A neural network takes an input vector, such as pixel values or a window of accelerometer samples, and passes it through layers of weighted sums and nonlinear functions until it produces an output, for example a class label or a predicted value. Training adjusts the weights to reduce error on labeled examples. Networks with many layers form the basis of deep learning.
For IoT the practical questions are size and speed. Weights have to fit in device memory and each inference must finish within the sampling interval. Techniques such as quantization shrink the network so it can run on a microcontroller or a small NPU. A small convolutional network that recognizes a few spoken commands is a typical embedded example.
Part of the IoT Glossary. Spotted an error or a missing term? Tell us.