**Course Name : Deep Learning with TensorFlow**

**Module 1:- Intro to Tensorflow**

**Question 1 : Which statement is FALSE about TensorFlow?**

- TensorFlow is well suited for handling Deep Learning Problems
**TensorFlow library is not proper for handling Machine Learning Problems**- TensorFlow has a C/C++ backend as well as Python modules
- TensorFlow is an open source library
- All of the above

**Question 2 : What is a Data Flow Graph?**

**A representation of data dependencies between operations**- A cartesian (x,y) chart
- A graphics user interface
- A flowchart describing an algorithm
- None of the above

**Question 3 : What is the main reasons of increasing popularity of Deep Learning?**

- The advances in machine learning algorithms and research.
- The availability of massive amounts of data for training computer systems.
- The dramatic increases in computer processing capabilities.
**All of the above**

**Question 4 : Which statement is TRUE about TensorFlow?**

**Runs on CPU and GPU**- Runs on CPU only
- Runs on GPU only

**Question 5 : Why is TensorFlow the proper library for Deep Learning?**

- It will benefit from TensorFlow’s auto-differentiation and suite of first-rate optimizers
- It provides a collection of trainable mathematical functions that are useful for neural networks.
- It has extensive built-in support for deep learning
**All of the above**

**Module 2 :- Convolutional Networks**

**Question 1 : What can be achieved with “convolution” operations on Images?**

- Noise Filtering
- Image Smoothing
- Image Blurring
- Edge Detection
**All of the above**

**Question 2 : For convolution, it is better to store images in a TensorFlow Graph as:**

**Placeholder**- CSV file
- Numpy array
- Variable
- None of the above

**Question 3 : Which of the following statements is TRUE about Convolution Neural Networks (CNNs)?**

- CNN can be applied ONLY on Image and Text data
**CNN can be applied on ANY 2D and 3D array of data**- CNN can be applied ONLY on Text and Speech data
- CNN can be applied ONLY on Image data
- All of the above

**Question 4 : Which of the following Layers can be part of Convolution Neural Networks (CNNs)**

- Dropout
- Softmax
- Maxpooling
- Relu
**All of the above**

**Module 3 :- Recurrent Neural Networks**

**Question 1 : What is a Recurrent Neural Network?**

**A Neural Network that can recur to itself, and is proper for handling sequential data**- An infinite layered Neural Network which is proper for handling structured data
- A special kind of Neural Network to predict weather
- A markovian model to handle temporal data

**Question 2 : What is NOT TRUE about RNNs?**

- RNNs are VERY suitable for sequential data.
- RNNs need to keep track of states, which is computationally expensive.
**RNNs are very robust against vanishing gradient problem.**

**Question 3 : What application(s) is(are) suitable for RNNs?**

- Estimating temperatures from weather data
- Natural Language Processing
- Video context retriever
- Speech Recognition
**All of the above**

**Question 4 : Why are RNNs susceptible to issues with their gradients?**

- Numerical computation of gradients can drive into instabilities
- Gradients can quickly drop and stabilize at near zero
- Propagation of errors due to the recurrent characteristic
- Gradients can grow exponentially
**All of the above**

**Question 5 : What is TRUE about LSTM gates?**

- The Read Gate in LSTM, determine how much old information to forget
- The Write Gate in LSTM, reads data from the memory cell and sends that data back to the network.
**The Forget Gate, in LSTM maintains or deletes data from the information cell.**- The Read Gate in LSTM, is responsible for writing data into the memory cell.

**Module 4 :- Restricted Boltzmann Machine**

**Question 1 : What is the main application of RBM?**

- Data dimensionality reduction
- Feature extraction
- Collaborative filtering
**All of the above**

**Question 2 : How many layers does an RBM (Restricted Boltzmann Machine) have?**

- Infinte
- 4
**2**- 3
- All of the above

**Question 3 : How does an RBM compare to a PCA?**

- RBM cannot reduce dimensionality
- PCA cannot generate original data
- PCA is another type of Neural Network
**Both can regenerate input data**- All of the above

**Question 4 : Which statement is TRUE about RBM?**

- It is a Boltzmann machine, but with no connections between nodes in the same layer
- Each node in the first layer has a bias
- The RBM reconstructs data by making several forward and backward passes between the visible and hidden layers
- At the hidden layer’s nodes, X is multiplied by a W (weight matrix) and added to h_bias
**All of the above**

**Question 5 : Which statement is TRUE statement about an RBM?**

- The objective function is to maximize the likelihood of our data being drawn from the reconstructed data distribution
- The Negative phase of an RBM decreases the probability of samples generated by the model
- Contrastive Divergence (CD) is used to approximate the negative phase of an RBM
- The Positive phase of an RBM increases the probability of training data
**All of the above**

**Module 5 :- Autoecoders**

**Question 1 : what is the difference between Autoencoders and RBMs?**

- Autoencoders are used for supervised learning, but RBMs are used for unsupervised learning.
**Autoencoders use a deterministic approach, but RBMs use a stochastic approach.**- Autoencoders have less layeres than RBMs.
- All of the above

**Question 2 : Which of the following problems cannot be solved by Autoencoders:**

- Dimensionality Reduction
**Time series prediction**- Image Reconstruction
- Emotion Detection
- All of the above

**Question 3 : What is TRUE about Autoencoders:**

- Help to Reduce the Curse of Dimensionality
- Used to Learn the Most important Features in Data
- Used for Unsupervised Learning
**All of the Above**

