1 of 15

Building an Age Guesser using Transfer Learning

Arnav Garg, Shail Mirpuri, Darren Tsang

2 of 15

Task

3 of 15

Convolutional Neural Nets (CNNs)

4 of 15

Gradient Descent

5 of 15

Transfer Learning

6 of 15

VGG-19

Consists of:

3 x 3 filters with a stride of 1 and zero-padding
Max Pooling Layer with a stride of 2
19 layers

In order to perform transfer learning, we had to select a pre-trained model that best suited our objective. Through our research process in the project, we looked at a range of different models such as ResNet and the Inception Model. At the end, however, we decided to go for the VGG network due to its simplicity and computational efficiency especially when compared to the previously mentioned ResNet and Inception models.

The VGG 19 model is a model created during ImageNet, and consists of 19 layers. In this model we have simple network architecture, with each layer consisting of either 3x3 filters with a stride of 1 and zero padding or max pooling layers with the stride of 2. This means that in our max pooling layers we would be shrinking the dimensions of our image by a factor of 2. The channels, however, will be growing as we move along the model and increase the number of filters. With more than 130 million parameters, we decided that the VGG network would be sufficient for us to perform transfer learning on our dataset.

7 of 15

The IMDB-WIKI Dataset

500,000+ images of celebrities from IMDB and Wikipedia

Collected by researchers and publicly available online

Discarded images where the age did not make sense

Negative ages
Ages over 100

Created subsets from cleaned dataset for training, validation, and testing

Training used to train model
Validation used during training process to ensure overfitting does not happen
Testing used at the very end to evaluate how well our model does on new, unseen data

https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/

8 of 15

Hyperparameters

Learning Rate: 0.01
Batch Size: 10
Trainable: Last Layer
Optimizer: Adam
Random Sample 50,000 images to train our dataset (about 10% of total dataset)

Shail

In order to perform transfer learning effectively, we had to set some hyperparameters to achieve our desired goal. One of these was the learning rate, which determines how quickly the model learns or changes its weights in backpropagation, in short it determines how big the steps we towards achieving a global minima of the loss function. We decided to set the learning rate to 0.01, and use a dynamic learning rate by selecting Adam as our optimizer, which decays the learning rate over time. This means that at the beginning of training the model we would be taking large steps, however as we come closer to the optimum weights we take smaller and smaller steps until we find the actual minima. This is done in order to prevent stepping over the actual minima and missing it.

Apart from this, we also decided to keep the last layer as our only trainable layer. We decided this after reading several articles on transfer learning, which stressed the importance of tuning the final few layers. Since our model was already pre-trained for another task, we would use its pre-trained weights and just change the final layer, where the classification occurs. Finally our original dataset consisted of 500,000 images, however this was way too much to even process using a GPU on google Colab. So we decided to take randomly sample of 50,000 images from ages 16 to 65 to train our models and see how it performs. We also set aside 5,000 images each from original dataset for validation and testing of our model on the same age classes.

9 of 15

Results

Good Predictions

Actual Age

Predicted Age

10 of 15

Results

Bad Predictions

Actual Age

Predicted Age

11 of 15

Results

12 of 15

Results

Mean Absolute Error: 7.83 years
Standard Deviation of Absolute Error: 6.84

13 of 15

Results

14 of 15

Challenges

Real age vs apparent age
Massive dataset, low computing power
Age skew in dataset

15 of 15

Future Work

Make it work on uncropped images
Multiple people in images
Train for longer periods of time on full dataset
Build user application to expand dataset
Use loss function that better accounts for close guesses
Look into newer, more advanced network architectures (ResNet, Inception)