Deep Learning

763 pages in the complete text edition.

All subject text · Markdown Original content · JSON

MT2 · Chapter

MT2 · Formulas

MT2 · Guide

MT2 · Index

MT2 · Mocks

MT2 · Section

Examination schedule and scopeMT2 preparation and examination plan
Recommended reading orderMT2 preparation and examination plan
Preparation sessions before 13 OctoberMT2 preparation and examination plan
Mock-paper conventionsMT2 preparation and examination plan
Writing technical answersMT2 preparation and examination plan
Before maths: what are we trying to learn?Mathematics revision
Negative numbers are directionMathematics revision
Fractions decimals and percentagesMathematics revision
Brackets before everything elseMathematics revision
Powers roots and very small numbersMathematics revision
Variables functions and symbolsMathematics revision
Exponentials and natural logarithmsMathematics revision
Calculator routineMathematics revision
A complete arithmetic calculationMathematics revision
Loss: how one wrong prediction becomes a numberMathematics revision
A vector is a list and a matrix is a tableThe maths behind a network
Matrix multiplication: every row has a jobThe maths behind a network
A graph and a slopeThe maths behind a network
Derivatives without a calculus courseThe maths behind a network
Derivatives: from an ordinary change to a local slopeThe maths behind a network
Partial derivatives hold other things stillThe maths behind a network
The chain rule is multiplying linked effectsThe maths behind a network
The derivative of prediction errorThe maths behind a network
Loss scaling: derive the gradient instead of guessing a factorThe maths behind a network
Probability mean and variance in plain languageThe maths behind a network
Mean, variance and standard deviation from ordinary arithmeticThe maths behind a network
Standardization: what the answer 1 actually meansThe maths behind a network
Chain rule: the complete calculationThe maths behind a network
What does learning from data meanNeural-network foundations
Your first complete training stepNeural-network foundations
Classification regression and learning typesNeural-network foundations
Training validation and test dataNeural-network foundations
Parameters and hyperparametersNeural-network foundations
What one artificial neuron doesNeural-network foundations
Model answer: explain an artificial neuronNeural-network foundations
Activation functions and how to calculate themNeural-network foundations
Model answer: why a network needs nonlinear activationNeural-network foundations
Choosing a lossNeural-network foundations
Forward propagation through multiple neuronsNeural-network foundations
Backpropagation with one hidden neuronNeural-network foundations
More outputs and vectorized backpropagationNeural-network foundations
Model answer: forward propagation, backpropagation and updatingNeural-network foundations
Counting parameters in a dense networkNeural-network foundations
What you should understand before optimizationNeural-network foundations
Complete sigmoid training exampleNeural-network foundations
Cross-entropy for probability distributionsNeural-network foundations
Reading an optimizer formula without getting lostOptimization algorithms
Optimizer notation: old value, memory, new valueOptimization algorithms
What an optimizer is trying to doOptimization algorithms
The ordinary gradient descent updateOptimization algorithms
Batch stochastic and mini batch gradient descentOptimization algorithms
Model answer: batch, stochastic and mini-batch trainingOptimization algorithms
A tiny dataset shows why the methods differOptimization algorithms
Learning rate and common training problemsOptimization algorithms
A moving average remembers the recent pastOptimization algorithms
Momentum remembers movementOptimization algorithms
Nesterov looks before correctingOptimization algorithms
Model answer: momentum and NesterovOptimization algorithms
AdaGrad adds every squared gradientOptimization algorithms
RMSProp remembers recent squared gradientsOptimization algorithms
Adam remembers both direction and squared sizeOptimization algorithms
Adam bias correction: why the denominator is one minus beta to a powerOptimization algorithms
Model answer: AdaGrad, RMSProp and AdamOptimization algorithms
One comparison to rebuild all optimizer formulasOptimization algorithms
AdaGrad and RMSProp with a negative gradientOptimization algorithms
Recognizing overfitting and underfittingRegularization and generalization
L1 and L2 add a cost for large weightsRegularization and generalization
Model answer: regularization and the L1–L2 differenceRegularization and generalization
Dropout removes activations during trainingRegularization and generalization
Early stopping uses the best validation pointRegularization and generalization
Model answer: generalization and early stoppingRegularization and generalization
Choosing a remedyRegularization and generalization
Inverted dropout during training and inferenceRegularization and generalization
Model answer: dropout with training and inference explainedRegularization and generalization
