Saatvik Rao
R Yesshu Dhurandhar
| Models | Training Time | Training Loss | Training Accuracy | Testing Accuracy | No. of model parameters | No. of epochs |
|---|---|---|---|---|---|---|
| VGG1 | 1749.3782160282135 | 0.5594148635864258 | 69.9999988079071 | > 55.000 | 40961153 | 20 |
| VGG3 | 80.03491282463074 | 0.203384131193161 | 97.50000238418579 | > 92.500 | 10333505 | 20 |
| VGG3 with data augmentation | 105.11191153526306 | 0.5282458066940308 | 73.7500011920929 | > 70.000 | 10333505 | 25 |
| Transfer learning using VGG16 or VGG19 | 151.68954753875732 | 0.02160074934363365 | 99.37499761581421 | > 97.500 | 17926209 | 10 |
| MLP | 68.79708528518677 | 0.2559237778186798 | 89.99999761581421 | > 92.500 | 15370497 | 15 |
Number of Parameters in VGG1:
33332+10010032128+128+128+1 = 40146017 (s = 2, p = 2 : at pooling stage)
⌊(n+2p-f)/s⌋ +1 x ⌊(n+2p-f)/s⌋ +1
Number of model parameters in the case of VGG1 is very high because there is only one convulation layer which is not decreasing the input size to the MLP as compared to VGG3. Therefore, the number of parameters used for MLP layers in VGG1 is very high.
We observe that as we increase the complexity of the model, the training accuracy increase, indicating that the model is fitting the training data better. Similarly, the testing accuracy also improves as we increase the complexity of the model.
Also, the testing accuracy for VGG3 model with data augmentation is lesser than orginal model becuase of:
- augmented images not captured the true variations in the data
- VGG3 architecture may not be deep enough
- introduced noise or irrelevant features
If the data size would have been very high, we would have got general model with higher accuracy.
Also, we observe that the training time increases as we increase the complexity of the model (for VGG3, VGG3 with Data Augmentation, VGG16), which is expected as more complex models require more computation time. VGG1 has highest training time becuase of high number of model parameters, therefore, it took more time for same numer of epochs as VGG3.
However, in some cases we can see that the test accuracy is higher than test accuracy and similarly, test loss is less than training loss. This is due to the small size of the dataset as well as the dataset itself. Also, the division of images for train and test varies the trainnig and testing accuracy for each corresponding epochs.
We can improve the performance of deep learning models based on VGG architecture using data augmentation. We can create variations of the images that can improve the ability of the fit models to generalize what they have learned to new images. This helps in better generalization and reduces overfitting.
But, the effectiveness of data augmentation in improving the performance of a model is largely dependent on the model's capacity to learn complex features and patterns in the data. VGG1 is a relatively simple model with only a few layers and a small number of parameters, which limits its capacity to learn complex features and patterns in the data. Therefore, data augmentation may not be as effective in improving the performance of VGG1 as it would be for more complex models like VGG3 and higher.
Overfitting can occur when the number of epochs is too high, resulting in a model that performs well on the training data but not on new data. On the other hand, underfitting can occur when the number of epochs is too low, leading to a model that has not learned enough from the training data and is not able to make accurate predictions. Therefore, it is crucial to find the optimal number of epochs that balances between the model's ability to learn from the training data and its generalization to new data.
As we can see in tha case of VGG 16, as we increase the no of epochs the train accuracy increases while test accuracy remains constant (or decreases).
Yes, the model gets confused in a few instances. For example, in some instances, when the tail of squirrel is hidden, it sometimes classifies it as rat rather than a squirrel, which is likely due to the similar physical appearance of the squirrel and rat in that particular image.
Such a MLP with comparable parameters to VGG16 model cannot learn complex features very well, becuase MLPs do not consider the spatial behaviour of the data like CNNs. The model is more susceptible to overfit due to large number of layers and neurons.
As discussed in class, the CNNs are Shift invariance (ouput to stay constant no matter how the input is transformed) and Shift equivariance (the output to undergo exactly the same transformation as applied to the input). However, in the case of MLP, it is not shift invariance and shift equivariance. Therefore, for image data, where the prediction should not depend on the position and orientation, CNNs are better than MLP.