NeuralNet Studio is an interactive web dashboard designed to visualize, monitor, and optimize deep learning models. This project specifically explores the integration of Forgetting Theory (inspired by cognitive psychology and the Ebbinghaus forgetting curve) into neural network optimization to improve generalization, control weight dynamics, and combat catastrophic forgetting or severe overfitting.
The application combines a powerful PyTorch backend for model manipulation with a clean, dynamic frontend built using Flask and Chart.js to track training and validation metrics in real time.
- Real-time Training Diagnostics: Interactive visualization of Training vs. Validation loss curves to instantly catch overfitting thresholds.
- Cognitive Forgetting Integration: Built-in experimentation frameworks for simulating memory decay, structural pruning, and adaptive regularization based on forgetting dynamics.
- On-the-Fly Configuration: Dynamically tweak configurations like total epochs, learning rates, and batch sizes before initializing training loops.
- Matplotlib Export: One-click visual extraction of standard publication-ready loss curves for academic documentation.
- Instant Inference Engine: An built-in "Predict" panel to test localized model inferences instantly post-training.
In traditional training configurations, networks frequently over-optimize on niche training data noise—visible when training loss plummets while validation loss climbs. This project introduces engineered "forgetting" dynamics to resolve this bottleneck:
- Weight Decay & Memory Fade: Using controlled L2 penalties to continuously decay low-impact parameters, effectively treating minor noise as forgettable artifacts.
- Elastic Overwrite Protection: Implementing consolidation metrics to ensure core generalized patterns are retained while redundant paths are recycled.
- Synaptic Dropout Schedules: Leveraging adaptive dropout constraints to prevent static co-adaptation of features.
- Backend: Python, Flask, PyTorch (NoisyLinear, Leaky ReLU, Adam Optimizer)
- Frontend: HTML5, CSS3, JavaScript, Chart.js, TailwindCSS / Custom Theme
- Logging/Plotting: Matplotlib