Reading image shapes before doing CNN arithmeticConvolutional neural networks
A computer sees an image as numbersConvolutional neural networks
The convolution calculationConvolutional neural networks
Stride padding and floorConvolutional neural networks
Convolution over several channelsConvolutional neural networks
One filter over multiple channels: all contributions go into one cellConvolutional neural networks
Choosing more or fewer filtersConvolutional neural networks
Pooling needs comparisons or averagesConvolutional neural networks
Parameters are different from activation valuesConvolutional neural networks
One complete CNN pipelineConvolutional neural networks
CNN versus a fully connected image networkConvolutional neural networks
Model answer: CNN structure and parameter sharingConvolutional neural networks
Output size and parameter count are differentConvolutional neural networks
Model answer: shape, activations and parametersConvolutional neural networks
Faculty example: a complete binary-image convolutionConvolutional neural networks
Faculty example: combine three RGB channel sumsConvolutional neural networks
Max, min and average pooling on the same lecture gridConvolutional neural networks
Faculty CNN table: shapes, activations and 62958 parametersConvolutional neural networks
Even-sized kernels: SAME padding is not always symmetricConvolutional neural networks
AlexNet and the reason for its componentsCNN architectures and transfer learning
Inception examines several spatial scalesCNN architectures and transfer learning
Why one by one convolution saves workCNN architectures and transfer learning
ResNet adds a shortcutCNN architectures and transfer learning
Transfer learning from a small datasetCNN architectures and transfer learning
Model answer: feature extraction and fine-tuningCNN architectures and transfer learning
How to draw the architecture diagramsCNN architectures and transfer learning
Inception concatenation and residual additionCNN architectures and transfer learning
Model answer: AlexNet, Inception and ResNetCNN architectures and transfer learning
Applications of convolutional neural networksCNN architectures and transfer learning
Reading LeNet and AlexNet layer diagramsCNN architectures and transfer learning
Top-1 and top-5 error: what an ImageNet result meansCNN architectures and transfer learning
VGG16: small filters in a regular sequenceCNN architectures and transfer learning
Global average pooling: one average for each channelCNN architectures and transfer learning
The faculty's 480-channel bottleneck numericalCNN architectures and transfer learning
Batch normalization inside a residual networkCNN architectures and transfer learning
Residual-block order and architecture qualificationsCNN architectures and transfer learning
Choosing frozen layers from dataset size and similarityCNN architectures and transfer learning
Four ways to reuse a pretrained modelCNN architectures and transfer learning
Start hereRecurrent neural networks
What makes a sequence differentRecurrent neural networks
Model answer: why recurrent networks use a hidden stateRecurrent neural networks
The vanilla RNN equations from scratchRecurrent neural networks
A complete vector RNN calculationRecurrent neural networks
Sequence mappings and applicationsRecurrent neural networks
Shared weights and RNN applicationsRecurrent neural networks
Start herePrerequisite repair and teacher backprop examples
MT1A Q2Prerequisite repair and teacher backprop examples
MT1A Q3Prerequisite repair and teacher backprop examples
MT1A Q4Prerequisite repair and teacher backprop examples
MT1B Q3Prerequisite repair and teacher backprop examples
MT1B Q4Prerequisite repair and teacher backprop examples
Teacher vectorized ReLU examplePrerequisite repair and teacher backprop examples
Teams photos full sigmoid examplePrerequisite repair and teacher backprop examples
Start hereF25 selected MT2 practice
F25 Q1cF25 selected MT2 practice
F25 Q1dF25 selected MT2 practice
F25 Q2bF25 selected MT2 practice
F25 Q3bF25 selected MT2 practice
F25 Q4aF25 selected MT2 practice
F25 Q5bF25 selected MT2 practice
F25 Q6aF25 selected MT2 practice
F25 Q7bF25 selected MT2 practice
Start hereR26 selected MT2 practice
R26 Q1cR26 selected MT2 practice
R26 Q1dR26 selected MT2 practice
R26 Q2bR26 selected MT2 practice
R26 Q3bR26 selected MT2 practice
R26 Q4aR26 selected MT2 practice
R26 Q5aR26 selected MT2 practice
R26 Q5bR26 selected MT2 practice
R26 Q6aR26 selected MT2 practice
R26 Q7bR26 selected MT2 practice
Start hereF23 selected MT2 practice
F23 Q1bF23 selected MT2 practice
F23 Q1dF23 selected MT2 practice
F23 Q3bF23 selected MT2 practice
F23 Q4aF23 selected MT2 practice
F23 Q5aF23 selected MT2 practice
F23 Q6bF23 selected MT2 practice
F23 Q7aF23 selected MT2 practice
F23 Q7bF23 selected MT2 practice
Start hereR24 selected MT2 practice
R24 Q1cR24 selected MT2 practice
R24 Q1dR24 selected MT2 practice
R24 Q2bR24 selected MT2 practice
R24 Q3bR24 selected MT2 practice
R24 Q4bR24 selected MT2 practice
R24 Q5aR24 selected MT2 practice
R24 Q6bR24 selected MT2 practice
R24 Q7bR24 selected MT2 practice
Start hereF22 selected MT2 practice
