| Contents | A correspondence between tropical rational functions, which are naturally associated with polytopes, and integer-valued neural networks (IVNNs) with $\text{ReLU}_ t$ activations was introduced by L. Zhang et al. in their paper ``Tropical Geometry of Deep Neural Networks'' . Here, an IVNN refers to a neural network with integer weights and real biases, while $\text{ReLU}_ t$ is defined by $\text{ReLU}_ t(x)=\max(x,t)$ for $t\in\mathbb{R}\cup\{-\infty\}$. As a consequence of this connection, the linear regions of an IVNN can be related to vertices on the upper faces of the polytope associated with the corresponding tropical polynomial, providing a tropical geometric derivation of a preexisting upper bound on the number of linear regions of a neural network. In recent work, we explored an additional connection with order theory, using posets and their associated order polytopes to construct structured tropical polynomials and derive corresponding pooling filters for convolutional neural networks. Beyond these connections, type decompositions of tropical polynomials, which organize terms according to their combinatorial type and frequency structure, provide another natural direction for investigation. Understanding how these structures interact with tropical representations of neural networks may offer new perspectives on the structure and theoretical analysis of neural network models. |
