![References [1/2]
1.Rothe, R., Timofte R., Van Gool, L. (2015). IMDB-WIKI – 500k+ Face Images with Age and Gender Labels.
Retrieved on May 10, 2021, from https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/.
2.Rothe, R., Timofte, R., Van Gool, L. (2016). Deep expectation of real and apparent age from a single image
without facial landmarks. Retrieved on May 10, 2021, from
https://data.vision.ee.ethz.ch/cvl/publications/papers/articles/eth_biwi_01299.pdf.
3.Parkhi, O., Vedaldi, A., Zisserman, A. (n.d.). VGG Face Descriptor. Retrieved on May 10, 2021, from
https://www.robots.ox.ac.uk/~vgg/software/vgg_face/.
4.Krizhevsky, A., Sutskever, I., and Hinton, G.E. (2012) ImageNet Classification with Deep Convolutional Neural
Networks. In NIPS, pages 1106–1114, 2012.
5.LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., and Jackel, L.D. (1989).
Backpropagation Applied to Handwritten Zip Code Recognition. In Neural Computation, vol. 1, no. 4, pp. 541–551.
6.Mathias, M., Benenson, R., Pedersoli, M., and Van Gool, L. (2014). Face Detection without Bells and Whistles. In
Proc. ECCV. Retrieved on May 10, 2021, from https://link.springer.com/content/pdf/10.1007%2F978-3-319-10593-
2_47.pdf.
7.Serengil, S. (2019). Apparent Age and Gender Prediction in Keras. Retrieved on May 10, 2021, from
https://sefiks.com/2019/02/13/apparent-age-and-gender-prediction-in-keras/.
8.Abadi, M., et al. 2016. TensorFlow: A System for Large-Scale Machine Learning. In 12th USENIX Symposium on
Operating Systems Design and Implementation (OSDI’16), pp. 265–283. Retrieved on May 10, 2021, from
https://www.usenix.org/system/files/conference/osdi16/osdi16-abadi.pdf
9.Chollet, F., & others. (2015). Keras. GitHub. Retrieved on May 10, 2021, from
https://github.com/fchollet/keras.
10.Van Rossum, G. and Drake Jr, F. L. (1995). Python Reference Manual. Centrum voor Wiskunde en Informatica
Amsterdam.
YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda](https://image.slidesharecdn.com/younger2-210630043622/85/Younger-Predicting-Age-with-Deep-Learning-28-320.jpg)
![References [2/2]
1. McKinney, W.. (2010). Data Structures for Statistical Computing in Python. In Proceedings of the 9th Python in
Science Conference (Vol. 445, pp. 51–56).
1.Deng, J., Dong, W., Socher, R., Li, L., Li, K., and Fei-Fei, L. (2009). Imagenet: A Large-Scale Hierarchical Image
Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255.
2.Kluyver, T., et al. (2016). Jupyter Notebooks – a publishing format for reproducible computational workflows. In
F. Loizides & B. Schmidt (Eds.), Positioning and Power in Academic Publishing: Players, Agents and Agendas, pp. 87–
90.
3.Agarwal, P. (2020). Age Detection Using Facial Images: traditional Machine Learning vs. Deep Learning. Towards
Data Science. Retrieved on May 10, 2021, from https://towardsdatascience.com/age-detection-using-facial-
images-traditional-machine-learning-vs-deep-learning-2437b2feeab2.
4.Parkhi, O., Vedaldi, A., and Zisserman, A. (2015). Deep Face Recognition. British Machine Vision Conference. In
Xianghua Xie, Mark W. Jones, and Gary K. L. Tam, editors, Proceedings of the British Machine Vision Conference
(BMVC), pages 41.1-41.12.
5.Chauhan, Nagesh, (2019). Predict Age and Gender Using Convolutional Neural Network and OpenCV. KDnuggets.