F22 Q1aF22 selected MT2 practice
F22 Q1bF22 selected MT2 practice
F22 Q1eF22 selected MT2 practice
F22 Q3aF22 selected MT2 practice
F22 Q3bF22 selected MT2 practice
F22 Q4aF22 selected MT2 practice
F22 Q4bF22 selected MT2 practice
F22 Q5bF22 selected MT2 practice
F22 Q6aF22 selected MT2 practice
F22 Q6bF22 selected MT2 practice
Start hereR23 selected MT2 practice
R23 Q1aR23 selected MT2 practice
R23 Q1bR23 selected MT2 practice
R23 Q1dR23 selected MT2 practice
R23 Q2aR23 selected MT2 practice
R23 Q2bR23 selected MT2 practice
R23 Q4aR23 selected MT2 practice
R23 Q4bR23 selected MT2 practice
R23 Q6aR23 selected MT2 practice
R23 Q6bR23 selected MT2 practice
R23 Q7bR23 selected MT2 practice
Start hereMT2 numerical practice from direct to mixed
How to solve a mixed numerical without guessingMT2 numerical practice from direct to mixed
Drill 1 Signs and bracketsMT2 numerical practice from direct to mixed
Drill 2 Half sum versus meanMT2 numerical practice from direct to mixed
Drill 3 Natural log and probabilityMT2 numerical practice from direct to mixed
Drill 4 Dot product and biasMT2 numerical practice from direct to mixed
Drill 5 Matrix rows receive inputsMT2 numerical practice from direct to mixed
Drill 6 Chain rule on a tiny lossMT2 numerical practice from direct to mixed
Drill 7 Sigmoid with binary cross entropyMT2 numerical practice from direct to mixed
Drill 8 Activation contrastMT2 numerical practice from direct to mixed
Drill 9 Softmax and true class lossMT2 numerical practice from direct to mixed
Drill 10 Sigmoid neuron gradientMT2 numerical practice from direct to mixed
Drill 11 Full ReLU backpropagationMT2 numerical practice from direct to mixed
Drill 12 Two output sigmoid backpropagationMT2 numerical practice from direct to mixed
Drill 13 A negative ReLU blocks this pathMT2 numerical practice from direct to mixed
Drill 14 Batch and sequential SGDMT2 numerical practice from direct to mixed
Drill 15 Incomplete mini batchMT2 numerical practice from direct to mixed
Drill 16 GD variation 1MT2 numerical practice from direct to mixed
Drill 17 GD variation 2MT2 numerical practice from direct to mixed
Drill 18 Momentum variation 1MT2 numerical practice from direct to mixed
Drill 19 Momentum variation 2MT2 numerical practice from direct to mixed
Drill 20 Nesterov variation 1MT2 numerical practice from direct to mixed
Drill 21 Nesterov variation 2MT2 numerical practice from direct to mixed
Drill 22 AdaGrad variation 1MT2 numerical practice from direct to mixed
Drill 23 AdaGrad variation 2MT2 numerical practice from direct to mixed
Drill 24 RMSProp variation 1MT2 numerical practice from direct to mixed
Drill 25 RMSProp variation 2MT2 numerical practice from direct to mixed
Drill 26 Adam variation 1MT2 numerical practice from direct to mixed
Drill 27 Adam variation 2MT2 numerical practice from direct to mixed
Drill 28 Regularized objectiveMT2 numerical practice from direct to mixed
Drill 29 L1 update on positive and negative weightsMT2 numerical practice from direct to mixed
Drill 30 L2 updateMT2 numerical practice from direct to mixed
Drill 31 Inverted dropoutMT2 numerical practice from direct to mixed
Drill 32 Convolution with a negative coefficientMT2 numerical practice from direct to mixed
Drill 33 Padding stride and channelsMT2 numerical practice from direct to mixed
Drill 34 Max and average poolingMT2 numerical practice from direct to mixed
Drill 35 Complete small CNNMT2 numerical practice from direct to mixed
Drill 36 Reverse a spatial formulaMT2 numerical practice from direct to mixed
Drill 37 Bottleneck cost with new valuesMT2 numerical practice from direct to mixed
Drill 38 Residual additionMT2 numerical practice from direct to mixed
Drill 39 Two RNN stepsMT2 numerical practice from direct to mixed
Drill 40 Mixed convolution pool dense and lossMT2 numerical practice from direct to mixed
Drill 41 Mixed backpropagation and L2MT2 numerical practice from direct to mixed
Drill 42 Mixed dropout and a momentum updateMT2 numerical practice from direct to mixed
Drill 43 Mixed CNN dimensions softmax and lossMT2 numerical practice from direct to mixed
Drill 44 Inception concatenation and parameter countMT2 numerical practice from direct to mixed
Drill 45 Projection shortcutMT2 numerical practice from direct to mixed
Revising solved examplesMT2 numerical practice from direct to mixed
Start hereTwo complete MT2 mock papers
MT2 Mock A: full questions and answersTwo complete MT2 mock papers
MT2 Mock B: full questions and answersTwo complete MT2 mock papers
The one page method mapFormula reference and exam revision
All formulas: meanings, use and substitutionFormula reference and exam revision
Calculator sequence without model guessworkFormula reference and exam revision
What to write if the exam page feels blankFormula reference and exam revision
Answer lengths and partial credit habitsFormula reference and exam revision
Methods to revise before the examinationFormula reference and exam revision
Learning and the numbers we adjustCourse vocabulary in plain words
Images and sequencesCourse vocabulary in plain words
Reading a question preciselyCourse vocabulary in plain words

