Machine Learning Concepts

Overview

Machine learning, and more specifically a subclass known as Supervised Learning, refers to a family of algorithms where a computer learns to recognize patterns in data by transforming sample data points, known as training data, into predicted values or outcomes. They are called learning algorithms because the algorithm learns its rules for transforming input data into outcomes by reducing the error between the known results and predicted values. A well-trained machine learning algorithm, also known as a model, is a powerful data processing technique that can accurately predict outcomes given new data points never seen before by the model.

Machine learning concepts

• Machine learning algorithms can be as simple as linear regression, where a line of best-fit is calculated based on numerical input data plotted against known outcomes, or as sophisticated as neural networks, a mathematical abstraction of biological neurons that have been successful in advancing the capabilities of artificial intelligence based on their ability to solve complex problems such as speech and visual pattern recognition. Deep neural networks, associated with a subclass of machine learning algorithms known as Deep Learning , take the neural network concept to its natural conclusion by layering 10s or even 100s of neural network component layers on top of each other. Recently, a specialized type of deep network known as a Deep Convolutional Neural Network has surpassed human-level performance on some image recognition tasks. These networks, which were inspired by the structure of a feline visual cortex, are composed of many layers of optical feature filters. These filters are learned from raw image data, essentially transposing low-level pixel values into high-level representations.

• Regardless of the complexity of a supervised machine learning algorithm, the key to a well-trained model relies on a set of training data that is sufficiently large and representative of the input and outcomes in which the model will be used. In the context of computer vision, if the end goal of a model is to accurately recognize if an image contains a cat, a good training data set will have a large number of images of cats of various sizes, colors, and shapes, as well as a sufficient amount of negative examples so the model can learn to both recognize any cat and recognize when a cat is not present.

• While a training data set is a collection of data used to build the machine learning model, a validation data set is data that has never been processed by the learning algorithm and is used to verify that the trained model is accurate. In machine learning, there is a risk of “teaching for the test” or over-tuning model features to fit the training data set best. To avoid the problem of so-called overfitting, validation data are used to verify the accuracy after the model is trained.

• In the Cogniac system, the process of uploading media, capturing to an application, and providing model detection feedback generates continuously growing sets of image data with each media item associated with ground-truth outcomes represented by subjects. In the Cogniac system, new media data are automatically assigned to either a training or validation set, and each model’s accuracy against the validation set is reported through application performance metrics. Users can quickly and easily create training and validation data sets with this feedback loop while simultaneously training deep-learning models to solve their visual automation tasks.