Multilayer Perceptron (MLP)
Dense feed-forward network for nonlinear relationships in tabular features.
Model development / technical proof
Types of Neural Networks — and what training one actually looks like.
Neural-network names are easy to list. The harder part is turning data into a reproducible experiment, establishing a baseline, reading the training evidence, making a defensible optimization, and proving whether the change helped. These working examples show that process end to end.
The model-development workflow
Training is a measurement process, not a single command.
Each repository uses the same controlled workflow so the evidence is comparable and auditable. A client can inspect the code, baseline, optimization decision, rerun, charts, environment metadata, and production notes.
1. DefineBusiness outcome, data, constraints, success metric.
2. BaselineStart with a defensible reference model.
3. DiagnoseRead loss, validation behavior, and failure modes.
4. OptimizeChange one controlled variable for a documented reason.
5. ProveRetest, preserve artifacts, explain implications.
Working experiments
Ten neural-network families, ten inspectable model-training case studies.
The datasets are intentionally synthetic and redistributable. The purpose is to make the engineering process visible without exposing client data or requiring specialized infrastructure.
Convolutional Neural Network (CNN)
Learns local visual features used in inspection, classification, medical imaging, and computer vision.
Recurrent Neural Network (RNN)
Carries a recurrent state through time for sequence-dependent predictions.
Long Short-Term Memory (LSTM)
Uses gated memory to preserve useful sequence information over longer spans.
Gated Recurrent Unit (GRU)
A streamlined gated recurrent architecture for ordered data.
Autoencoder
Learns a compact representation and reconstructs the input.
Variational Autoencoder (VAE)
Learns a probabilistic latent representation for reconstruction and controlled generation.
Generative Adversarial Network (GAN)
Trains a generator and discriminator together, making stability and diagnostics central to the workflow.
Transformer
Uses self-attention to model relationships across positions in parallel.
Graph Neural Network (GCN-style)
Uses message passing so predictions incorporate both node features and graph structure.
Apply the same process to your problem
Have a model that is underperforming—or data you think could support one?
A HerbDev technical review starts with the objective, available data, current model behavior, infrastructure constraints, and the definition of success. The first goal is not to make the model larger. It is to identify the smallest defensible experiment that reveals what to do next.