TEE · Chapter

TEE · Formulas

TEE · Guide

TEE · Index

TEE · Mocks

TEE · Section

Coverage and examination structureTEE preparation and examination plan
Recommended reading orderTEE preparation and examination plan
Theory answers and numerical solutionsTEE preparation and examination plan
Practice resourcesTEE preparation and examination plan
All recovered TEE papers: unit, topic and marks mapTEE preparation and examination plan
Before maths: what are we trying to learn?Mathematics revision
Negative numbers are directionMathematics revision
Fractions decimals and percentagesMathematics revision
Brackets before everything elseMathematics revision
Powers roots and very small numbersMathematics revision
Variables functions and symbolsMathematics revision
Exponentials and natural logarithmsMathematics revision
Calculator routineMathematics revision
A complete arithmetic calculationMathematics revision
Loss: how one wrong prediction becomes a numberMathematics revision
A vector is a list and a matrix is a tableThe maths behind a network
Matrix multiplication: every row has a jobThe maths behind a network
A graph and a slopeThe maths behind a network
Derivatives without a calculus courseThe maths behind a network
Derivatives: from an ordinary change to a local slopeThe maths behind a network
Partial derivatives hold other things stillThe maths behind a network
The chain rule is multiplying linked effectsThe maths behind a network
The derivative of prediction errorThe maths behind a network
Loss scaling: derive the gradient instead of guessing a factorThe maths behind a network
Probability mean and variance in plain languageThe maths behind a network
Mean, variance and standard deviation from ordinary arithmeticThe maths behind a network
Standardization: what the answer 1 actually meansThe maths behind a network
Chain rule: the complete calculationThe maths behind a network
What does learning from data meanNeural-network foundations
Your first complete training stepNeural-network foundations
Classification regression and learning typesNeural-network foundations
Training validation and test dataNeural-network foundations
Parameters and hyperparametersNeural-network foundations
Artificial intelligence machine learning and deep learningNeural-network foundations
What one artificial neuron doesNeural-network foundations
Model answer: explain an artificial neuronNeural-network foundations
Perceptrons and the thresholdNeural-network foundations
Why XOR needs more than one perceptronNeural-network foundations
Activation functions and how to calculate themNeural-network foundations
Model answer: why a network needs nonlinear activationNeural-network foundations
Choosing a lossNeural-network foundations
Forward propagation through multiple neuronsNeural-network foundations
Backpropagation with one hidden neuronNeural-network foundations
More outputs and vectorized backpropagationNeural-network foundations
Model answer: forward propagation, backpropagation and updatingNeural-network foundations
Counting parameters in a dense networkNeural-network foundations
What you should understand before optimizationNeural-network foundations
Complete sigmoid training exampleNeural-network foundations
Cross-entropy for probability distributionsNeural-network foundations
Biological neurons and the McCulloch-Pitts modelNeural-network foundations
Reading an optimizer formula without getting lostOptimization algorithms
Optimizer notation: old value, memory, new valueOptimization algorithms
What an optimizer is trying to doOptimization algorithms
The ordinary gradient descent updateOptimization algorithms
Batch stochastic and mini batch gradient descentOptimization algorithms
Model answer: batch, stochastic and mini-batch trainingOptimization algorithms
A tiny dataset shows why the methods differOptimization algorithms
Learning rate and common training problemsOptimization algorithms
A moving average remembers the recent pastOptimization algorithms
Momentum remembers movementOptimization algorithms
Nesterov looks before correctingOptimization algorithms
Model answer: momentum and NesterovOptimization algorithms
AdaGrad adds every squared gradientOptimization algorithms
RMSProp remembers recent squared gradientsOptimization algorithms
Adam remembers both direction and squared sizeOptimization algorithms
Adam bias correction: why the denominator is one minus beta to a powerOptimization algorithms
Model answer: AdaGrad, RMSProp and AdamOptimization algorithms
One comparison to rebuild all optimizer formulasOptimization algorithms
