Podcast
Questions and Answers
What is the main challenge in creating a rule-based computational system for image classification?
What is the main challenge in creating a rule-based computational system for image classification?
- Understanding consumer experiences depicted in the images (correct)
- Determining the color channels in an image
- Identifying the pixel values in images
- Processing large volumes of labeled training data
In an RGB image, how many integers represent the value of each pixel?
In an RGB image, how many integers represent the value of each pixel?
- Two integers per pixel
- Four integers per pixel
- One integer per pixel
- Three integers per pixel (correct)
How is a greyscale image represented in terms of pixel values?
How is a greyscale image represented in terms of pixel values?
- As a binary value of 0 or 1
- As a single value ranging from 0 to 255 (correct)
- Using three values, one for each color channel
- As a vector of arbitrary length
What dimension does each RGB image in the dataset have?
What dimension does each RGB image in the dataset have?
How many numbers are in the single dimensional vector representing each image?
How many numbers are in the single dimensional vector representing each image?
What is the primary function of the convolution operation in a Convolutional Neural Network?
What is the primary function of the convolution operation in a Convolutional Neural Network?
What is the purpose of padding in the context of convolutional layers?
What is the purpose of padding in the context of convolutional layers?
Which activation function is mentioned as commonly used in Convolutional Neural Networks?
Which activation function is mentioned as commonly used in Convolutional Neural Networks?
What does the stride in a convolution operation determine?
What does the stride in a convolution operation determine?
What is the main purpose of pooling layers in CNNs?
What is the main purpose of pooling layers in CNNs?
How many fully connected layers are used for classification in the described CNN application?
How many fully connected layers are used for classification in the described CNN application?
What is the result of using several convolutions on the same image in a CNN?
What is the result of using several convolutions on the same image in a CNN?
Which operation is associated with max pooling in CNNs?
Which operation is associated with max pooling in CNNs?
What is a primary challenge that arises with an increasing number of hidden layers in deep neural networks?
What is a primary challenge that arises with an increasing number of hidden layers in deep neural networks?
Which of the following is a significant feature of Convolutional Neural Networks (CNN)?
Which of the following is a significant feature of Convolutional Neural Networks (CNN)?
Which loss function is used in the training of a neural network?
Which loss function is used in the training of a neural network?
What was a significant architectural feat achieved by Alexnet in the field of neural networks?
What was a significant architectural feat achieved by Alexnet in the field of neural networks?
What does backpropagation in neural network training primarily utilize?
What does backpropagation in neural network training primarily utilize?
What does the term 'tied weights' refer to in the context of CNNs?
What does the term 'tied weights' refer to in the context of CNNs?
Why is the learning rate significant in neural network training?
Why is the learning rate significant in neural network training?
What major change does CNN architecture implement compared to traditional neural networks?
What major change does CNN architecture implement compared to traditional neural networks?
What is a key step managers should take in relation to their brands?
What is a key step managers should take in relation to their brands?
Which method is considered to be more accurate in understanding brand image?
Which method is considered to be more accurate in understanding brand image?
What is the primary classification problem in the data mining task mentioned?
What is the primary classification problem in the data mining task mentioned?
What type of data is needed for training the model in the data mining task?
What type of data is needed for training the model in the data mining task?
How do consumers typically communicate about brands on social media?
How do consumers typically communicate about brands on social media?
What are the identified classes for analyzing consumption experience in the data mining task?
What are the identified classes for analyzing consumption experience in the data mining task?
What social media platform is specifically mentioned for data collection related to brand images?
What social media platform is specifically mentioned for data collection related to brand images?
What is one of the steps managers need to take regarding brand authors?
What is one of the steps managers need to take regarding brand authors?
What is the first step in hyper-parameter tuning?
What is the first step in hyper-parameter tuning?
What is a purpose of comparing the average probabilities of each class across all images?
What is a purpose of comparing the average probabilities of each class across all images?
What is K-Fold Cross Validation used for in hyper-parameter tuning?
What is K-Fold Cross Validation used for in hyper-parameter tuning?
What does Randomized Search CV optimize for?
What does Randomized Search CV optimize for?
What should be done with the best hyperparameter values after identification?
What should be done with the best hyperparameter values after identification?
Which of the following statements about hyperparameter tuning is false?
Which of the following statements about hyperparameter tuning is false?
What is the main advantage of using Grid Search CV?
What is the main advantage of using Grid Search CV?
Why might a validation set not be representative of the data?
Why might a validation set not be representative of the data?
What is the benefit of combining training and validation data in k-fold cross-validation?
What is the benefit of combining training and validation data in k-fold cross-validation?
What does obtaining the average probability of each class help to understand?
What does obtaining the average probability of each class help to understand?
