notesum.ai

Published at November 11

Scientific machine learning in ecological systems: A study on the predator-prey dynamics

cs.LG

cs.AI

stat.ML

Released Date: November 11, 2024

Authors: Ranabir Devgupta¹, Raj Abhijit Dandekar², Rajat Dandekar², Sreedath Panat²

Aff.: ¹RCC Institute of Information Technology; ²Vizuara AI Labs

Arxiv: http://arxiv.org/abs/2411.06858v1

Criteria

Neural ODE

UDE

Forecasting Breakdown Point

35% training data

30% training data

Optimizers Used

Adam (LR = 0.001)

BFGS for fine-tuning

Adam for 500 iterations,

BFGS for 100 iterations

Adam (LR = 0.001)

RMSProp for fine-tuning

Adam for 20,000 iterations,

RMSProp for 5,000 iterations

Loss Decrease (%)

85% after Adam and BFGS

Loss function: sum of squared differences

90% after Adam and RMSProp

Loss function: sum of squared differences

Neural Network Architecture

3 hidden layers with 100 neurons each

Activation function: RBF

Dense Network

3 hidden layers with 10 neurons each

Activation function: ReLU

Shallow network

Training Time (approx)

50 minutes

20 minutes

Parameter Count

\sim

35,000 parameters

(due to 100 neurons per layer)

\sim

3,500 parameters

(due to 10 neurons per layer)

Sensitivity to Data Scarcity

High

Low

Model Robustness on Small Data

Poor

Strong

Forecast Accuracy

Breaks down beyond

35% training data

Accurate forecasts even with

30% training data

Sensitivity to Noise

Unable to recover underlying trend

with noise (Gaussian,

\sigma=0.3

)

Able to capture underlying trend

despite noise (Gaussian,

\sigma=0.3

)