AdaGrad and RMSProp with a negative gradientOptimization algorithms
Recognizing overfitting and underfittingRegularization and generalization
L1 and L2 add a cost for large weightsRegularization and generalization
Model answer: regularization and the L1–L2 differenceRegularization and generalization
Dropout removes activations during trainingRegularization and generalization
Early stopping uses the best validation pointRegularization and generalization
Model answer: generalization and early stoppingRegularization and generalization
Choosing a remedyRegularization and generalization
Inverted dropout during training and inferenceRegularization and generalization
Model answer: dropout with training and inference explainedRegularization and generalization
Reading image shapes before doing CNN arithmeticConvolutional neural networks
A computer sees an image as numbersConvolutional neural networks
The convolution calculationConvolutional neural networks
Stride padding and floorConvolutional neural networks
Convolution over several channelsConvolutional neural networks
One filter over multiple channels: all contributions go into one cellConvolutional neural networks
Choosing more or fewer filtersConvolutional neural networks
Pooling needs comparisons or averagesConvolutional neural networks
Parameters are different from activation valuesConvolutional neural networks
One complete CNN pipelineConvolutional neural networks
CNN versus a fully connected image networkConvolutional neural networks
Model answer: CNN structure and parameter sharingConvolutional neural networks
Output size and parameter count are differentConvolutional neural networks
Model answer: shape, activations and parametersConvolutional neural networks
Faculty example: a complete binary-image convolutionConvolutional neural networks
Faculty example: combine three RGB channel sumsConvolutional neural networks
Max, min and average pooling on the same lecture gridConvolutional neural networks
Faculty CNN table: shapes, activations and 62958 parametersConvolutional neural networks
Even-sized kernels: SAME padding is not always symmetricConvolutional neural networks
AlexNet and the reason for its componentsCNN architectures and transfer learning
Inception examines several spatial scalesCNN architectures and transfer learning
Why one by one convolution saves workCNN architectures and transfer learning
ResNet adds a shortcutCNN architectures and transfer learning
Transfer learning from a small datasetCNN architectures and transfer learning
Model answer: feature extraction and fine-tuningCNN architectures and transfer learning
How to draw the architecture diagramsCNN architectures and transfer learning
Inception concatenation and residual additionCNN architectures and transfer learning
Model answer: AlexNet, Inception and ResNetCNN architectures and transfer learning
Applications of convolutional neural networksCNN architectures and transfer learning
Reading LeNet and AlexNet layer diagramsCNN architectures and transfer learning
Top-1 and top-5 error: what an ImageNet result meansCNN architectures and transfer learning
VGG16: small filters in a regular sequenceCNN architectures and transfer learning
Global average pooling: one average for each channelCNN architectures and transfer learning
The faculty's 480-channel bottleneck numericalCNN architectures and transfer learning
Batch normalization inside a residual networkCNN architectures and transfer learning
Residual-block order and architecture qualificationsCNN architectures and transfer learning
Choosing frozen layers from dataset size and similarityCNN architectures and transfer learning
Four ways to reuse a pretrained modelCNN architectures and transfer learning
Complete the entire sequence chapterRecurrent neural networks
What makes a sequence differentRecurrent neural networks
Model answer: why recurrent networks use a hidden stateRecurrent neural networks
Reading recurrent memory and gatesRecurrent neural networks
The vanilla RNN equations from scratchRecurrent neural networks
A complete vector RNN calculationRecurrent neural networks
How an RNN is trainedRecurrent neural networks
BPTT with every arithmetic step visibleRecurrent neural networks
Vanishing and exploding gradientsRecurrent neural networks
Model answer: backpropagation through timeRecurrent neural networks
LSTM a memory path with three gatesRecurrent neural networks
Reading an LSTM equation one operation at a timeRecurrent neural networks
LSTM complete numerical with supplied gatesRecurrent neural networks
LSTM numerical when gates must first be calculatedRecurrent neural networks
