Examinable Topics for Revision
ACTL3143 & ACTL5111 Deep Learning for Actuaries
Exam details
- Exam is on Inspera
- Open book
- No handwritten answers, maybe have a calculator handy
- You’ll have 1.5 hours to complete plus 15 mins reading
- Complete the IT preparation checklist (MFA, speed test, read policies)
Lecture 1: AI
- artificial intelligence
- artificial intelligence vs machine learning
- classification problem
Deep Blue- labelled/unlabelled data
- machine learning
minimax algorithm- pseudocode
- regression problem
- targets
Lecture 1: Python
- default arguments
- dictionaries
- f-strings
- function definitions
Google Colaboratoryhelp
- list
pip install ...
range
- slicing
- tuple
type
- whitespace indentation
- zero-indexing
Lecture 2: Deep Learning
- activations, activation function
- artificial neural network
- biases (in neurons)
- callbacks
- cost/loss function
- deep/shallow network, network depth
- dense or fully-connected layer
- early stopping
- epoch
- feed-forward neural network
Keras, Tensorflow, PyTorchmatplotlib, seaborn- neural network architecture
- overfitting
perceptron- ReLU activation
- representation learning
- training/validation/test split
- universal approximation theorem
- weights (in a neuron)
Tutorial 2: Forward Pass
- batches, batch size
- forward pass of network
- gradient-based learning
- learning rate
- stochastic (mini-batch) gradient descent
Lecture 3: Mixed Topics
Categorical Variables
- entity embeddings
- Input layer
- Keras functional API
- nominal variables
- ordinal variables
- Reshape layer
- skip connection
- wide & deep network
Classification
- accuracy
- confusion matrix
- cross-entropy loss
- metrics
- sigmoid activation
- sofmax activation
Lecture 4: Computer Vision
- channels
- computer vision
- convolutional layer & CNN
- error analysis
- filter/kernel
- flatten layer
- max pooling
- MNIST
- stride
- tensor (rank, dimension)
Tutorial 4: Backpropagation
- backpropagation
- partial derivatives
Lecture 5: Natural Language
- bag of words
- lemmatization
- one-hot embedding
- stop words
- vocabulary
- word embeddings/vectors
- word2vec
Week 6: Uncertainy Quantification
- aleatoric and epistemic uncertainty
- Bayesian neural network
- deep ensembles
- dropout
- ensemble model
- CANN
- GLM
- MDN
- mixture distribution
- Monte Carlo dropout
- posterior sampling
proper scoring rule- uncertainty quantification
variational approximation
Lecture 7: Recurrent Networks
- GRU
- LSTM
- recurrent neural networks
- SimpleRNN
Lecture 8: Transfer Learning
AlexNet, GoogLeNet, Inception- ImageNet
- fine-tuning
- transfer learning
Lecture 8-9: Generative Networks
- autoencoder (
variational) - beam search
- bias
ChatGPT (& RLHF)DeepDreamgenerative adversarial networks- greedy sampling
Hugging Face- language model
- latent space
neural style transfer- softmax temperature
- stochastic sampling
Lecture 9: Interpretability
- global interpretability
Grad-CAM- inherent interpretability
- LIME
- local interpretability
- permutation importance
- post-hoc interpretability
SHAP values
StoryWalls
Go AI: Basic Python- Sydney Temperature Forecasting: Basic MLP
- Victorian Crash Severity: Classification & entity-embedding
- Hurricane damage: Convolutional neural networks and hyperparameter tuning
- US Police reports: NLP with bag-of-words, TF-IDF, permutation importance
- Health Insurance Premiums: Uncertainty Quantification
Generative networks experimenting