**Question 4 : What are Autoencoders:**

- A Neural Network that is designed to replace Non-Linear Regression
**A Neural Network that is trained to attempt to copy its input to its output**- A Neural Network that learns all the weights by using labeled data
- A Neural Network where different layer inputs are controlled by gates
- All of the Above

**Question 5 : What is a Deep Autoencoder:**

**An Autoencoder with Multiple Hidden Layers**- An Autoencoder with multiple input and output layers
- An Autoencoder stacked with Multiple Visible Layers
- An Autoencoder stacked with over 1000 layers
- None of the Above

**Deep Learning with TensorFlow Cognitive Class Final Exam Answers:-**

**Question 1 : Why use a Data Flow graph to solve Mathematical expressions?**

**To create a pipeline of operations and its corresponding values to be parsed**- To represent the expression in a human-readable form
- To show the expression in a GUI
- Because it is the only way to solve mathematical expressions in a digital computer
- None of the above

**Question 2 : What is an Activation Function?**

**A function that triggers a neuron and generates the outputs**- A function that models a phenomenon or process
- A function to normalize the output
- All of the above
- None of the above

**Question 3 : Why is TensorFlow considered fast and suitable for Deep Learning?**

- It is suitable to operate over large multi-dimensional tensors
- It runs on CPU
- Its core is based on C++
- It runs on GPU
**All of the above**

**Question 4 : Can TensorFlow replace Numpy?**

- None of the above
- No, whatsoever
- With only Numpy we can’t solve Deep Learning problems, therefore, TensorFlow is required
- Yes, completely
**Partially for some operations on tensors, such as minimization**

**Question 5 : What is FALSE about Convolution Neural Networks (CNNs)?**

**They fully connect to all neurons in all of the layers**- They connect only to neurons in the local region (kernel size) of input images
- They build feature maps hierarchically in every layer
- They are inspired by human visual systems
- None of the above

**Question 6 : What is the meaning of “Strides” in Maxpooling?**

- The number of pixels the kernel should add
**The number of pixels the kernel should move**- The size of the kernel
- The number of pixels the kernel should remove
- None of the above

**Question 7 : What is TRUE about “Padding” in Convolution?**

**size of the input image is reduced for the “VALID” padding**- Size of the input image is reduced for the “SAME” padding
- Size of the input image is increased for the “SAME” padding
- Size of the input image is increased for the “VALID” padding
- All of the above

**Question 8 : Which of the following best describes the Relu Function?**

- (-1,1)
- (0,5)
**(0, Max)**- (-inf,inf)
- (0,1)

**Question 9 : Which are types of Recurrent Neural Networks? (Select all that apply)**

- LSTM
- Hopfield Network
- Recursive Neural Network
- Deep Belief Network
**Elman Networks and Jordan Networks**

**Question 10 : Which is TRUE about RNNs?**

- RNNs can predict the future
**RNNs are VERY suitable for sequential data**- RNNs are NOT suitable for sequential data
- RNNs are ONLY suitable for sequential data
- All of the above

**Question 11 : What is the problem with RNNs and gradients?**

- Numerical computation of gradients can drive into instabilities
- Gradients can quickly drop and stabilize at near zero
- Propagation of errors due to the recurrent characteristic
- Gradients can grow exponentially
**All of the above**

**Question 12 : What type of RNN would you use in an NLP project to predict the next word in a phrase? (only one is correct)**

- Bi-directional RNN
- Neural history compressor
**Long Short-Term Memory**- Echo state network
- None of the above

**Question 13 : Which one does NOT happen in the “forward pass” in RBM?**

**Making a deterministic decision about returning values into network.**- Multiplying inputs by weights, and adding an overall bias, in each hidden unit.
- Applying an activation function on the results in hidden units.
- Feeding the nework with the input images converted to binary values.

**Question 14 : Which one IS NOT a sample of CNN application?**

- Creating art images using pre-trained models
- Object Detection in images
- Coloring black and white images
**Predicting next word in a sentence**

**Question 15 : Select all possible uses of Autoencoders and RBMs (select all that apply):**

**Clustering****Pattern Recognition****Dimensionality Reduction****Predict data in time series**

**Question 16 : Which technique is proper for solving Collaborative Filtering problem?**

- DBN
**RBM**- CNN
- RNN

**Question 17 : Which statement is TRUE for training Autoencoders?**

- The Size of Last Layer must be at least 10% of the Input Layer Dimension
**The size of input and Last Layers must be of the Same Dimensions**- The Last Layer must be Double the size of Input Layer Dimension
- The Last Layer must be half the size of Input Layer Dimension
- None of the Above

**Question 18 : To Design a Deep Autoencoder Architecture, what factors are to be considered?**

- The size of the centre-most layer has to be close to number of Important Features to be extracted
- The centre-most layer should have the smallest size compared to all other layers
- The Network should have an odd number of layers
- All the layers must be symmetrical with respect to the centre-most layer
**All of the Above**

**Question 19 : With is TRUE about Back-propogation?**

- It can be used to train LSTMs
- It can be used to train CNNs
- It can be used to train RBMs
- It can be used to train Autoencoders
**All of the Above**

**Question 20 : How can Autoencoders be improved to handle higly non-linear data?**

- By using Genetic Algorithms
**By adding more Hidden Layers to the Network**- By using Higher initial Weight Values
- By using Lower initial Weight Values
- All of the Above