Retrieved on May 10, 2021 from https://www.kdnuggets.com/2019/04/predict-age-gender-using-convolutional-
neural-network-opencv.html
6.Sharma, Sagar (2017). Epoch vs. Batch Size vs Iterations. Towards Data Science. Retrieved on May 10, 2021
from https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-4dfb9c7ce9c9
7.West, Jeremy; Ventura, Dan; Warnick, Sean (2007). Spring Research Presentation: A Theoretical Foundation for
Inductive Transfer. Brigham Young University, College of Physical and Mathematical Sciences. Archived from the
original on 2007-08-01. Retrieved May 10, 2021
8.Serengil, S. (2019). Deep Face Recognition with Keras. Retrieved on May 10, 2021, form
https://sefiks.com/2018/08/06/deep-face-recognition-with-keras/
YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda](https://image.slidesharecdn.com/younger2-210630043622/85/Younger-Predicting-Age-with-Deep-Learning-29-320.jpg)
Younger: Predicting Age with Deep Learning
- 1. YOUNGER PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Mentor: AJ Sanchez, Ph.D. May 2021
- 2. Executive Summary Problem: to develop and evaluate image-based supervised models to predict the age of the person in a given image, using deep neural nets. Hypothetical Stakeholder: A beauty company, ‘Younger’, requested an age predictor app to demonstrate the value of their ‘age-defying’ product line. Results: Using a customized Convolutional Neural Network (CNN) and Transfer Learning, we were able to estimate a subject’s age from a photograph with an average error of about 7 years. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 3. The Code All project code is available on my Github repository: https://github.com/premonish/YOUNGER The project was developed in Python using appropriate libraries on Google Colab Pro to take advantage of fast GPU-based architecture TensorFlow and Keras were used for preprocessing and implementing the Convolutional Neural Network (CNN). YOUNGER: GITHUB YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 4. The Dataset The dataset used is called the “IMDB-WIKI – 500k+ face images With Age and Gender Labels” related to this paper is available for download for academic research only. https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/ Created from scraped images from IMDb and Wikipedia of famous people. Massive public dataset of people’s faces labeled with ages and gender. Images are centered and cropped. Photos from “IMDB-WIKI” Dataset YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 5. Practical Age Prediction Marketing Applications Ex: Quividi uses interactive signage which plays a specific ad/video based on age and gender Law Enforcement Estimating age to create a suspect profile Access Control Prevent underage access to prohibited materials – alcohol, nightclubs, cars, curfew enforcement, etc. Photo by Milan Malkomes on Unsplash YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 6. Challenges Many factors affect apparent age including: Makeup Photo resolution Photo processing (Filters) Camera Angle Lighting Expression Lifestyle: • Smoking, drinking, sun exposure, etc. Photo by Persnickety Prints on Unsplash YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 7. Dataset Challenges Dataset is imbalanced Ages 0-7 and ages 81-100 are underrepresented We will focus on predicting for ages 8-80 only In the future, we can collect more data for underrepresented age classes YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 8. ‘Average Face’ Images To explore the dataset, we created composite images ‘averaging’ the pixel values across various subsets: 1,000 females, 1,000 males, and all images from age 60-69 YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 9. Convolutional Neural Networks CNNs can achieve state-of-the-art results in image classification Modeled on the biological visual cortex of animals Multi-layered feature extraction Enhanced by large datasets, and GPU computing Convolution uses matrix multiplication against a filter for feature extraction Transfer Learning can expedite related tasks Photo by Vincent Ledvina on Unsplash YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 10. VGG-Face CNN Architecture Created for face identity recognition by Oxford’s Visual Geometry Group (VGG) Can identify 2,622 faces Transfer Learning is possible since the task of age prediction has similar lower-level features (eyes, mouths, noses, headshapes, etc.) VGG-Face Architecture Overview Source: https://sefiks.com /2018/08/06/deep -face-recognition- with-keras/ YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Convolutional Neural Networks
