Understanding how artificial neurons mimic biological neural networks
Dendrites receive signals from other neurons. The cell body integrates these signals. If the combined signal exceeds a threshold, the neuron fires an action potential through the axon to other neurons.
Inputs (x₁, x₂, ...) are weighted (w₁, w₂, ...) and summed. A bias (b) is added. The result passes through an activation function f(·) to produce the output y.
Integration: Both biological and artificial neurons combine multiple inputs into a single output.
Threshold: Both have a threshold mechanism - biological neurons fire when membrane potential exceeds threshold, artificial neurons activate when weighted sum exceeds bias.
Non-linearity: Both introduce non-linearity - biological through action potential generation, artificial through activation functions.
Activation functions are crucial components of artificial neurons. They serve multiple important purposes:
1. Introduce Non-linearity: Without activation functions, a neural network would just be a series of linear transformations, which can be represented by a single matrix multiplication. Activation functions allow networks to learn complex, non-linear patterns.
2. Control Output Range: Different activation functions bound the output to specific ranges (e.g., sigmoid outputs between 0 and 1, tanh between -1 and 1), which helps with numerical stability and interpretation.
3. Enable Learning: The gradients of activation functions are essential for backpropagation, the algorithm that allows neural networks to learn from data.
In the calculation above, we used the sigmoid function. Below, you can explore other common activation functions and their properties.