Basics Flashcards
Difference deep Learning and traditional CV approaches
Deep Learning uses an end-to-end approach, which allows the neural network to learn by itself, using the training data to improve the model, while traditional computer vision relies on handcrafted features
What the difference between supervised and unsupervised learning?
Supervised learning uses labeled data to train the model, while unsupervised learning uses unlabeled data to find patterns and structures within the data
Difference between classification and regression
Classification is used to predict a class or a label for the input data. Regression is used to predict a numeric value for the input data
Explain Underfitting and Overfitting. Name the causes.
Underfitting is when the model is not complex enough to capture the distributions of the data. The model has a bad generalization for not fitting the distribution well, i.e predicts very badly the result for the input data. That happens when the model is not complex enough or needs more training.
Overfitting is when the model is too complex and is fitted to well from the training data, making it difficult to predict the results for unseen data. This can be caused by a too complex model or missing of regularization.
Why are non-linear activation functions used in NN?
Non-linear activation functions are used in neural networks to introduce non-linearity, enabling the model to learn complex and non-linear relationships in the data
2 advantages of ReLu activation functions over Sigmoid and Tahn
1 - not vanishing gradient.
2 - It is computationally effective as it involves simpler mathematical operations than sigmoid and tanh.
Write the formula for precision, recall and F1-score and name on task for which those metrics can be used
Precision = True Positives / (True Positives + False Positives)
Recall = True Positives / (True Positives + False Negatives)
F1-score = 2 * (Precision * Recall) / (Precision + Recall)
Those metrics can be used in image classification.
Update formula for stochastic gradient descendent
θ <- θ − η⋅∇_θL(θ;x(i);y(i))
Formula leaky ReLu. What param alpha is normal?.Alpha = 1, reasonable?
f(x)=max(0.01*x , x).
Normal aplha = 0,01. Alpha=1 is not reasonable, it would become an Identity function.
Which loss function for classification commonly used together with Softmax normalization? Name, formula and define variables
Negative Log-Likelihood.
(Look the formula online)
N: Number of samples in the dataset.
y_i : Ground-truth class label for sample i.
p iy i : Predicted probability of the true class for sample
What is the triplet loss. Write its formula.
The goal is to minimize the distance between the ancor image to the positive mach, while maximize the distance between the ancor to the negative math
What is data augmentation and give 2 examples
Is a technique to use the dataset to transform the images and then have new data.
1 - Horizontal reflexion
2 - Random crops on the image
3 - Scale
4 - flip the image
Key idea behind data augmentation?
Increase the database, by applying random transforming to existing data
Name a reason why it can make sense to apply image augmentations/transformations at test time
Applying augmentations at test time improves model robustness and generalization to handle diverse real-world scenarios
What is dropout and what is it used for? Which hyperparam does a dropout layer have
Dropout is a regularization technique that randomly deactivates neurons during training to prevent overfitting in neural networks.
The hyperparameter of a dropout layer is the dropout rate, representing the probability of setting a neuron’s output to zero during training.