- 11. Performance Metrics • MAE – Mean Absolute Error – De facto performance metric for age prediction in literature • MAPE – Mean Absolute Percentage Error • Test Set Classification Accuracy • Test Set Classification Loss (categorical cross entropy) Photo by Thom Milkovic on Unsplash YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 12. Model Constants Fix constants for model comparison: Epochs = 200 Batch Size = 128 Training Set = 3,525 Test Set = 1,512 Training Set : Test Set Ratio = 70:30 Photo by Julius Drost on Unsplash YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 13. Model Overview • CNN 1 = VGG-Face ‘out of the box’ with SGD optimizer • CNN 2 = VGG-Face with Adam optimizer • CNN 3 = VGG-Face with Adam optimizer with pretrained weights • CNN 4 = VGG-Face with Adam optimizer with pretrained weights & data augmentation YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Clarisse Croset on Unsplash
- 14. CNN 1 - Baseline Model • CNN 1 = VGG-Face ‘out of the box’ with SGD optimizer YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Test Set Model Performance Metrics MAE MAPE Test Accuracy Test Loss CNN 1 11.22 44.43% 0.04696 3.92303 CNN 2 7.94 25.50% 0.08466 4.31565 CNN 3 7.78 25.23 % 0.06019 3.65387 CNN 4 7.54 24.74 % 0.06878 3.59017 Observations The Baseline model has an MAE of 11.22, MAPE of 44.43%, Test Set Accuracy of 4.69%, and Test Set Loss of 3.92303. We can observe the training set diverge from the validation set on accuracy and loss, possibly indicating overfitting. Fig 1. CNN1 Training Set / Test Set Loss Fig 2. CNN1 Training Set / Test Set Accuracy
- 15. CNN 2 • CNN 2 = VGG-Face with Adam Optimizer YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Test Set Model Performance Metrics MAE MAPE Test Accuracy Test Loss CNN 1 11.22 44.43% 0.04696 3.92303 CNN 2 7.94 25.50% 0.08466 4.31565 CNN 3 7.78 25.23 % 0.06019 3.65387 CNN 4 7.54 24.74 % 0.06878 3.59017 Observations The CNN 2 model has an MAE of 7.94 and MAPE of 25.50% a vast improvement from the baseline model on both metrics. The Test Set Accuracy is up to 8.466%, however the Test Set loss increased which is not a good sign. Observing figure 3 and figure 3 there is an early divergence of the training set from the test set which may indicate more extreme overfitting than the baseline model. Fig 3. CNN2 Training Set / Test Set Loss Fig 4. CNN2 Training Set / Test Set Accuracy
- 16. CNN 3 • CNN 3 = VGG-Face with Adam Optimizer with pretrained weights YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Test Set Model Performance Metrics MAE MAPE Test Accuracy Test Loss CNN 1 11.22 44.43% 0.04696 3.92303 CNN 2 7.94 25.50% 0.08466 4.31565 CNN 3 7.78 25.23 % 0.06019 3.65387 CNN 4 7.54 24.74 % 0.06878 3.59017 Observations The CNN 3 model has a slightly improved MAE of 7.78, MAPE of 25.23% and Test Set Loss of 3.65387. However, the Test Set Accuracy has decreased to 6.019%. Again, figures 5 and 6 indicate overfitting. Fig 5. CNN3 Training Set / Test Set Loss Fig 6. CNN3 Training Set / Test Set Accuracy
