In this paper, we propose a novel unified neural computational unit, the APTx neuron, which integrates nonlinear activations and linear transformations into a single learnable representation. The APTx neuron is derived from the APTx activation function and does not require a separate activation layer, resulting in an architecture that is both computationally efficient and elegant. The proposed neuron has a functional form $y = \sum_{i=1}^{n} ((\alpha_i + \tanh(\beta_i x_i)) \cdot \gamma_i x_i) + \delta$, where all parameters $\alpha_i$, $\beta_i$, $\gamma_i$, and $\delta$ are learnable. We validate the APTx neuron-based architecture on the MNIST dataset, achieving a test accuracy up to 96.69% in only 20 epochs with about 332K learnable parameters. These results highlight the superior expressive power and computational efficiency of APTx neurons compared to conventional neurons, and suggest a new paradigm for integrated neuron design and architectures built on top of them.
