Embedded Files
Best and Worst Cases - Spiral Graphs
Best and Worst Cases - Spiral Graphs
Terminology
Epochs: Number of times dataset is passed through
LR: Size of the updates during training
Batch Size: Number of training examples
Betas: controls exponential decay rate for the first/second moment estimates
Weight-Decay: helps prevent overfitting by penalizing large weights
Clip_grad: controls whether gradient clipping is applied during training
Log-Interval: how frequently training progress is logged
Ansatz Exploration
Found that QCNN Ansatz produced the most consistent and satisfactory results in comparison to QAOA Ansatz and BrickBuilder Ansatz
QCNN Ansatz Modifications:
Layering different Ansatz
Refactoring QCNN code - no change to accuracies, but consistently decreased the losses
Encoder Exploration - QCNN Ansatz and AngleEncoder
Motivation for Testing: Improve train and test accuracy.
entanglement_depth [test results shown]
Plot for the Second Best Accuracy Achieved: 69%
* Observations:
Both training accuracy and test accuracy improved steadily, especially after the second epoch.
Training loss decreased consistently, showing the model is learning effectively within a short number of epochs.
* Steps we decided to take for the next model:
Increasing the number of epochs to push for even higher accuracy.
Possible exploration of larger datasets for better generalization.
Plot showing a significant increase in both training and test accuracy, along with a stable decrease in training loss.
Plot for the Best Accuracy Achieved: 70%
* Modifications:
Increased Epochs: Allowing the model more iterations over the data to improve learning.
Larger Batch Size: By increasing the batch size, we stabilized the training process and enhanced model performance.
After experimenting with various parameter settings, we reached a significant accuracy improvement of 70%
Plot for the Lowest Accuracy: 54%
Potential causes:
* Learning rate too high (0.01)
A high learning rate can cause the model to make large jumps in the parameter space, which may lead to skipping over optimal solutions, preventing the model from converging effectively.
* Limited number of epochs (only 4)
With each epoch, the model updates its parameters based on all the examples in the dataset.
Training loss decreases while both training and test accuracy remain unchanged.
Page updated
Google Sites
Report abuse