Skip to content

Echoes: Exploring Echo State Networks

Echoes is a research project comparing Echo State Networks (ESNs) with traditional Recurrent Neural Networks (RNNs) for video classification tasks using PyTorch and the UCF101 dataset.

Project Goals

  • Implement and compare different neural network architectures:
  • Echo State Networks (ESNs)
  • Traditional RNNs
  • LSTM/GRU networks
  • Other relevant architectures
  • Evaluate performance on video classification tasks
  • Analyze computational efficiency and training dynamics
  • Document best practices and findings

Getting Started

New to the project? Start with Installation to set up your environment, then check out Quick Start to run your first experiment.

Key Features

Flexible Architecture Comparisons

Train and compare multiple network architectures with configurable parameters and automatic tracking.

Persistent Storage

Experiments are tracked with MLflow, logs saved with TensorBoard, and model artifacts stored on persistent Azure disk.

Cloud Integration

Run GPU experiments on Azure VMs with automatic provisioning, data transfer, and cleanup.

Interactive Visualization

View model architectures, training metrics, and comparisons through the web dashboard.

Project Structure

echoes/
├── data/               # Dataset preparation and loading
├── models/             # Model implementations (ESN, RNN, LSTM, etc.)
├── experiments/        # Training and evaluation scripts
├── scripts/            # Utility scripts (Azure automation, deployment, etc.)
├── dashboard/          # Web dashboard for visualization
├── docs/               # Documentation (this site)
└── README.md          # Quick reference

Dataset

We use the UCF101 dataset: 13,320 video clips across 101 action categories. The full dataset is stored on persistent Azure disk at /mnt/echoes_data/ucf101/.

Next Steps