This project is a clean, modular C++ implementation of core neural network components, built from the ground up. It includes an end-to-end pipeline for training and evaluating models on an audio classification task — specifically vowel recognition — using a small dataset of audio sequences.
The primary focus is on understanding and implementing the internals of neural networks, not on optimizing performance. The classification task is simply a demonstration of how the framework can be used.
- Modular Neural Network Framework
NeuralNetworkclass to manage training and inferenceLayerclass supporting various activationsLossfunctions: MSE, cross-entropyRegularizersupport: L1, L2, ElasticNetOptimizerimplementations: SGD, Adam
- Preprocessing Pipeline
- Train/test split
- One-hot encoding of labels
- Standardization of input features
- Model Training
- Denoising Autoencoder
- Feedforward Classifier using the encoder from the autoencoder for feature extraction
- Minimal External Dependencies
- Only depends on Eigen for linear algebra
- 500 audio examples of vowel sounds (balanced across classes)
- Dataset available here: https://drive.google.com/file/d/1r0zLOusossMFZ0ZOQcye_iOE0Zoqrdlq/view?usp=drive_link
-
Preprocess the data
- Split into training/test sets
- One-hot encode labels
- Standardize features
-
Train a denoising autoencoder
- Corrupt input with noise
- Train to reconstruct clean inputs
-
Train a classifier
- Use the encoder from the autoencoder as the first few layers (for feature extraction)
- Add a classification head (fully connected layer + softmax)
- No hyperparameter tuning (by design)
-
Evaluate the classifier
- Achieves ~70% accuracy on the test set
- Eigen (for matrix operations)
To install Eigen:
# On Ubuntu
sudo apt install libeigen3-dev
# Or clone manually
git clone https://gitlab.com/libeigen/eigen.git