GRU two gates and one stateRecurrent neural networks
Comparing the three recurrent modelsRecurrent neural networks
Model answer: LSTM and GRU memoryRecurrent neural networks
Sequence mappings and applicationsRecurrent neural networks
Recurrent networks: the complete relationshipRecurrent neural networks
Gate logic without memorizing a storyRecurrent neural networks
Start hereCurrent MT1 and the two teacher backprop examples
MT1A Q1aCurrent MT1 and the two teacher backprop examples
MT1A Q1bCurrent MT1 and the two teacher backprop examples
MT1A Q2Current MT1 and the two teacher backprop examples
MT1A Q3Current MT1 and the two teacher backprop examples
MT1A Q4Current MT1 and the two teacher backprop examples
MT1B Q1aCurrent MT1 and the two teacher backprop examples
MT1B Q1bCurrent MT1 and the two teacher backprop examples
MT1B Q2Current MT1 and the two teacher backprop examples
MT1B Q3Current MT1 and the two teacher backprop examples
MT1B Q4Current MT1 and the two teacher backprop examples
Teacher vectorized ReLU exampleCurrent MT1 and the two teacher backprop examples
Teams photos full sigmoid exampleCurrent MT1 and the two teacher backprop examples
Start hereFinal examination December 2025 fully solved
F25 Q1aFinal examination December 2025 fully solved
F25 Q1bFinal examination December 2025 fully solved
F25 Q1cFinal examination December 2025 fully solved
F25 Q1dFinal examination December 2025 fully solved
F25 Q2aFinal examination December 2025 fully solved
F25 Q2bFinal examination December 2025 fully solved
F25 Q3aFinal examination December 2025 fully solved
F25 Q3bFinal examination December 2025 fully solved
F25 Q4aFinal examination December 2025 fully solved
F25 Q4bFinal examination December 2025 fully solved
F25 Q5aFinal examination December 2025 fully solved
F25 Q5bFinal examination December 2025 fully solved
F25 Q6aFinal examination December 2025 fully solved
F25 Q6bFinal examination December 2025 fully solved
F25 Q7aFinal examination December 2025 fully solved
F25 Q7bFinal examination December 2025 fully solved
Start hereRe examination February 2026 fully solved
R26 Q1aRe examination February 2026 fully solved
R26 Q1bRe examination February 2026 fully solved
R26 Q1cRe examination February 2026 fully solved
R26 Q1dRe examination February 2026 fully solved
R26 Q2aRe examination February 2026 fully solved
R26 Q2bRe examination February 2026 fully solved
R26 Q3aRe examination February 2026 fully solved
R26 Q3bRe examination February 2026 fully solved
R26 Q4aRe examination February 2026 fully solved
R26 Q4bRe examination February 2026 fully solved
R26 Q5aRe examination February 2026 fully solved
R26 Q5bRe examination February 2026 fully solved
R26 Q6aRe examination February 2026 fully solved
R26 Q6bRe examination February 2026 fully solved
R26 Q7aRe examination February 2026 fully solved
R26 Q7bRe examination February 2026 fully solved
Start hereFinal examination December 2023 fully solved
F23 Q1aFinal examination December 2023 fully solved
F23 Q1bFinal examination December 2023 fully solved
F23 Q1cFinal examination December 2023 fully solved
F23 Q1dFinal examination December 2023 fully solved
F23 Q2aFinal examination December 2023 fully solved
F23 Q2bFinal examination December 2023 fully solved
F23 Q3aFinal examination December 2023 fully solved
F23 Q3bFinal examination December 2023 fully solved
F23 Q4aFinal examination December 2023 fully solved
F23 Q4bFinal examination December 2023 fully solved
F23 Q5aFinal examination December 2023 fully solved
F23 Q5bFinal examination December 2023 fully solved
F23 Q6aFinal examination December 2023 fully solved
F23 Q6bFinal examination December 2023 fully solved
F23 Q7aFinal examination December 2023 fully solved
F23 Q7bFinal examination December 2023 fully solved
Start hereRe examination February 2024 fully solved
R24 Q1aRe examination February 2024 fully solved
R24 Q1bRe examination February 2024 fully solved
R24 Q1cRe examination February 2024 fully solved
R24 Q1dRe examination February 2024 fully solved
R24 Q2aRe examination February 2024 fully solved
R24 Q2bRe examination February 2024 fully solved
R24 Q3aRe examination February 2024 fully solved
R24 Q3bRe examination February 2024 fully solved
R24 Q4aRe examination February 2024 fully solved
R24 Q4bRe examination February 2024 fully solved
R24 Q5aRe examination February 2024 fully solved
R24 Q5bRe examination February 2024 fully solved
R24 Q6aRe examination February 2024 fully solved
R24 Q6bRe examination February 2024 fully solved
R24 Q7aRe examination February 2024 fully solved
R24 Q7bRe examination February 2024 fully solved