- 17. CNN 4 • CNN 4 = VGG-Face with Adam Optimizer with pretrained weights & data augmentation YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Test Set Model Performance Metrics MAE MAPE Test Accuracy Test Loss CNN 1 11.22 44.43% 0.04696 3.92303 CNN 2 7.94 25.50% 0.08466 4.31565 CNN 3 7.78 25.23 % 0.06019 3.65387 CNN 4 7.54 24.74 % 0.06878 3.59017 Observations The CNN 4 model has an improved MAE of 7.54, MAPE of 24.74% and Test Set Loss of 3.59017. The Test Set Accuracy has improved to 6.878%. In figure 7 the training set accuracy and the test set accuracy moving together. Similarly, in figure 8 the training set loss and the test set loss moving together until around 100 epochs. These are good signals that the model is not overfitting. We selected this as our “best” model to solve our age prediction problem. Fig 7. CNN4 Training Set / Test Set Loss Fig 8. CNN4 Training Set / Test Set Accuracy
- 18. Model Performance Comparison Test Set Model Performance Metrics MAE MAPE Test Accuracy Test Loss CNN 1 11.22 44.43% 0.04696 3.92303 CNN 2 7.94 25.50% 0.08466 4.31565 CNN 3 7.78 25.23 % 0.06019 3.65387 CNN 4 7.54 24.74 % 0.06878 3.59017 CNN 4 demonstrates the best performance on all metrics except Test Accuracy where CNN 2 performs best. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 19. Business Results CNN 4 achieved an MAE of 7.54 years. The average age prediction is 7.54 years away from the actual value. The estimator can be improved with more, balanced training data. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Markus Spiske on Unsplash
- 20. App Demo YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 21. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda App Demo
- 22. Conclusion We were able to create a functional image-based age predictor model using open-source software, a massive public dataset and free cloud-based GPUs. Python, TensorFlow, Keras, Google Colab, etc., enable a very exciting world of AI innovation open to many people willing to learn. Public source code and public datasets further research and can enable new learners to develop faster. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Charles Deluvio on Unsplash
- 23. Recommendations We recommend to our client that we create the beta version of our app with disclaimer that the model will be optimized over time We recommend that we focus on high- quality and greater quantity of data to create a more robust predictor for a wider age range. We would like to do a survey of other pretrained models and how well they perform with our age prediction problem. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Isaac Quesada on Unsplash
- 24. Future Work Data Acquisition – expand data set to include more people age 0-8 and 81 – 100. More high-quality data would increase the model’s performance and generalizability Experiment with other pretrained CNN models and fine-tune the promising models YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Egor Vikhrev on Unsplash
- 25. Lessons Learned 1. We are in a very exciting time in history where people can build impressive tools using open-source software 2. Python, TensorFlow, Keras, Colab and public datasets can accomplish the seemingly impossible. 3. Class imbalance has been one of the most difficult challenges to overcome on this project. Downsampling and data augmentation were both used to address class imbalance. 4. Keep it simple. Beware of feature creep. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Hannah Olinger on Unsplash
- 26. Acknowledgements I’d like to thank my incredible Springboard Data Science Mentor, AJ Sanchez, Ph.D. Chief Data Scientist and Principal Software Engineer at Exodus Software Services, Inc., for patiently guiding me along in this project. Also, my wife is pretty cool. Thank you for your inspiration, Pinky! YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Photo by Howie R on Unsplash
- 27. YOUNGER PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Mentor: AJ Sanchez, Ph.D. May 2021 THANK YOU FOR YOUR TIME! Question or Comments? Email: prem@prem5.com linkedin.com/in/prem-ananda/ YOUNGER: GITHUB
