Neural ODE 🧩⏱️

Reproducible continuous-time deep-learning benchmark. DL 2025 team project. Reproducible comparison of continuous‑time Neural ODE models against discrete‑time baselines (ResNet, RNN, LSTM) on irregular time‑series forecasting tasks.

   

Neural ODE compares continuous-depth models against classic discrete-time baselines (ResNet-18, GRU, Time-LSTM) on irregular time-series and vision.

📊 Results

Task / Metric	ResNet-18	GRU (Δt)	Time-LSTM	Neural ODE
MNIST – Acc ↑	99.31 %	97.40 %	99.23 %	99.17 %
Params	11.7 M	1.2 M	11.7 M	0.21 M
Latency ↓	11 ms	0.49 ms	3 ms	3 ms
CIFAR-10 – Acc ↑	95.1 %	48.70 %	94.8 %	74.2 %
Latency ↓	12 ms	0.83 ms	4 ms	201 ms
PhysioNet 2012 – AUROC ↑	0.742	0.786	0.693	0.754
Latency ↓	7 ms	1.69 ms	0.6 ms	2.0 ms

🗂️ Datasets

Name	Domain	Samples	Δt pattern
PhysioNet ICU 2012	41 vital-sign channels	8 k patients	Highly irregular
MNIST / CIFAR-10	Vision (28×28 / 32×32×3)	60 k / 50 k images	Uniform

🏗️ Repository Layout

├── 1806.07366v5.pdf            # main paper
│
├── DL_NeuralODE.pdf      
├── GRU_implementation.ipynb
├── LSTM_MNIST_CIFAR-10_PhysioNet.ipynb
│
├── NeuralODE.ipynb            # initial Neural ODE draft
├── odenet_cifar10_metric.py   # Neural ODE on CIFAR10
├── odenet_mnist_metric.py     # Neural ODE on MNIST
├── odenet_physionet.py        # Neural ODE on PhysioNet
│
├── ResNet_CIFAR.ipynb        # ResNet on CIFAR
├── ResNet_MNIST.ipynb        # ResNet on MNIST
│
├── utils.py          
└── report.pdf                # project defense presentation

---

Team & Roles

Member	Role	Responsibility
@dpetrov835 Dmitry	ResNet Lead	Deep residual baseline
@Sklaveman Artemiy	RNN Lead	GRU baseline
@Akmuhammet01 Akmuhammed	LSTM Lead	LSTM baseline
@alina2002200 Alina	ODE Lead	Neural ODE module
@rainbowbrained	Data & MLOps	Dataset, CI, sweeps, evaluation