Deep learning owes its massive success to the mathematical smoothness of backpropagation. If you want to optimize a network's weights, you calculate a gradient and take a step down a continuous loss landscape. But what happens when you want to optimize the structure of the network or the hardware it runs on?
Consider choosing the number of layers in a model, the integer size of a convolution kernel, or the physical memory buffer size of an accelerator chip. These parameters are fundamentally discrete. You cannot have a layer count of 3.4, or a hardware tiling strategy of 5.1. As noted by Hanxiao Liu et al. in their foundational work on DARTS: Differentiable Architecture Search, this non-differentiability creates a massive barrier, forcing researchers to abandon elegant gradient descent in favor of slow, black-box search loops.
The problem deepens when we look at hardware-software co-design. As Yonggan Fu et al. point out in Auto-NBA: Efficient and Effective Search Over the Joint Space of Networks, Bitwidths, and Accelerators, hardware parameters like data tiling strategies and memory allocations lack analytical gradients entirely. This creates a "chicken-and-egg" optimization nightmare: how do you optimize a discrete hardware architecture for a neural network whose optimal structure is itself dependent on that hardware? Even auxiliary training structures present discrete bottlenecks; Alexander Detkov et al. note in Reparameterization through Spatial Gradient Scaling that determining the optimal configuration of auxiliary branches for structural reparameterization is a complex discrete optimization task that lacks an analytical solution.
To bridge this gap, researchers have developed three core strategies: smoothing the discrete landscape into a continuous one, probing the landscape using zero-order numerical gradients, or abandoning gradients entirely in favor of evolutionary and reinforcement learning frameworks.
Smoothing the Landscape via Continuous Relaxations
To bring the power of gradients to discrete spaces, one prominent strategy is to transform the discrete "staircase" into a smooth "ramp." If the optimization landscape can be made continuous, standard backpropagation can be used to update structural parameters alongside network weights.
One way to achieve this is through probabilistic relaxations. For instance, Will Grathwohl et al. utilize the "Concrete relaxation" trick in Gradient-based Optimization of Neural Network Architecture to make discrete choices like network depth and width differentiable. Similarly, Griffin Lacey leverages a differentiable Gumbel-Softmax sampling strategy in Stochastic Layer-Wise Precision in Deep Neural Networks to dynamically allocate layer-wise bit-widths within a fixed bit budget, bypassing the non-differentiability of quantization choices.
Another approach is to map complex, non-Euclidean structures into a continuous latent space. Xiaoyu Lu et al. demonstrate this in Structured Variationally Auto-encoded Optimization by training a grammar-aware Variational Autoencoder (VAE) to embed discrete architectures into a smooth, low-dimensional space where Bayesian optimization can operate.
Because small structural changes can cause discontinuous jumps in performance, Jovita Lukasik et al. introduce Smooth Variational Graph Embeddings for Efficient Neural Architecture Search, which explicitly smooths the latent space to make performance prediction more reliable. Understanding this geometry is critical; Fabrizio Pittorino et al. show in Architecture-Aware Minimization (A2M): How to Find Flat Minima in Neural Architecture Search that analytically biasing the search toward "flat minima" in the architecture space ensures that small structural changes do not cause catastrophic drops in generalization.
Discrete Space: [Layer=3] -------------> [Layer=4] (No gradient between them)
\ ^
Continuous Latent: \--> ( Smooth Space ) --/ (Gradients flow smoothly)
Probing the Dark with Zero-Order Optimization
Relaxations are powerful, but they often introduce severe gradient approximations that distort the true loss landscape, leading to optimization instability. What if we bypass analytical gradients entirely while retaining a gradient-like update rule?
This is the core intuition behind zero-order (gradient-free) numerical estimation. Instead of calculating an exact mathematical derivative through backpropagation, we can probe the objective function by injecting stochastic noise and observing how the performance changes. Zechun Liu et al. apply this concept in Joint Multi-Dimension Pruning via Numerical Gradient Update, treating discrete architectural parameters (like channel counts and depths) as a continuous pruning vector and updating them via estimated numerical gradients.
