||
"Technical debt", "Code smell",
|
||
"Reproducibility", "Replicability",
|
||
"Documentation debt",
|
||
"Bus factor",
|
||
"Code review",
|
||
"Version control",
|
||
"Git", "Experiment tracking",
|
||
"Pipeline", "Orchestration"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Technical Debt Interest",
|
||
"expression": "Time_saved_now vs Time_cost_later ratio",
|
||
"variables": "Short-term vs long-term time costs",
|
||
"application": "Debt assessment"
|
||
},
|
||
{
|
||
"name": "Team Velocity",
|
||
"expression": "Story_points / Sprint_duration",
|
||
"variables": "Completed work, Time period",
|
||
"application": "Productivity measurement"
|
||
},
|
||
{
|
||
"name": "Code Coverage",
|
||
"expression": "Tested_lines / Total_lines × 100%",
|
||
"variables": "Code line counts",
|
||
"application": "Test completeness"
|
||
},
|
||
{
|
||
"name": "Documentation Completeness",
|
||
"expression": "Documented_functions / Total_functions",
|
||
"variables": "Function counts",
|
||
"application": "Documentation quality"
|
||
}
|
||
]
|
||
},
|
||
"conclusion": {
|
||
"concepts": [
|
||
"Future trends and directions",
|
||
"Emerging technologies",
|
||
"Career development in ML systems",
|
||
"Continuous learning strategies",
|
||
"Community engagement",
|
||
"Open challenges in ML systems",
|
||
"Research opportunities",
|
||
"Industry evolution",
|
||
"Technology readiness levels",
|
||
"Career pathways",
|
||
"Professional development",
|
||
"Open-source contribution",
|
||
"Emerging paradigms",
|
||
"Cross-disciplinary integration",
|
||
"Future of ML systems engineering"
|
||
],
|
||
"terminology": [
|
||
"Emerging paradigms",
|
||
"Technology readiness",
|
||
"Career pathways",
|
||
"Professional development",
|
||
"Community contribution",
|
||
"Open-source",
|
||
"Research frontiers",
|
||
"Industry trends",
|
||
"Innovation cycles"
|
||
],
|
||
"formulas": [
|
||
{
|
||
"name": "Technology Adoption Curve",
|
||
"expression": "Adopters(t) = N × (1 - e^(-kt))",
|
||
"variables": "Time, adoption rate (k), population (N)",
|
||
"application": "Adoption modeling"
|
||
},
|
||
{
|
||
"name": "Research Impact",
|
||
"expression": "Citations × Implementation_adoptions",
|
||
"variables": "Academic and practical impact",
|
||
"application": "Impact assessment"
|
||
},
|
||
{
|
||
"name": "Career Growth Rate",
|
||
"expression": "Skills_gained / Time_period",
|
||
"variables": "Skill acquisition, Time",
|
||
"application": "Professional development"
|
||
}
|
||
]
|
||
},
|
||
"emerging_topics": {
|
||
"concepts": [],
|
||
"terminology": [],
|
||
"formulas": []
|
||
},
|
||
"emerging_challenges": {
|
||
"concepts": [],
|
||
"terminology": [],
|
||
"formulas": []
|
||
}
|
||
},
|
||
"cumulative_map": {
|
||
"introduction": {
|
||
"chapter_number": 1,
|
||
"new_concepts": ["Machine Learning Systems Engineering", "AI vs ML vs Deep Learning distinctions", "Supervised, Unsupervised, Reinforcement Learning", "ML Systems vs traditional software systems", "Data-driven vs rule-based approaches", "ML development lifecycle", "AI pervasiveness in modern applications", "Symbolic AI and rule-based systems", "Expert Systems and knowledge representation", "Statistical Learning paradigm shift", "Shallow Learning (SVM, decision trees)", "Deep Learning revolution", "Neural network fundamentals", "Perceptron history", "Backpropagation introduction", "AlexNet and ImageNet breakthrough", "Foundation models concept", "Model-Data-Infrastructure triangle", "ML system lifecycle phases", "Production deployment challenges"],
|
||
"new_terminology": ["Model", "Dataset", "Features", "Labels", "Training", "Validation", "Testing", "Overfitting", "Underfitting", "Generalization", "Bias", "Variance", "Bias-variance tradeoff", "Loss function", "Optimization", "Hyperparameters", "Parameters", "Inference", "Serving", "Perceptron", "Activation function", "Gradient descent", "Epochs", "Iterations", "Batch size", "Learning rate"],
|
||
"new_formulas": [
|
||
{
|
||
"name": "Accuracy",
|
||
"expression": "Accuracy = Correct Predictions / Total Predictions",
|
||
"variables": "Correct Predictions, Total Predictions",
|
||
"application": "Model evaluation"
|
||
},
|
||
{
|
||
"name": "Error Rate",
|
||
"expression": "Error = 1 - Accuracy",
|
||
"variables": "Accuracy",
|
||
"application": "Model evaluation"
|
||
},
|
||
{
|
||
"name": "Train/Test Split",
|
||
"expression": "70% train, 15% validation, 15% test (typical)",
|
||
"variables": "Dataset size",
|
||
"application": "Data partitioning"
|
||
}
|
||
],
|
||
"assumed_knowledge": {
