5

Question 1

What is the intuition behind L2 loss function?

Answer 1

Measures the squared difference between the predicted value and the actual value, encouraging the model to minimize large errors.

Question 2

What is the intuition behind L1 loss function?

Answer 2

Measures the absolute difference between the predicted and actual values. This makes it more robust to outliers because it does not square the differences, which would amplify large errors.

Question 3

What is the intuition behind Smooth L1 loss function?

Answer 3

Combination of L1 and L2 loss, and it behaves differently depending on the size of the error. When the error is small, it behaves like L2 loss (squared differences), and when the error is large, it behaves like L1 loss (absolute differences). This makes it more robust to outliers than L2 while not being as harsh as L1 for small errors.

Question 4

What is the intuition behind Negative Log Likelihood loss function?

Answer 4

Calculates the likelihood of the predicted class being correct. For a given class, if the model is confident and correct, the loss will be small. If the model is wrong, the loss increases sharply.

Question 5

What is the intuition behind Cross Entropy loss function?

Answer 5

Compares the predicted probability distribution of classes with the actual distribution (one-hot encoded labels). It is particularly effective when used with softmax activations.

Question 6

What is the intuition behind Binary Cross-Entropy loss function?

Answer 6

Measures the difference between two probability distributions, one being the predicted probability for class 1 and the actual label for the class

Question 7

What is the intuition behind Cosine Embeddings Loss function?

Answer 7

Calculates the cosine of the angle between the predicted and actual vectors. If they are similar (i.e., their angle is small), the cosine similarity will be close to 1, and the loss will be small.

Question 8

Why do we need to augment the data?

Answer 8

Improving the performance of machine learning models by increasing data diversity, improving generalization, preventing overfitting, and making models more robust to variations in real-world data.

Question 8

“Black box” issues

Answer 8

Lack of Interpretability (not transparent processes)
Visualizing Filters
Visualizing Activations
Maximally Activating Patches
Occlusion Experiments (masking part of the picture)
Saliency Maps

Question 9

Advantages and disadvantages of Deep Learning network

Answer 9

Advantages: easy to implement, tailored approach, performance boost, can help with overfitting

Disadvantages: time consuming, trial and error, relational gaps, can lead to overfitting