tanh is a hyperbolic activation function that converts a numeric input into a value between -1 and 1. In software engineering, it is commonly used in machine learning to introduce controlled nonlinearity into neural networks. Its zero-centered output can help models represent both positive and negative signals, although its gradients become very small for extreme inputs.
tanh
What is tanh in software and machine learning?
tanh is the hyperbolic tangent function, an activation function that converts any numeric input into an output between -1 and 1. In machine learning, it adds nonlinearity so neural networks can learn patterns that linear calculations cannot represent.
In code, developers commonly apply tanh to each value produced by a neural-network layer. Its zero-centered outputs let models pass both negative and positive signals forward.
tanh formula, range, and key properties
The formula is tanh(x) = (e^x - e^-x) / (e^x + e^-x), also written as sinh(x) / cosh(x). It accepts every real number, returns a value strictly between -1 and 1, and tanh(0) = 0. It is an odd function: tanh(-x) = -tanh(x).
As x becomes very positive, tanh approaches 1; as x becomes very negative, it approaches -1. Its derivative is 1 - tanh(x)^2, which training systems use to calculate weight changes.
How tanh works as an activation function
tanh smoothly limits a layer's calculated value.
- A neuron receives a weighted input, such as 1.
- The network applies tanh: tanh(1) is approximately 0.76.
- The bounded output passes to the next layer as a signed signal.
- During training, the derivative measures responsiveness to the input.
- The optimizer updates weights to reduce prediction error.
tanh vs. tan vs. sigmoid
Despite similar names, tanh and tan solve different problems. tanh is not arctan, the inverse tangent function.
| Function | Output range | Behavior and common use |
|---|---|---|
| tanh | -1 to 1 | Zero-centered, smooth, and saturates near its limits. Used in some hidden and recurrent neural-network layers. |
| tan | Unbounded | Has repeating discontinuities. Used in mathematical calculations, not typical neural-network activation layers. |
| sigmoid | 0 to 1 | Not zero-centered. Often used for probabilities and binary-output layers. |
Benefits of tanh
tanh is useful when a model needs controlled signed values.
- Bounded output keeps hidden signals in a predictable range.
- Zero-centered outputs can make optimization more balanced than strictly positive activations.
- It is smooth and differentiable, supporting gradient-based training.
- It is commonly suitable for recurrent architectures with controlled hidden states.
Limitations of tanh
Its useful range creates a trade-off.
- Large positive or negative inputs saturate near 1 or -1.
- In saturated regions, very small gradients can slow learning in deep networks.
- ReLU-family activations are often preferred for many deep hidden layers.
- Match the activation to the task and output requirements, not only its bounded range.
Using tanh in AI application development
Use framework-provided tanh implementations instead of hand-coding the exponential formula, which can be numerically unstable for extreme inputs. Test training curves, output distributions, and validation results before keeping an activation choice. Teams building AI features through app design and development can use Emergent to build, test, and ship application workflows while exploring the AI agent builder landscape for suitable approaches.
Start Building
on Emergent today
on Emergent today
Start Building