- 28. References [1/2] 1.Rothe, R., Timofte R., Van Gool, L. (2015). IMDB-WIKI – 500k+ Face Images with Age and Gender Labels. Retrieved on May 10, 2021, from https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/. 2.Rothe, R., Timofte, R., Van Gool, L. (2016). Deep expectation of real and apparent age from a single image without facial landmarks. Retrieved on May 10, 2021, from https://data.vision.ee.ethz.ch/cvl/publications/papers/articles/eth_biwi_01299.pdf. 3.Parkhi, O., Vedaldi, A., Zisserman, A. (n.d.). VGG Face Descriptor. Retrieved on May 10, 2021, from https://www.robots.ox.ac.uk/~vgg/software/vgg_face/. 4.Krizhevsky, A., Sutskever, I., and Hinton, G.E. (2012) ImageNet Classification with Deep Convolutional Neural Networks. In NIPS, pages 1106–1114, 2012. 5.LeCun, Y., Boser, B., Denker, J.S., Henderson, D., Howard, R.E., Hubbard, W., and Jackel, L.D. (1989). Backpropagation Applied to Handwritten Zip Code Recognition. In Neural Computation, vol. 1, no. 4, pp. 541–551. 6.Mathias, M., Benenson, R., Pedersoli, M., and Van Gool, L. (2014). Face Detection without Bells and Whistles. In Proc. ECCV. Retrieved on May 10, 2021, from https://link.springer.com/content/pdf/10.1007%2F978-3-319-10593- 2_47.pdf. 7.Serengil, S. (2019). Apparent Age and Gender Prediction in Keras. Retrieved on May 10, 2021, from https://sefiks.com/2019/02/13/apparent-age-and-gender-prediction-in-keras/. 8.Abadi, M., et al. 2016. TensorFlow: A System for Large-Scale Machine Learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI’16), pp. 265–283. Retrieved on May 10, 2021, from https://www.usenix.org/system/files/conference/osdi16/osdi16-abadi.pdf 9.Chollet, F., & others. (2015). Keras. GitHub. Retrieved on May 10, 2021, from https://github.com/fchollet/keras. 10.Van Rossum, G. and Drake Jr, F. L. (1995). Python Reference Manual. Centrum voor Wiskunde en Informatica Amsterdam. YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 29. References [2/2] 1. McKinney, W.. (2010). Data Structures for Statistical Computing in Python. In Proceedings of the 9th Python in Science Conference (Vol. 445, pp. 51–56). 1.Deng, J., Dong, W., Socher, R., Li, L., Li, K., and Fei-Fei, L. (2009). Imagenet: A Large-Scale Hierarchical Image Database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255. 2.Kluyver, T., et al. (2016). Jupyter Notebooks – a publishing format for reproducible computational workflows. In F. Loizides & B. Schmidt (Eds.), Positioning and Power in Academic Publishing: Players, Agents and Agendas, pp. 87– 90. 3.Agarwal, P. (2020). Age Detection Using Facial Images: traditional Machine Learning vs. Deep Learning. Towards Data Science. Retrieved on May 10, 2021, from https://towardsdatascience.com/age-detection-using-facial- images-traditional-machine-learning-vs-deep-learning-2437b2feeab2. 4.Parkhi, O., Vedaldi, A., and Zisserman, A. (2015). Deep Face Recognition. British Machine Vision Conference. In Xianghua Xie, Mark W. Jones, and Gary K. L. Tam, editors, Proceedings of the British Machine Vision Conference (BMVC), pages 41.1-41.12. 5.Chauhan, Nagesh, (2019). Predict Age and Gender Using Convolutional Neural Network and OpenCV. KDnuggets. Retrieved on May 10, 2021 from https://www.kdnuggets.com/2019/04/predict-age-gender-using-convolutional- neural-network-opencv.html 6.Sharma, Sagar (2017). Epoch vs. Batch Size vs Iterations. Towards Data Science. Retrieved on May 10, 2021 from https://towardsdatascience.com/epoch-vs-iterations-vs-batch-size-4dfb9c7ce9c9 7.West, Jeremy; Ventura, Dan; Warnick, Sean (2007). Spring Research Presentation: A Theoretical Foundation for Inductive Transfer. Brigham Young University, College of Physical and Mathematical Sciences. Archived from the original on 2007-08-01. Retrieved May 10, 2021 8.Serengil, S. (2019). Deep Face Recognition with Keras. Retrieved on May 10, 2021, form https://sefiks.com/2018/08/06/deep-face-recognition-with-keras/ YOUNGER: PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda
- 30. YOUNGER PREDICTING AGE WITH DEEP LEARNING Springboard Data Science Prem Ananda Mentor: AJ Sanchez, Ph.D. May 2021
Editor's Notes
- #2: Thank you for being here and taking the time to listen to my presentation my name is Prem and I'm a data scientist finishing up my data science career track with springboard data science. The name of this project is younger predicting age with deep learning.
