ML Flashcards
Artificial Intelligence
Program that can sense, reason, act & adapt
Machine Learning
Algorithms where performance improve as they re exposed to more data
Deep Learning
Multilayered neural networks learn from lots of data
Types of analytics in order of complexity
Descriptive
Predictive
Prescriptive
Ability to process large data because of ________. Using:
1.
2.
3.
Infrastructure
- CLoud Services
- GPU: Handles graphic rendering tasks
- TPU: Gives high performance & pure efficiency when running tensor flow
Types of ML (3 types)
Supervised - training data
Unsupervised - no training data
Reinforcement - interaction, +/-ve feedback
Summary of what to use:
Continuous = …
Discrete = …
…Accuracy
…Confusion Matrix
Support Vector Machines use
Discriminative classifier
Clear margin of separation between categories
Used in small clean data sets
Don’t suffer from over fitting as much as other methods
When should you not use Support Vector Machines
large data sets because required training time is higher in noisy data sets
Decision trees
Used for regression lasts - continous variable decision trees
Used for classification - Categorical variable decision trees
Why use them?
→ involves stratifying or segmenting the predictor space spare into # of regions
→ good for non-linear data.
Random Forests
Belong in the general category of Ensemble Methods
Increases predictive accuracy but sometimes at the expense of explainability
Ways to increase accuracy
Bagging:
Learns from each other independently in parallel & combines them for determining model average
Boosting:
Learns sequentially & adaptively to increase model predictions of learning algorithm
Pruning
Decreases size of decision trees by removing sections of the tree that are non-critical & redundant to classify instances.
Decreases of the final classifier .’. increase accuracy by decrease of overfitting
Why use Unsupervised Learning
Easier to obtain unlabelled data
Takes place in real time
Decrease complexity in comparison to supervised
finds all kinds of unknown patterns in data
K-means cluster analysis
Help with data-driven insights
Deep domain knowledge in required
No right or wrong matter for interpretation
Isolation Forest Algorithm
Basic idea is to overfit decision tree models
- Grows in random decisions see model until exact instance is in its own leaf
- Model is forced to keep splitting
Dimensionality reduction
Decrease features to easily understandable dimensions
Helps project data in dimensions the brain can understand
Deep Learning
Can be supervised or unsupervised
Uses multilayered neural networks called DNN to simulate complex decision making power of human brain
Convolutional neural nets
What’s the difference between loss function and optimisation?
Loss function: Measures how far a model’s predictions are from the actual values. Lower loss indicates a better fit.
Optimization: The process of adjusting a model to minimize the loss function and improve its performance.
Analogy: Imagine training a dog to fetch. The loss function is like the distance between the ball and where the dog drops it. Optimization is the process of training the dog to minimize that distance (bring the ball back closer each time).
Backpropagation
A method in deep learning to adjust the internal parameters of a neural network. It calculates the error between predictions and actual values, then propagates it backwards through the network to update weights and biases, making the model learn from its mistakes.
Analogy: Imagine a team game where players need to improve their coordination. Backpropagation is like a coach reviewing plays, identifying where things went wrong (high error), and guiding each player (weights and biases) on how to adjust for the next round.
Transfer Learning
Reusing a pre-trained neural network model on a new task. The pre-trained model (usually trained on a massive dataset) acts as a starting point, with its weights fine-tuned for the specific needs of the new task. This saves time and data compared to training a new model from scratch.
Analogy: Learning to ride a bike. Once you know the basics (balance, pedaling), you can transfer that knowledge to ride a motorcycle (similar concept, new controls). Transfer learning leverages pre-existing knowledge (pre-trained model) for a new task (riding a motorcycle).
Pros and Cons of Deep Learning
Pros:
High Accuracy: Can achieve state-of-the-art performance on complex tasks like image recognition, natural language processing, and speech recognition.
Feature Learning: Discovers patterns and features from raw data automatically, eliminating feature engineering.
Scalability: Handles large and complex datasets effectively.
Cons:
High Computational Cost: Requires powerful GPUs and large amounts of data for training, making it resource-intensive.
Black Box Problem: Deep models can be complex and difficult to interpret, making it challenging to understand their decision-making process.
Overfitting: Prone to overfitting if not trained properly, leading to poor performance on unseen data.
Stemming
Popular way to decrease size of vocab in Neural language tasks by conflating words with related meanings
Aims to convert words with same stem/root to single word types