Start hereFinal examination November 2022 fully solved
F22 Q1aFinal examination November 2022 fully solved
F22 Q1bFinal examination November 2022 fully solved
F22 Q1cFinal examination November 2022 fully solved
F22 Q1dFinal examination November 2022 fully solved
F22 Q1eFinal examination November 2022 fully solved
F22 Q2aFinal examination November 2022 fully solved
F22 Q2bFinal examination November 2022 fully solved
F22 Q3aFinal examination November 2022 fully solved
F22 Q3bFinal examination November 2022 fully solved
F22 Q4aFinal examination November 2022 fully solved
F22 Q4bFinal examination November 2022 fully solved
F22 Q5aFinal examination November 2022 fully solved
F22 Q5bFinal examination November 2022 fully solved
F22 Q6aFinal examination November 2022 fully solved
F22 Q6bFinal examination November 2022 fully solved
F22 Q7aFinal examination November 2022 fully solved
F22 Q7bFinal examination November 2022 fully solved
Start hereRe examination January 2023 fully solved
R23 Q1aRe examination January 2023 fully solved
R23 Q1bRe examination January 2023 fully solved
R23 Q1cRe examination January 2023 fully solved
R23 Q1dRe examination January 2023 fully solved
R23 Q1eRe examination January 2023 fully solved
R23 Q2aRe examination January 2023 fully solved
R23 Q2bRe examination January 2023 fully solved
R23 Q3aRe examination January 2023 fully solved
R23 Q3bRe examination January 2023 fully solved
R23 Q4aRe examination January 2023 fully solved
R23 Q4bRe examination January 2023 fully solved
R23 Q5aRe examination January 2023 fully solved
R23 Q5bRe examination January 2023 fully solved
R23 Q6aRe examination January 2023 fully solved
R23 Q6bRe examination January 2023 fully solved
R23 Q7aRe examination January 2023 fully solved
R23 Q7bRe examination January 2023 fully solved
Start hereNumerical practice from direct to mixed
How to solve a mixed numerical without guessingNumerical practice from direct to mixed
Drill 1 Signs and bracketsNumerical practice from direct to mixed
Drill 2 Half sum versus meanNumerical practice from direct to mixed
Drill 3 Natural log and probabilityNumerical practice from direct to mixed
Drill 4 Dot product and biasNumerical practice from direct to mixed
Drill 5 Matrix rows receive inputsNumerical practice from direct to mixed
Drill 6 Chain rule on a tiny lossNumerical practice from direct to mixed
Drill 7 Sigmoid with binary cross entropyNumerical practice from direct to mixed
Drill 8 Activation contrastNumerical practice from direct to mixed
Drill 9 Softmax and true class lossNumerical practice from direct to mixed
Drill 10 Sigmoid neuron gradientNumerical practice from direct to mixed
Drill 11 One perceptron learning updateNumerical practice from direct to mixed
Drill 12 XOR using two hidden unitsNumerical practice from direct to mixed
Drill 13 Full ReLU backpropagationNumerical practice from direct to mixed
Drill 14 Two output sigmoid backpropagationNumerical practice from direct to mixed
Drill 15 A negative ReLU blocks this pathNumerical practice from direct to mixed
Drill 16 Batch and sequential SGDNumerical practice from direct to mixed
Drill 17 Incomplete mini batchNumerical practice from direct to mixed
Drill 18 GD variation 1Numerical practice from direct to mixed
Drill 19 GD variation 2Numerical practice from direct to mixed
Drill 20 Momentum variation 1Numerical practice from direct to mixed
Drill 21 Momentum variation 2Numerical practice from direct to mixed
Drill 22 Nesterov variation 1Numerical practice from direct to mixed
Drill 23 Nesterov variation 2Numerical practice from direct to mixed
Drill 24 AdaGrad variation 1Numerical practice from direct to mixed
Drill 25 AdaGrad variation 2Numerical practice from direct to mixed
Drill 26 RMSProp variation 1Numerical practice from direct to mixed
Drill 27 RMSProp variation 2Numerical practice from direct to mixed
Drill 28 Adam variation 1Numerical practice from direct to mixed
Drill 29 Adam variation 2Numerical practice from direct to mixed
Drill 30 Regularized objectiveNumerical practice from direct to mixed
Drill 31 L1 update on positive and negative weightsNumerical practice from direct to mixed
Drill 32 L2 updateNumerical practice from direct to mixed
Drill 33 Inverted dropoutNumerical practice from direct to mixed
Drill 34 Convolution with a negative coefficientNumerical practice from direct to mixed
Drill 35 Padding stride and channelsNumerical practice from direct to mixed
Drill 36 Max and average poolingNumerical practice from direct to mixed
Drill 37 Complete small CNNNumerical practice from direct to mixed
