Begrepp Flashcards
Supervised learnning
In Supervised learning, we can only make a decision on how to update the internal parameters by knowing what our expected output should be, I.E “learning by example”.
Basically you need to have some training data to provide to the training examples
- Includes target outcomes
- Trained to recognize patterns(that lead to certain outcomes based on historical data(examples))
- (such examples: credit evaluation or interest rate prediction)
Unsupervised learning
In unsupervised learning, we don’t know what our output should be, except for some pre-programmed target such as “trying to win at tic tac toe”
It’s learning from data without any examples
- No known outcomes
- Learns to recognize patterns in the form of similarities
- Examples: Customer segmentation or marget basket analysis
Sample
Some kind of incoming data to be analysed
- example: a JPG
Feature
Some quantifiable data from the sample
- Example: Color, height, width, pixel data, etc
Labels
- Some useful information about the sample that we wish to categorize
- Example: looking at this picture, this is a person
Model
The output of some learning algorithm
- The parameterization of an algorithm that can be run against new data after it has been trained
Reinforcement learning
No specific target given, will instead explore many loops (solutions) to find the most optimal/best reward, based on feedback from the environment
- Based on agents, states, rewards, environment
- Similar to playing a game many times and learning from interaction
- Central/important to modern AI
- Examples: self-driving cars or game-playing computers(bottar)
How do you know when to stop training an algorithm? (in case of supervised learning)
- Seperate out some test data to assess the model accuracy
What is Accuracy?
- The number of correct classifications / Total number of test cases
The learning algorithm analyses the training data and produces a predictor function that can be used for mapping new examples to outputs.
What does this sentence actually mean?
This is a fundamental process in machine learning.
Learning Algorithm:
- This is a set of rules or procedures used by a machine learning model to learn from data. The algorithm processes the data and identifies patterns or relationships within it.
Analyses the Training Data:
- Training data is a dataset used to teach the model. The learning algorithm examines this data to understand its structure and the relationships between its elements. For instance, in a dataset of housing prices, the algorithm might analyze features like location, size, and number of bedrooms.
Produces a Predictor Function:
- Based on the analysis of the training data, the algorithm creates a function (or model). This function is designed to make predictions. For example, in the housing prices scenario, the predictor function would estimate the price of a house based on its features.
Mapping New Examples to Outputs:
- The term “mapping” refers to the process of taking new data (new examples that were not part of the training set) and applying the predictor function to this data to generate outputs (predictions). For instance, if you provide the characteristics of a new house to the trained model, it will use the predictor function to estimate the house’s price.
Summary:
- the sentence describes how a machine learning algorithm takes a set of training data, learns from it by identifying patterns, creates a model (predictor function) based on this learning, and then uses this model to make predictions about new, unseen data.
List some examples of Supervised machine learning approaches
- Linear regression
- Logistic regression
- Decision trees
- Nearest neighbor
- Support vector machine (SVM)
- Artificial neural networks (ANN): Most popular form of Machine learning approach
List some examples of Unsupervised machine learning approaches
- Apriori algorithm.
- K-means.
- Principle components analysis (PCA).
- Kohonen self-organizing maps (SOM).
What is Classification?
It’s a key part of AI, used for many different ai applications
- Look at some attributes about an object and decide how to label it (classify it)
- Deeply useful for making sense of big data
What is cluster analysis?
An analysis technique which groups objects by some selected attributes so that each object is similar to the other objects in a cluster different from objects in all other clusters
Used for:
- Classification
- Simplifying data
- Identifying relationships
Different Methods for cluster analysis:
- K-means
- K-nearest-neihhbor
- Mean-shift
- DBSCAN
Artificial Neural Networks
They are biologically inspired methods of computation:
- Based on models from neuroscience and cognitive psychology ideas
Similar to how we think the human nervous system works:
- Series of weighted connections between neurons (aka small processing units)
- Connection weights adjusted until desired output reached
- Basically a series of nonlinear equations
Explain what everything is in this image:
Xi
- The number of inputs
Wi
- The number of weights applied to each input
F(x)
- Is the activation function
We multiply the weights, to each input and sum this together, the sum is then compared to a certain value in the neruon(Threshold value)
If the sum is greather than this threshold, the output is 1. Otherwise it is 0(neuron doesn’t fire).
What is a XOR network
It’s what forms a neural network, aka multiple layers of connected perceptrons
What is artificial neural networks
- Multi-layer feedforward neural networks are the core model for most neural networks
- Backpropagation is one of the most popular supervised learning techniques. and is used in deep neural networks (aka very large multi-layer networks)
- Recursive and recurrent connectivity in neural nets and self-organising maps are other model sof neural networks.
What are the advantages and disadvantages of using artificial neural networks?
Advantages:
- Can learn complex patterns: The more layers, the more complex.
- Can handle non-linear data
- A series of non-linear regressions, can basically model any function
- Can handle redundant attributes, learns importance through weighting.
Disadvantages:
- Very susceptible to overfitting: The more layers, the higher the risk.
- Missing values must be imputed or the corresponding records removed
- Sensitive to noise
- Increased complexity of network requires more data and increases training time
- Traditional ANNs failed at high-dimensional input tasks such as image processing.
