A neural network architecture that uses many neurons per layer rather than many layers, enabling it to learn diverse feature representations in parallel, often offering better performance on structured tabular data than deep narrow alternatives.
Also known as wide neural network, overparameterized network
A wide network is a neural network characterized by having many neurons (units) in each hidden layer, as opposed to a deep network which has many layers with fewer neurons per layer. Width and depth are the two primary dimensions along which neural network architectures are scaled. Wide networks can represent many different features in parallel across their neurons, while deep networks learn hierarchical abstractions by stacking many transformations in sequence.
The relationship between width, depth, and model capability is nuanced. Universal approximation theory establishes that a single-layer network with sufficient width can approximate any continuous function, but in practice deep networks tend to be more parameter-efficient for complex hierarchical tasks like image recognition and language understanding. Wide networks tend to perform better on lower-dimensional tabular data where the relevant features are not deeply hierarchical, and on tasks where parallel feature extraction is more valuable than sequential abstraction.
Wide networks have gained renewed theoretical interest through the neural tangent kernel (NTK) framework, which shows that infinitely wide networks have tractable training dynamics that can be analyzed as kernel methods. The double descent phenomenon—where model performance first improves, then temporarily degrades, then improves again as model size (including width) increases through an interpolation threshold—has also been studied in the context of wide networks. Modern large language models combine both extreme width and extreme depth, but the relative contribution of each dimension continues to be an active research area.
Wide networks matter to agencies primarily as a concept for understanding why AI tools perform differently on different data types. Many of the structured prediction tasks relevant to agency work—predicting campaign performance from tabular campaign data, scoring audience segments, modeling bidding behavior—favor wider, shallower architectures over the deep transformer models that dominate natural language and vision tasks. The architecture of an AI tool is often a signal about what data types and tasks it was designed for.
Width versus depth is a relevant vendor evaluation question. When an agency evaluates an AI prediction tool for campaign optimization, understanding whether the underlying model is wide, deep, or a hybrid helps assess whether it is well-matched to the data type. A very deep transformer-style model applied to simple tabular campaign data may be unnecessarily complex and prone to overfitting, while a properly-regularized wide network or gradient-boosted tree ensemble may outperform it with less data and compute.
Wide networks are often more interpretable. A wide, shallow network’s behavior is easier to analyze than a deep network’s: there are fewer non-linear transformations between input and output, and methods like feature importance scores and attention weights are more straightforwardly meaningful. For agency use cases where client stakeholders need to understand why a model makes particular recommendations, the interpretability advantage of wider, shallower architectures is practically relevant.
An agency analytics team compares two campaign outcome prediction models: a deep transformer-style model marketed as state-of-the-art, and a wide two-layer feedforward network with 512 neurons per layer. Both are evaluated on the same tabular campaign dataset of 50,000 rows with 40 features. The wide network achieves slightly higher validation accuracy and trains in 20 minutes; the deep transformer requires 4 hours of training and achieves marginally lower accuracy. The team selects the wide network not only for its performance and efficiency but because it supports cleaner feature importance analysis, allowing them to explain to the client which campaign parameters are most predictive of performance—a capability the deep model’s architecture does not support as cleanly.
The workshop covers how AI tools actually work, how to evaluate them, and how to apply them to real agency workflows.