Drill 38 Reverse a spatial formulaNumerical practice from direct to mixed
Drill 39 Bottleneck cost with new valuesNumerical practice from direct to mixed
Drill 40 Residual additionNumerical practice from direct to mixed
Drill 41 Two RNN stepsNumerical practice from direct to mixed
Drill 42 LSTM with a negative candidateNumerical practice from direct to mixed
Drill 43 Calculate gates before the cellNumerical practice from direct to mixed
Drill 44 GRU reset then mixNumerical practice from direct to mixed
Drill 45 BPTT shared weightsNumerical practice from direct to mixed
Drill 46 Mixed convolution pool dense and lossNumerical practice from direct to mixed
Drill 47 Mixed backpropagation and L2Numerical practice from direct to mixed
Drill 48 Mixed dropout and a momentum updateNumerical practice from direct to mixed
Drill 49 Mixed CNN dimensions softmax and lossNumerical practice from direct to mixed
Drill 50 Three step LSTM retention and replacementNumerical practice from direct to mixed
Drill 51 Inception concatenation and parameter countNumerical practice from direct to mixed
Drill 52 Projection shortcutNumerical practice from direct to mixed
Revising solved examplesNumerical practice from direct to mixed
Start hereComplete TEE paper rehearsal
Rehearsal Q1aComplete TEE paper rehearsal
Rehearsal Q1bComplete TEE paper rehearsal
Rehearsal Q1cComplete TEE paper rehearsal
Rehearsal Q1dComplete TEE paper rehearsal
Rehearsal Q2aComplete TEE paper rehearsal
Rehearsal Q2bComplete TEE paper rehearsal
Rehearsal Q3aComplete TEE paper rehearsal
Rehearsal Q3bComplete TEE paper rehearsal
Rehearsal Q4aComplete TEE paper rehearsal
Rehearsal Q4bComplete TEE paper rehearsal
Rehearsal Q5aComplete TEE paper rehearsal
Rehearsal Q5bComplete TEE paper rehearsal
Rehearsal Q6aComplete TEE paper rehearsal
Rehearsal Q6bComplete TEE paper rehearsal
Rehearsal Q7aComplete TEE paper rehearsal
Rehearsal Q7bComplete TEE paper rehearsal
The one page method mapFormula reference and exam revision
All formulas: meanings, use and substitutionFormula reference and exam revision
Calculator sequence without model guessworkFormula reference and exam revision
What to write if the exam page feels blankFormula reference and exam revision
Answer lengths and partial credit habitsFormula reference and exam revision
Methods to revise before the examinationFormula reference and exam revision
Learning and the numbers we adjustCourse vocabulary in plain words
Images and sequencesCourse vocabulary in plain words
Reading a question preciselyCourse vocabulary in plain words
R25 Q1a · ML versus DL and architecture choiceFebruary 2025 re-exam · complete solved paper
R25 Q1b · Training and validation errorsFebruary 2025 re-exam · complete solved paper
R25 Q1c · CNN hierarchy and stride-two convolutionFebruary 2025 re-exam · complete solved paper
R25 Q1d · BPTT and recurrent strengthsFebruary 2025 re-exam · complete solved paper
R25 Q2a · Activation, ReLU and numeric outputsFebruary 2025 re-exam · complete solved paper
R25 Q2b · Vanishing/exploding gradients and GRUFebruary 2025 re-exam · complete solved paper
R25 Q3a · 4–3–2 forward pass and dropoutFebruary 2025 re-exam · complete solved paper
R25 Q3b · Four CNN design explanationsFebruary 2025 re-exam · complete solved paper
R25 Q4a · 64×64 CNN shape and parameter tableFebruary 2025 re-exam · complete solved paper
R25 Q4b · Bipolar AND perceptronFebruary 2025 re-exam · complete solved paper
R25 Q5a · ResNet shortcut and Inception1×1February 2025 re-exam · complete solved paper
R25 Q5b · ReLU RNN state with declared orientationFebruary 2025 re-exam · complete solved paper
R25 Q6a · 3–2–1 forward pass and update w7,w8February 2025 re-exam · complete solved paper
R25 Q6b · AdaGrad stagnation versus RMSPropFebruary 2025 re-exam · complete solved paper
R25 Q7a · Long-text sentiment architectureFebruary 2025 re-exam · complete solved paper
R25 Q7b · Transfer learningFebruary 2025 re-exam · complete solved paper
R25 Q7c · Learning rate and two lossesFebruary 2025 re-exam · complete solved paper
Faculty Assignment2 Q1 · One momentum update with an omitted coefficientFaculty Assignment 2 · worked practice
Faculty Assignment2 Q2 · Specified momentum updateFaculty Assignment 2 · worked practice
Faculty Assignment2 Q3 · 28×28 CNNFaculty Assignment 2 · worked practice
Faculty Assignment2 Q4 · Batch linear regression updateFaculty Assignment 2 · worked practice
Faculty Assignment2 Q5 · 32×32 three-convolution CNNFaculty Assignment 2 · worked practice
Faculty Assignment2 Q6 · 64×64 two-convolution CNNFaculty Assignment 2 · worked practice