Back to Projects
Deep Learning

Aquatic Toxicity Prediction

Deep neural networks study for predicting acute aquatic toxicity to Daphnia Magna.

Course: Neural NetworksCo-authors: Diego Quezada
Aquatic Toxicity Prediction

Objectives

  • 1Predict acute aquatic toxicity to Daphnia Magna using deep neural networks.
  • 2Analyze the impact of hyperparameters (learning rate, optimizers, initialization) on training convergence.
  • 3Compare regularization techniques (L1, L2, Dropout) to reduce overfitting.
  • 4Evaluate Extreme Learning Machines as an alternative to traditional deep learning.

Conclusions

  • High Learning Decay values generate very slow training.
  • SGD increases generalization slowly and continuously. Adam and RMSprop learn fast but overfit.
  • Incorrect learning rate can cause training divergence.
  • L1 rule generates zero-value weights, L2 generates scattered weights centered at zero.
  • Dropout effectively decreases overfitting.
  • Extreme Learning Machines have remarkable performance with no overfitting.

Technologies

  • Pandas
  • NumPy
  • Matplotlib
  • Plotly
  • Seaborn
  • Scikit-learn
  • TensorFlow
  • Keras