ML

Turing Test

A test based on a process where series of judges attempt to discern interactions between a human from a machine. Test if a computer can think like a human KAPTCHA

Data Labeling

Process of detecting and tagging data samples -part of preprocessing -useful for supervised learning -Errors in data labeling can impact the model -HITL (human in the loop) humans do it

Chat bot

A chatbot is deep learning, ML, and AI software that can simulate a conversation with a user through websites, apps, telephone. Simulate like a human. natural evolution as NLP chatbot needs to identify user's intent and extract data and relevant entities in the user's request chatbot returns appropriate response through NLP

reinforcement learning

A form of learning based on the receipt of rewards. The agent will be punished for bad actions and rewarded for good actions. The agent will be learned incrementally

Artificial Neural Networks (ANNs)

A neural network is a network of neurons (artificial neural network) composed of artificial neurons. Mimics the brain's neurons using nodes, connections, and weights to find output from a specific input a neuron processes the input and passes it to the next input. learn on their own. generate characteristics from what they process. uses pixels weights need to be taught by doing backpropagation

Human Pose Estimation

Computer vision-human pose estimation is the task of determining the position and orientation of the human's body in an image CNN-based regression problem can predict pose despite hidden through estimation

NLP

Natural Language Processing is the subfield of artificial intelligence with the aspect of linguistics for computers. read decipher and understand human language 1-NLP apply algorithms to identify and extract natural language rules that unstructured language data is converted to a form computers can understand 2-computer uses algorithms to extract meaning from every sentence and collect important data

RNN

Recurrent Neural Network Class of artificial neural networks specialized for processing sequential information -Main perk is ability to memorize results from previous computations for current computations -useful for NLP from the state becoming a representation of all the words processed through the RNN -knows order of words

Classification

Type of supervised learning Assign a label to data (categorical data) -linear classifiers -support vector machines -kernel estimation -decision trees -etc

What is data science

a study to discover hidden patterns from raw data. maybe predict

computer vision (cv)

subfield of AI that trains computers to interpret and understand the visual world. includes methods for acquiring, processing, analyzing digital images unsolved problem +neuroscience understand content of visual images object detection: where are the objects in the pic object classification: what category is the object object segmentation: what pixels belong in the object object identification: which type of object is in this photograph object verification: is the object in the pic object landmark detection: key points for the object in pic object recognition: what object are in the photo and where are they

Image segmentation

subfield of computer vision used to locate objects and boundaries mostly used to partition an image into regions to find some object -divide image into segments or groups with similar attributes -create a pixel-wise mask for objects -has exact pixel-to-pixel detection

Machine Learning

subset of AI it is a set of algorithms to learn from data without being explicitly uses statistical methods to improve using large amounts of data compares new instances to previous experiences learn through supervised, unsupervised, or semi supervised

Supervised Learning

Learns a function that maps an input to an output based on example input output pairs used for regression and classification

Transfer Learning

Machine Learning method where a model can be reused as a starting point for another task. -first train a base network on a base dataset and task -repurpose the learned feature and transfer to the second target network to be trained -useful for NLP and CV

semi-supervised learning

Uses both supervised and unsupervised learning think of it like an exam seen material before, but exam you dont have answers.

Clustering

instance of unsupervised learning group a set of objects (cluster) that are more similar to each other. Before grouping need to find similar examples and can measure the similarity by combining the example's feature data into a similarity measure unlabeled data -kmeans -mean shift

Logistic regression

machine learning technique used for binary classification problems binary logistic regression -> 2 possible outcomes spam or not spam multinomial logistic regression-> three or more categories without ordering ordinal logistic regression-> categories with ordering logistic function (sigmoid function)

Self-Supervised Learning

train a model using labels that are naturally part of the input data rather than separate external labels -labels automatically labeled, labels do not need to be done by a human -learns without using explicitly provided labels -not supervised learning because the model is not learning the inherent structure of the data -not centered around clustering etc... can predict through natural evolution and consequence of its actions kind of like a baby learning from unlabeled data with learned labels. label is within the data?

