As a product manager looking to develop AI products, it is important to understand the mechanics of ML. Model architecture, training data and model evaluation are all key elements of ML and are instrumental in successful AI product management. When using AI and products, there are numerous uncertainties that that can alter the course of development and by understand the core concepts, we can stay ahead of the unknowns and plan development accordingly.

Key Concepts:

Overview of Modelling:

Modelling in ML consists of constructing the core components that will serve as the framework for machine learning model. Neural networks are some of the most common models used in ML.

https://s3-us-west-2.amazonaws.com/secure.notion-static.com/f222a43c-3e61-42a6-ba53-06f4d37b478d/Untitled.png

A single neural network is made up of numerous computational layers. Each of these layers will pass a decision to the subsequent layer until reaching the output layer, which will produce a final decision for the input data. The data that goes into the network is a set of values. These values, for example, could be pixel contained in an image or audio waveform.  The above image shows an example of a network with three layers, two hidden layers and an output layer. When counting the number of layers, we generally count all the layers after the input layer. Each layer is made up of some number of nodes, each of which has its own parameters that describe the data flow through the network.

Neural networks are sometimes referred to as deep neural networks, simply because of the deep number of layers contained in the network. Before a network is trained, there is an architecture that is defined and used throughout the life of the model. There’s a process known as neural architecture search that further optimises the creation of a neural network by swapping out and testing different architectures as the model is trained.

In Summary: Neural network is a series of layers, each of which contains different nodes which will perform various calculations. The structure of these nodes and layers as well as how they’re connected is what is knows as the architecture of the network.

Why Neural Networks?

Neural networks are made up of neurons, sometimes called nodes. One neuron is responsible for processing some input data and producing an output. This is modeled after the neurons in our brain, which process input signals and produce output signals. For a neural network that processes visual data, such as a set of images, these inputs will be either spatial information or color information. When these color components or shapes are combined, which happens inside a neuron in the form of an equation (ex. 0.5red and 0.5blue = 1*purple), it produces an output signal that can do something like help classify the initial input!

This explanation and the below gif, were taken from Cezanne Camacho's blog post on neural networks