Deep learning quesions
Budget: $10 – $30 USD
Q1:
A set of radiomic features extracted from single-energy CT scans and their corresponding labels
are given. It is an imbalance binary classification problem. Build a model for classification with
the following requirements.
(a) Handle the imbalanced dataset with over sampling.
(b) Perform feature analysis and selection (30 most significant features)
(c) Apply recursive feature elimination with a random forest classifier to select ten most
discriminative features in (b).
(d) Use ensemble learning models like Random Forest and Multi-layer Perceptron and optimize
them with grid searching.
(e) Use nested cross validation with a 20% for the test.
(f) Use Repeated K-fold with 10 splits and 10 repeats for the test.
(g) Calculate Classification metrics including (accuracy, F1, precision, recall) and AUC
Provide Code and report readability
"""""""""""""""""""""""""""""""""""""
Q2:
Using CIFAR10 image dataset,
(a) Build a deep neural network having 20 hidden layers of 100 neurons. Use He initializations
and RELU activation function. Use Adam optimization with early stopping and train the network.
(b) Then add Batch normalization and compare the learning curve: Is it converging faster than
before? Does it produce a better model? How does it affect training speed?
(c) Use grid-searching for hyperparameters and model’s architecture.
- You can load the dataset with keras.datasets.cifar10.loaddata().
The dataset is composed of 60,000 32x32-pixel color image (48,000 for training & 12,000 for
testing) with 10 classes, so you will need a SoftMax output layer with 10 neurons.
You need to submit your reproducible python code as well as a report including the results.
A set of radiomic features extracted from single-energy CT scans and their corresponding labels
are given. It is an imbalance binary classification problem. Build a model for classification with
the following requirements.
(a) Handle the imbalanced dataset with over sampling.
(b) Perform feature analysis and selection (30 most significant features)
(c) Apply recursive feature elimination with a random forest classifier to select ten most
discriminative features in (b).
(d) Use ensemble learning models like Random Forest and Multi-layer Perceptron and optimize
them with grid searching.
(e) Use nested cross validation with a 20% for the test.
(f) Use Repeated K-fold with 10 splits and 10 repeats for the test.
(g) Calculate Classification metrics including (accuracy, F1, precision, recall) and AUC
Provide Code and report readability
"""""""""""""""""""""""""""""""""""""
Q2:
Using CIFAR10 image dataset,
(a) Build a deep neural network having 20 hidden layers of 100 neurons. Use He initializations
and RELU activation function. Use Adam optimization with early stopping and train the network.
(b) Then add Batch normalization and compare the learning curve: Is it converging faster than
before? Does it produce a better model? How does it affect training speed?
(c) Use grid-searching for hyperparameters and model’s architecture.
- You can load the dataset with keras.datasets.cifar10.loaddata().
The dataset is composed of 60,000 32x32-pixel color image (48,000 for training & 12,000 for
testing) with 10 classes, so you will need a SoftMax output layer with 10 neurons.
You need to submit your reproducible python code as well as a report including the results.