By abandoning analytical gradient approximations, frameworks like the one proposed by Xiaoxing Wang et al. in ZARTS: On Zero-order Optimization for Neural Architecture Search achieve remarkable stability. Crucially, zero-order optimization allows researchers to incorporate non-differentiable hardware constraints—such as real-world latency and energy consumption—directly into the optimization objective without needing differentiable proxies.
Embracing Discontinuity via Evolution and Reinforcement Learning
When a search space is completely non-convex, discontinuous, or involves complex structural modifications (like adding or removing physical components), even numerical gradients can fail. In these rugged terrains, population-based and reward-driven exploration methods become essential because they do not rely on local smoothness.
Genetic and Evolutionary Algorithms (EAs) mimic natural selection to "breed" optimal designs. This paradigm dates back to early efforts like NeuroGENESYS by Steven A. Harp et al. (Designing Application-Specific Neural Networks Using the Genetic Algorithm) and has evolved into a robust field surveyed by Yuqiao Liu et al. (A Survey on Evolutionary Neural Architecture Search), Xun Zhou et al. (A Survey on Evolutionary Construction of Deep Neural Networks), and Yanan Sun et al. (Evolutionary Neural Architecture Search and Applications [Guest Editorial]). EAs navigate discrete choices by evaluating a population, selecting the fittest candidates, and applying mutations.
Beyond standard genetic algorithms, other derivative-free metaheuristics offer powerful alternatives:
- Particle Swarm Optimization: Fei Ye leverages swarm intelligence in Particle swarm optimization-based automatic parameter selection for deep neural networks and its applications in large-scale and high-dimensional data to automatically select discrete network parameters like hidden layer and neuron counts.
- Subnetwork Selection: Philipp Altmann et al. demonstrate in Finding Strong Lottery Ticket Networks with Genetic Algorithms that genetic algorithms can discover highly accurate, sparse subnetworks ("strong lottery tickets") within randomly initialized models without calculating any connection gradients.
- Topology Discovery: This need for discrete exploration extends to tensor network topologies, where Meraj Hashemizadeh et al. (Adaptive Tensor Learning with Tensor Networks) highlight the necessity of advanced discrete optimization techniques to navigate high-dimensional structural landscapes instead of relying on rigid manual heuristics.
This immunity to non-differentiability makes evolutionary search uniquely suited for highly constrained physical hardware environments. For example, Florentia Afentaki et al. use genetic-based training in Embedding Hardware Approximations in Discrete Genetic-Based Training for Printed MLPs to optimize discrete logic gate selections for printed electronics, where standard backpropagation is completely incompatible. The same principle allows Xudong Feng et al. (Evolutionary auto-design for aircraft engine cycle) to automate discrete structural changes in complex aerospace systems.
Alternatively, Reinforcement Learning (RL) treats structural design as a sequential decision-making problem. In DISeR: Designing Imaging Systems with Reinforcement Learning, Kushagra Tiwary et al. represent camera hardware configurations as a formal language, using an RL agent to navigate the massive, non-differentiable search space of physical imaging parameters.
As Olivier Sigaud highlights in a comprehensive survey (Combining Evolution and Deep Reinforcement Learning for Policy Search: a Survey), the frontier of this field lies in hybridization: combining population-based global exploration to find the right discrete structure, with gradient-based local exploitation to tune continuous internal weights.
Key Foundational Work
Learning to Compose Neural Networks for Question Answering (2016) — contributed 1 concept elements to this direction.
Learning Neural Network Architectures using Backpropagation (2016) — contributed 1 concept elements to this direction.
Designing Application-Specific Neural Networks Using the Genetic Algorithm (1989) — contributed 1 concept elements to this direction.