Machine Learning Vocab I

Batch Size

# of the set of samples used in one iteration of training

Accuracy

% of Correct classifications during training

AND

0 0 0 1

XOR

0 1 1 0

OR

0 1 1 1

NOR

1 0 0 0

XNOR

1 0 0 1

NAND

1 1 1 0

Model

A mathematical representation or a program that can learn from data and find patterns to make predictions

Linear Regression

A method of finding the "line of best fit" of the data points based on a set equation denoted as y=mx+b where m=(x-avg(x))(y-avg(y))/(x-avg(x))^2 b=avg(y)-m*avg(x)

PoseNet

A model that converts human image or video into skeleton position Allows you to gather data points from skeleton position Also captures human position and movement over time

Neural Network

A structure of a Machine Learning computing system that is modeled after the human brain. Pre-programmed inputs and outputs and provided, that are then put through some calculations to determine the optimal neurons to convert inputs to their proper outputs. Generally has three portions: The input layer, hidden layers and output layer. Input layer is responsible for your input such as an image or some numbers. The hidden layers are used for transforming and forming generalizations with the data. The output layer gives you your output.

Generative

Algorithm that takes an input, and uses randomness or pseudo randomness to create an output. Example: ChatGPT, which can "generate" essays, practice problems, and responses from the user's input, and create said output.

Gradient Descent

An optimization model that is used to train machine learning models and neural networks. It works by trying to minimize the loss function.

Dataset

Collection of raw data

FaceMesh

FaceMesh takes a 192x192 input image of a face and outputs 468 3D keypoints. The architecture of FaceMesh is based on MobileNet FaceMesh was trained on 3D morphable model (3DMM) rendered images to annotate real images. (Didn't use a dataset labeled by humans).

ImageNet

Image classifier pre-trained on thousands of images, and is able to classify them with relatively consistent accuracy.

JSON

Lightweight file that stores objects and data in JavaScript. Commonly used in web development.

HandPose

ML model that allows for palm detection and hand-skeleton finger tracking in the browser. It can detect a maximum of one hand at a time and provides 21 3D hand key points that describe important locations on the palm and fingers.

FaceAPI

ML model to solve face detection, face recognition, and face landmark detection

YOLO

Object detecting algorithm that only applies the algorithm to the image once rather than multiple times in different areas of an image( thus making it much quicker!).

Normalize

Organizing data in a database by establishing tables and relationships between them according to rules to make the database more adaptable by removing redundancy and inconsistent dependency

Speech Commands

Recognizes a pre-trained set of 18 words, not the entire dictionary. It has the 10 digits, up, down, left, right, stop, go, yes, and no. You can train your own words with the teachable machine website and upload those to p5js, or find other pre-trained vocabularies online.

Training Set

Samples used to train the machine (subset of dataset)

Classification

The action of distinguishing a certain input to be a certain type of output. A classification model is a model that does the aforementioned action (also called a classifier)

Prediction

The output that a classification model gives about a certain input. This can be inaccurate at times.

Test Samples

Unused samples to test the accuracy of the machine

KNN/Kmeans

Way of clustering data into groups. You decide the number of clusters, k, you want to divide your data points into You choose k points randomly from your data set and make those your initial centroids You assign all the other data points to the centroid that they are located closest to Within each cluster you recalculate the centroid You repeat this process of redefining the centroid and reclustering for specified amount of iterations

Learning Rate

a floating-point number that tells the gradient descent algorithm how strongly to adjust weights and biases on each iteration

Epoch

a full training pass over the entire training set such that each example has been processed once. (Size of the Data Set divided by the Batch Size is how many batches per epoch)

Overfit

a machine that classifies poorly (captures training samples too closely & incorrectly classifies test samples) It will be less generous with what it considers to fit into the category. Example: If only trained on metal forks, the bot will have a hard time distinguishing a plastic fork (anything that deviates from its ordinary training set) but has a easy time distinguishing a metal fork (exact thing that the bot was trained on)

Underfit

a machine that classifies poorly (doesn't capture complexity of training samples). It will often ignore members of the category that should be recognized as matches. model is underfit when it classifies poorly because the model hasn't captured the complexity of the training samples Example: machine won't even recognize the fork that was used in training, and will struggle to distinguish a plastic fork.

Feature

an individual measurable property or characteristic of a phenomenon

Quick Draw

dataset created by users that played the online game, created and trained more than 50 million human labeled drawings.

Loss

evaluation of how well the machine has learned to predict correct classifications (certainty of correct choice high = low loss)

SketchRNN

takes a stroke from human then makes it into a drawing