LESSON 11 - Unsupervised learning foundation

Question 1

What is one characteristic of unsupervised learning that distinguishes it from supervised learning?

Answer 1

Unsupervised learning does not require explicit labels, and it focuses on discovering statistical regularities and patterns in the environment without constant feedback.

Question 2

How is reinforcement learning different from supervised learning in terms of the feedback provided?

Answer 2

In reinforcement learning, the feedback is not in the form of correct or wrong but is more like +1 or -1, related to the outcomes of the actions taken.

Question 3

What is the main advantage of unsupervised learning, particularly in industrial settings?

Answer 3

The absence of the need for labels in unsupervised learning is advantageous, making it useful for tasks like anomaly detection in industrial settings.

Question 4

What is the causality issue mentioned in unsupervised learning, and how is it addressed?

Answer 4

The causality issue refers to the need for active observation rather than passive observation. Transfer learning is suggested as a solution, where unsupervised learning is followed by supervised learning.

Question 5

How does unsupervised learning contribute to transfer learning in a sequential manner?

Answer 5

Unsupervised learning is initially used to extract features and then followed by supervised learning, making it easier to perform tasks accurately when new examples are introduced.

Question 6

In the context of unsupervised learning, what role does feature extraction play in preparing for supervised learning?

Answer 6

Unsupervised learning involves describing objects and extracting features, which is useful for supervised learning where decisions about hyperplanes or lines are made for classification tasks.

Question 7

How does hierarchical processing in neural networks contribute to making classification problems linearly separable?

Answer 7

Hierarchical processing in neural networks, as seen in the example of car and plane manifolds, enables linear separability by transforming initially indistinguishable objects into distinguishable ones as the visual path progresses.

Question 8

What is the primary goal of learning representations in neural networks, and how is data compressed in the process?

Answer 8

The main goal of learning representations is to extract essential features. Data is compressed by removing unnecessary features, such as correlated ones, reducing the number of features while approximately capturing the relationships.

Question 9

How is Principal Component Analysis (PCA) used to find directions of maximum variability in a dataset?

Answer 9

PCA is a statistical technique that identifies directions of maximum variability in a dataset. It finds the first principal component, representing the direction with the highest variance, and subsequent components in decreasing order of variance.

Question 10

What is the purpose of decomposing and compressing a matrix in PCA, and what are the vectors representing principal components used for?

Answer 10

Decomposing and compressing a matrix in PCA is done to capture many variabilities while maintaining a representation similar to the original. The vectors representing principal components are used to reconstruct the data with fewer dimensions.

Question 11

In unsupervised learning, how does reinforcement learning differ from other approaches in terms of interaction with the environment?

Answer 11

Reinforcement learning involves active interaction with the environment, where actions are taken and feedback is received, marked by values like +1 or -1, rather than traditional correct or wrong feedback.

Question 12

What is the key difference between supervised learning and unsupervised learning when it comes to the availability of external teaching signals?

Answer 12

In supervised learning, there is a constant external teaching signal available, while in unsupervised learning, the agent tries to discover statistical regularities without explicit lab

Question 13

How does transfer learning leverage unsupervised learning in the context of feature extraction?

Answer 13

Transfer learning starts with unsupervised learning, extracting features and preparing the model. Later, supervised learning is applied to build upon the extracted features for specific tasks.

Question 14

Explain the role of clustering and feature extraction in unsupervised learning.

Answer 14

Clustering and feature extraction are techniques used in unsupervised learning to identify prototypes and reduce the number of features, respectively, without relying on explicit labels.

Question 15

What is the significance of hierarchical processing in neural networks, and how does it contribute to linear separability?

Answer 15

Hierarchical processing transforms initially indistinguishable objects into distinguishable ones, contributing to linear separability in neural networks, particularly in the context of visual paths and object recognition.

Question 16

How does the removal of unnecessary features contribute to the compression of data in neural networks?

Answer 16

The removal of unnecessary features, such as correlated or constant ones, reduces the number of dimensions in the data, effectively compressing it while capturing essential relationships.

Question 17

What does Principal Component Analysis (PCA) aim to achieve in a given dataset, and how are principal components represented?

Answer 17

PCA aims to find directions of maximum variability in a dataset. Principal components are vectors representing these directions, ordered by the amount of variance they capture.

Question 18

In the context of PCA, what is the role of orthogonal components when compressing images?

Answer 18

Orthogonal components are used to compress images in PCA. By representing images in terms of these orthogonal components, the original data can be reconstructed with reduced dimensions.

Question 19

How is unsupervised learning utilized in convolutional networks to map and compress images?

Answer 19

Convolutional networks map every pixel to a data sample, compressing images by projecting them into a lower-dimensional space using hidden layers, ultimately aiming for effective decompression.

Question 20

Explain the concept of self-supervised learning and its application in neural networks.

Answer 20

Self-supervised learning involves training a neural network to predict the next element or fill in omissions in data. It does not rely on traditional labels, making it a form of unsupervised learning with a focus on data prediction.

Question 21

How does unsupervised learning contribute to the analysis of psychological studies, as exemplified in the Big Five personality traits?

Answer 21

Unsupervised learning, specifically the extraction of features, can be applied to psychological studies like the Big Five, summarizing different aspects of personality by approximating them with a reduced set of main features.

Question 22

What distinguishes convolutional networks from other unsupervised learning models in terms of mapping pixels to data samples?

Answer 22

Convolutional networks map pixels to data samples and compress them using hidden layers, presenting an approach where images are projected into a lower-dimensional space.

Question 23

How is the gradual change in latent code utilized in generating smooth transitions between digits or faces?

Answer 23

The gradual change in latent code allows for smooth transitions between digits or faces, enabling the generation of images with subtle changes, providing a more continuous and natural appearance.

Question 24

What does the decrease in the loss function during training signify, and how can it be helpful in assessing the learning progress?

Answer 24

The decrease in the loss function during training indicates that the model is learning and fitting the data better. It can be helpful in assessing learning progress, with a lower loss indicating improved performance.

Question 25

In the context of neural networks, how is overfitting identified, and what measures can be taken to address it?

Answer 25

Overfitting is identified when there is a significant difference between training and test phase performance. Measures to address overfitting may include introducing a validation set or adjusting regularization techniques.