7. Machine Learning Applications and Image Classification

It's Predicting how much you are willing to pay and then customize the price we offer you for a service or product.

• Important objects un user generated images may appear partially or distorted

• Images that contain brand elements may be presented as a combination of text and visuals rather than just visuals.

• Users image sharing may happen in different contexts

• Understand the consumers motives behind image sharing behavior is difficult if we rely only on image classification algorithm 

• Practitioners are recommended to complement deep learning techniques with behavioral research behavior methods to make the most of their image datasets

Convolution -> Pooling -> Fully connected layer / Classification

• Advertising 
• Branding
• Product Design 
• Product Presentation

Count midway between neurons in input and output layer.

(Logistic curve), send very negative close to 0 and very positive numbers close to 1

It makes sure the activation number is between 0 and 1

They inform us about how hight the weighted sum needs to be before the neuron becomes active.

They tell us which pattern the first neuron if the first hidden layer captures from the input layer.

• Weight
• Bias
• Activation function
• Sigmoid function

mathematical functions that transform input data in ways that allow the model to make predictions, classify data, or recognize pattern.

A class of machine learning models inspired by the structure of the function of a human brain. They are designed to recognize patterns and learn from data.

For some tasks, comparing to human-level performance can be a good benchmark. If a task is extremely difficult or even for humans, achieving new-human performance might be considered very good.

In some domains the cost of wrong prediction can be very high. In such cases, even a very high accuracy might not be sufficient if the consequences of errors are severe.

in many fields, there are benchmarks datasets with known "good" accuracy levels. Comparing against these benchmarks can provide context for your models performance.

More complex problems. like image or speech recognition, might have lower acceptable accuracies compared to simpler problems.

Compare your models accuracy against a baseline. 

(simple model-random classifier, or classifier that always predicts the most frequent class)

• Baseline Accuracy 
• Problem Complexity 
• Domain-Specific Benchmarks
• Cost of Misclassification
• Human-level Performance

• Statistical (linear) approximation 
• Machine learning approximation

• Logistic Regression 
• Naive Bayes 
• Decision Tree 
• Neural Networks

Training data necessary to identify factors and factor relationship

Prediction for new data based on learned structure. (train vs. test)

No training data, structure is derived directly from the data set.

Hihgly personalized ad: Show exactly the product consumer has previously been engaged with.

potential drawback: Overpersonalization

Targeting with limited personalization: More concrete re-targeting by reminding consumer of context of previous encounter by showing for instance, category-specific ad.

Targeting with no personalization: Present only brand in new ad to remind potential customer of previous brand engagement.

A substet of machine learning which uses deep artificial neural networks as models and automatically builds a hierarchy of data representations.

Brain inspired machine learning models.

Algorithm that allow computers to learn from examples without being explicitly programmed.

Any technique that enable machines to solve a task in a way like humans do.

1. Artificial Intelligence (AI)
2. Machine Learning (ML)
3. Artificial neural networks (ANN)
4. Deep Learning (DL)