Regression

type of supervised learning estimate relationship between dependent and independent variable find which feature is important or not important continuous data **correlations -linear regression -polynomial regression -elastic net regression -ridge regression -lasso regression

YOLO

you only look once object detector algorithm only needs to image to pass one time through the network uses a neural network on the entire image to predict multiple boxes each containing a specific object simultaneously -image is divided into a X x X -if the center of an object is in one of the regions then detect the object -Each of the cells in the boxes are responsible for scoring the level o confidence of the object as well the type of object output will have labels attached to it. faster, but decreased accuracy in other algorithms

Energy-Based Learning

-use case: dont want to spend a lot of time and effort training (deep learning) and causing the same mistakes over and over try to match human based learning -A system that predicts whats going to happen exactly this deals with the unsureness of the future -energy based function that takes low values on the data manifold and higher values everywhere else ***What is the y most compatible with the X? -Can use Gan architecture for objective functions with an algorithm so that the volume of low energy is constant . then use encoders to train and limit the volume of space with low energy lower the probability to guess that

Genetic Algorithms

A search heuristic that mimics the process of natural evolution. This heuristic is routinely used to generate useful solutions to optimization and search problems. genetic algorithms is the simplest random-based algorithms for optimizing hyperparameters and datasets

Back propagation algorithm

A supervised learning algorithm for training a deep learning model. Most popular because it uses powerful calculation methods with low computation time Change the weights to achieve expected output. Does this by calculating the error. Gets the error and goes backward to the model and adjust all the weights in proportion to the error. Another example fed and repeat until the error is minimized A training algorithm for neural networks that passes error estimates from the output layer back to earlier neurons so they can adjust connection weights

Unsupervised learning

A type of learning to help find unknown patterns in a dataset without any labels two main method to unsupervised learning is -principal component analysis and cluster analysis cluster analysis - groups datasets with shared attributes

object detection

AI/ML and deep learning and computer vision locate instances of objects in pic or videos approximately identify what is in the image or vid most popular method is 2-staged neural network technique (R-CNN) -first identifies the region proposals or a part of the image that might contain an object -classify the object in the region another method is single-stage network -faster than the first method YOLO (you only look once) produces prediction for region across the entire image with anchor boxes

Artificial general intelligence vs AI

Artificial general intelligence that can learn to do different things multi-purpose. can learn, plan, reason, communicate in natural language today AI can only do a single tasks AGI doesnt exist today yet. What people are imagining

CNN

Convolutional Neural Network-gradient based learning Used often A NN for processing data like a grid? able to detect patterns and understand it reduce to more dimensions first layer is convolution layer-convolves the image using filters to detect edges these filters are taught to network during training activation layer-non linear activation function there is no learning here -non linear transformation over the input signal the transformed output is sent to next layer neurons as input and help decide if the neuron will fire or not to next layer pooling layer-reduces the number of parameters and computation in the network and control size of network max-pooling only take out the maximum from the pool of data 3 layers can be repeated multiple times called conv layers fully connected layer that learns a non-linear function from the pooling layer output flattens 3D to 1D then use vector for image classification etc,,

Deep Learning

Deep structured learning or hierarchical learning is a subset of artificial learning and machine learning -neural network has more layers than any other ML subset (deep) -need a lot of data -connect a lot of neurons and assign weights to each features -can use supervised, unsupervised, semi supervised, or reinforcement learning

Data Mining

Extract useful information from a lot of data to discover new, accurate, and useful patterns. Data mining is designed to extract rules from large quantities from data, while Machine learning teaches a computer how to learn given the parameters Data mining to determine particular outcome from total of data gathered ML perform tasks and use data to become smarter Data mining works with data, ML works with algorithms

How to prevent overfitting and underfitting

Find a good balance between variance and bias with cross validation: use mini train and test on folds for the data during training -more training or early stopping before overfitting happens

GANs

Generative Adversarial networks uses deep learning methods like CNN Unsupervised learning task in machine learning to automatically discover and learn the regularities or patterns in input data to generate new examples that can possibly been drawn from the original dataset -generator model used to generate new examples -discriminator model to classify examples as real or fake the two models trained until discriminator model is fooled about half the time. discriminator is rewarded or not change needed to model parameters, but generator is penalized and updates model parameters generator is rewarded if fools, but discriminator is penalized and changes model parameters high res photo is one example

Transformers

RNN used in NLP, but very slow and can't deal with long sequences and could only work in one word at a time -transformer replace RNN and CNN to attention -transformers would make use of parallelization like RNN during training to be faster and more accurate -attention key idea for transformer is the evaluation of the relationship between each input item and each output item -self attention is the evluation of the relationship between each input to other inputs tells how relevant the word is to other words in the same sentence -transformer networks use self-attention multiple time called multi headed attention and each head learns attention relationships independently -transformers are sequence to sequence model that uses encoders and decoders -encoder checks the degree of relevance of each word to other words in the sentence to be translated -decoder predicts the next word and executed multiple times -transformers can learn the sentences sumultaneously instead of RNN's one word at a time -faster training time than RNN -attention ignores order better for distant relationships -requires lots of data+memory+is new

Artificial Intelligence

Subfield of cs -intelligence demonstrated by machines in contrast to human intellect -describe machines that mimic learning