|
||
"concepts": [],
|
||
"terminology": [],
|
||
"formulas": []
|
||
},
|
||
"total_concepts_known": 20,
|
||
"calc_questions_possible": 3
|
||
},
|
||
"ml_systems": {
|
||
"chapter_number": 2,
|
||
"new_concepts": ["Cloud ML architecture", "Edge ML deployment", "Mobile ML constraints", "TinyML and embedded systems", "Deployment spectrum tradeoffs", "Data center infrastructure", "Power consumption constraints", "Memory hierarchy in ML systems", "Latency vs Throughput optimization", "Network connectivity requirements", "Resource-constrained deployment", "System reliability and monitoring", "MLOps practices", "Model versioning and governance", "A/B testing for ML", "Shadow mode deployment", "Production ML challenges", "ML technical debt"],
|
||
"new_terminology": ["Inference", "Serving", "Deployment", "Data pipeline", "Feature pipeline", "Model registry", "Model versioning", "A/B testing", "Shadow mode", "SLA (Service Level Agreement)", "MLOps", "Model governance", "Edge device", "Embedded system", "Data center", "Hyperscale", "PUE (Power Usage Effectiveness)", "Thermal limits", "Battery constraints"],
|
||
"new_formulas": [
|
||
{
|
||
"name": "Latency Percentiles",
|
||
"expression": "P50, P95, P99 calculations",
|
||
"variables": "Response time distribution",
|
||
"application": "Performance measurement"
|
||
},
|
||
{
|
||
"name": "Throughput",
|
||
"expression": "Throughput = Requests / Second",
|
||
"variables": "Number of requests, Time",
|
||
"application": "System performance"
|
||
},
|
||
{
|
||
"name": "System Availability",
|
||
"expression": "Availability = Uptime / Total Time × 100%",
|
||
"variables": "Uptime, Total Time",
|
||
"application": "Reliability measurement"
|
||
},
|
||
{
|
||
"name": "Power Efficiency",
|
||
"expression": "Performance / Watt",
|
||
"variables": "Computational performance, Power consumption",
|
||
"application": "Energy efficiency"
|
||
}
|
||
],
|
||
"assumed_knowledge": {
|
||
"concepts": ["Machine Learning fundamentals", "Model-Data-Infrastructure concepts"],
|
||
"terminology": ["Model", "Training", "Inference", "Dataset"],
|
||
"formulas": [
|
||
{
|
||
"name": "Accuracy",
|
||
"expression": "Accuracy = Correct Predictions / Total Predictions"
|
||
}
|
||
]
|
||
},
|
||
"total_concepts_known": 38,
|
||
"calc_questions_possible": 7
|
||
},
|
||
"nn_computation": {
|
||
"chapter_number": 3,
|
||
"new_concepts": ["Neural network architecture", "Biological vs artificial neurons", "Perceptron model", "Multi-layer perceptrons (MLP)", "Forward propagation", "Backward propagation algorithm", "Activation functions (ReLU, Sigmoid, Tanh)", "Gradient descent optimization", "Stochastic gradient descent (SGD)", "Deep vs shallow networks", "Representation learning", "Feature hierarchies", "Vanishing/Exploding gradients", "Weight initialization", "Loss function types", "Training vs inference phases", "Pre-processing and post-processing", "Network topology design"],
|
||
"new_terminology": ["Neurons", "Layers", "Weights", "Biases", "Forward pass", "Backward pass", "Learning rate", "Batch size", "Epochs", "Gradients", "Chain rule", "Hidden layers", "Input layer", "Output layer", "Activation", "Non-linearity", "Loss", "Cost function", "Optimizer", "Momentum", "Mini-batch", "Iteration"],
|
||
"new_formulas": [
|
||
{
|
||
"name": "Neuron Output",
|
||
"expression": "y = f(Σ(wi×xi) + b)",
|
||
"variables": "weights (w), inputs (x), bias (b), activation (f)",
|
||
"application": "Neural computation"
|
||
},
|
||
{
|
||
"name": "ReLU Activation",
|
||
"expression": "f(x) = max(0, x)",
|
||
"variables": "input x",
|
||
"application": "Non-linear activation"
|
||
},
|
||
{
|
||
"name": "Sigmoid Activation",
|
||
"expression": "σ(x) = 1/(1+e^(-x))",
|
||
"variables": "input x",
|
||
"application": "Probability output"
|
||
},
|
||
{
|
||
"name": "Parameter Count",
|
||
"expression": "params = (input_size + 1) × output_size",
|
||
"variables": "input_size, output_size",
|
||
"application": "Memory estimation"
|
||
},
|
||
{
|
||
"name": "Memory for Activations",
|
||
"expression": "batch_size × layer_size × precision_bytes",
|
||
"variables": "batch_size, layer_size, precision",
|
||
"application": "Memory planning"
|
||
}
|
||
],
|
||
"assumed_knowledge": {
|
||
"concepts": ["ML Systems architecture", "Cloud/Edge/Mobile deployment"],
|
||
"terminology": ["Model", "Training", "Inference", "Deployment", "MLOps"],
|
||
"formulas": ["Accuracy", "Throughput", "System metrics"]
|
||
},
|
||
"total_concepts_known": 56,
|
||
"calc_questions_possible": 12
|
||
}
|
||
}
|
||
}
|