Flashcards are hidden until you start studying
Study Notes
Convolutional Neural Networks
- Can be considered to have two components: Feature Mappings and Classifier Layers.
- Feature Mappings include Convolution, Activation functions (ReLU (max(0, 𝑥))), and Pooling.
- Classifier is made up of Fully connected layers (neurons) with some activation function.
- Convolution utilizes a filter/kernel that 'slides' along the input volume and transforms it.
- The same weights are used for the entire input volume.
- Stride is the step-size for the filter as it moves through the image (default value is 1).
- Padding can be used to utilize pixels less on the image perimeter.
Padding
- Uses 0's padding in the border.
- Performed especially in the final layers.
Pooling
- Transforms input volume by resizing it using operations such as min, max, avg.
- Max pooling takes the maximum of every block.
CNN Applications: Consumer Experience
- Uses 5 convolution layers and Max Pooling.
- Activation function: ReLU.
- Two fully connected layers for classification.
- Efficiently implemented using PyTorch library.
Brand Image and Consumer Experiences
- Brand image can be determined through customer surveys.
- More accurately, listen to consumer experiences on Instagram and Twitter.
- Consumers often tag brands in their posts.
- Consumers communicate about brands with each other using these images.
Image-Based Brand Image: Measuring Brand from Images
- Technique: Convolution Neural Networks.
- To know Brand Image: Simply conduct survey of the customers.
- More accurate: Listen to consumer experiences through social media, such as Instagram, Twitter etc.
Data Mining Task
- Classify the consumption experience of apparel and beverage brands.
- Classes: Glamor, Fun, Healthy, and Rugged.
- Consider this task as image classification: 4 Class Classification problem.
- Calculate average experience of all consumers in a given brand.
Data for Model Building
- Obtain tags from mTurk Annotators.
- Build an image classifier to predict the experience.
- Identify brand images by using brand hashtags in Instagram posts.
- Brands: Adidas, Nike, Gucci, Levi's, Coca-Cola, Pepsi, Fanta, etc.
Image Classification
- Can identify the consumer experiences in these images.
- The human visual system is incredibly good: the hallmark of intelligence.
- Difficult to make a rule-based computational system that recognizes experiences in the images.
Image Classification: ML Approach
- Use a large number of labeled images to train a classifier.
- Classify images as: Fun, Rugged, Glamor, Healthy.
Image Representation
- Each image is a matrix of pixel values.
- Greyscale images: single value per pixel (0.0 represents black, 255 represents white).
- RGB Image: Each pixel has three integers reflecting the contribution of red, green, blue color channels (0 to 255).
- Dataset: 227 pixels wide and 227 pixels tall, 227 x 227 x 3.
- Each image is represented by a single dimensional vector of RGB values.
Neural Networks
- Example: Multi-label classifier.
- Each image is a 154,587 vector.
Deep Neural Networks
- Deep Neural Network: Increasing number of hidden layers since the 1980s.
- Increasing the number of hidden layers results in more weight to learn, and optimization (Training) becomes harder.
Convolution Neural Networks (CNN)
- Utilizes spatial correlations in grid-like data (e.g. images).
- CNN reduces the number of weights through convolution and pooling operations.
CNN History: Alexnet
- CNN designed by Krizhevsky, Sutskever and Hinton.
- Won the ImageNet Large Scale Visual Recognition Challenge, 2012.
- Marks the rise of deep neural networks.
Image Net Challenge
- Other architectures have improved performance further.
Calculating Image-Based Brand Image
- Obtain average probability of each class across all images of a brand shared by users and sponsors.
- Compare the probabilities to understand the brand image among consumers and the brand image that the firm wishes to portray.
Hyper-Parameter Tuning
- Identify optimal value of parameters for the model.
- Create 3 splits of your data: Train(70%), Validation(10%) & Test Sets (20%).
- Use different hyperparameter values and train different models using Training Data.
- Evaluate performance of each model on Validation Set and choose the best hyper parameters.
- One validation set may not be representative of the data, combine training and validation data and partition them into "k" folds.
K-Fold Cross Validation
- Divide entire data into K folds.
- Train on training set and evaluate on validation set.
- Average performance over all the folds to estimate "true" test error.
Hyper-Parameter Tuning
- Train model on Training + Validation set to learn a new model.
- Grid SearchCV: Evaluates every possible combination of hyper-parameter values.
- Randomized search CV: Random samples of hyperparameter values are evaluated.
- Existing packages (sklearn) perform these efficiently.
References
- Churn management:
- Neural Networks:
- Hyper-Parameter Tuning:
Studying That Suits You
Use AI to generate personalized quizzes and flashcards to suit your learning preferences.
