cvpr maths123

You are training a deep neural network on a binary classification task, using a sigmoid activation in the final layer, and a mixture of tanh and ReLU activations for all other layers. You notice that the gradients to these layers completely go to zero very early on in training. Which of the following fixes could help? • Increase the size of your training set • Switch the ReLU activations with leaky ReLUs everywhere • Add Batch Normalization before every activation • Changing the SGD to adam optimizer At test time, a Dropout layer with a dropout probability α will: • Multiply the input tensor by α. • Drop neurons from the input with probability α. • Power normalizes the input tensor by α. • None of the above. If your input to the convolutional layer is 32 × 32 × 16, how many parameters are there in a single 1 × 1 convolution filter, including bias? • 2 • 17 • 1025 • 1 Which methods can be used to prevent overfitting? • Changing scheduler • Changing batch size • L2 regularization • Dropout A (32 × 32 × 3) image is passed through a convolutional layer with 32 filters, each of size (3 × 3). The padding size is 1 and the stride size is 4. What is the size of the output feature map? • (32 × 32 × 3) • (16 × 16 × 32) • (8 × 8 × 32) • (32 × 32 × 32) The questions in this section should be answered in maximum 4 sentences. Please be concise in your responses. Your project is in collaboration with the police, and they want you to design an AI system to identify criminals, given their data. Since being imprisoned is a severe punishment, it is very important for AI system to not incorrectly identify the criminals, and simultaneously ensure that the place is as safe as possible. What could be the most appropriate evaluation metric and why: Accuracy, Precision, Mean average precision, or F1 score? You are solving a face recognition (modeled as binary classification) that uses facial data to help users log into their devices. You train a classification model for user X until it achieves 98% accuracy on the validation set for same user. However, upon deployment (at test time) the model has a 50% misclassification rate. List the two most important factors for the differences in performance between the validation set and test set Pooling layers such as Max or Sum pooling in Deep Nets cause a loss of information between convolutional layers, why would we add Pooling layers to our network? Why do we need activation functions (ReLU etc) in neural networks? A medical company approaches you for deep learning expertise. They want to classify X-ray images into 2 classes (disease or healthy). In the given dataset the disease class is very underrepresented (1 disease for every 10 healthy). Precisely describe two techniques that you can use during training that help alleviate the class imbalance problem.