- #3: Let's begin with a high level overview in the form of an executive summary So what is the problem we are trying to solve the problem is to develop and evaluate image based supervised models to predict the age of a person in a given image using deep neural Nets. The hypothetical stakeholders is a beauty company named younger they've requested an age predictor app to demonstrate the value of their age divine product line. The results using a customized convolutional neural network and transfer learning we were able to estimate a subject page from a photograph with an average error of about seven years.
- #4: The code all of the project code is available on my GitHub repository here at github.com/premohnish/younger. This project was developed in Python using appropriate libraries under Google colab pro to take advantage of their fast GPU based architecture . Tensorflow and keras were used for preprocessing and implementing the convolutional neural network.
- #5: The data set the data set IMDb wiki 500K plus face images with age and gender labels related to this paper is available for download for academic research only.
- #6: Practical age prediction so when do people actually use age prediction in practice well we have different use cases for example in marketing there are applications where a targeted video ad can be based on age specifically Quividi. Another example where we see age prediction being used or potentially being used is law enforcement so we can estimate age to create the profile of a suspect for example the suspect is a 24 year old male. Another practical use of age prediction is access control for example we were trying to control who can access certain prohibited materials such as alcohol purchasing alcohol I'm going to nightclubs nightclub entrance driving cars you can't drive a car under certain age obviously and curfew enforcement if we have cameras that can automatically detect a person's age it's easier to enforce curfew.
- #7: there are many challenges there are many challenges when dealing with predicting age from an image some of the many factors that affect that effect apparent age include makeup when someone wears makeup they can look a lot younger than they actually are for photo resolution if account if a camera is too high resolution people can look or actually older than they are when there's more detail there and if the photo is too low resolution it's hard to get a good amount of texture and detail which can skew results. Photo processing such as filters if you've ever taken a photo on Instagram you know that there's a lot of filters that are designed to make people look younger specifically more flattering or they take wrinkles away thus making you look younger than your actual age other factors include lighting xpression and the person a person can look a lot older than their age in real life due to lifestyle choices such as smoking can age your person drinking can age a person etc drug use so there are many things
- #8: So we have this data set that was created and collected by researchers in computer vision and their intention was to create a data set the largest data set about creating age predictions and gender predictions. Some of the challenges of working with this data set like many datasets is this data set is imbalanced as we can see in in the histogram we have over representation of ages 20 to 60 and ages 0 to 7 and ages 81 to 100 are under represented. For the sake of this project we will narrow our focus on predicting for ages 8 to 80 only. In the future we can collect more data for under represented age classes
- #9: Since this is a computer vision problem we don't go too much into exploratory exploratory data analysis as there is not much exploratory data analysis to do however we can investigate trends or similarities and patterns inherent in the raw images. So to explore commonality‘s among the data set we subset the data set into female images male images and we further subset by age by subsetting we were able to average the pixel values of 1000 female images and we were able to composite an image averaging the pixel values and creating an X exported image and we did the same for male images and we did the same for male and female images age 60 to 69. Upon looking at the faces we can observe a certain femaleness about the female images as one would expect where the face shape is marked by long hair on the side and the lips are lighter in color compared to the male images whereas the lips are darker the hair appears to be slightly darker and their eyebrows parrot appeared to be heavier. In the older 60 to 69 age category we can see what we would associate as elderly features specifically we can see that there is a head shape that is rounder there is evidence of thinner or lighter colored hair.
- #10: why did we choose convolutional neural networks to approach this problem. Convolutional known neural networks CNN's can achieve state of the art results in image classification in fact imagenet the annual competition for computer vision in classifying images the winners these days our always using convolutional neural networks ever since about 2012 when adeep CNN was used to achieve a 10% increase over the the next competitor. Convolutional neural networks are modeled on the biological visual cortex of animals. CNN's feature multilayered feature extraction where the first layers are extracting simple features such as a diagonal line and higher level layers are extracting features with greater complexities such as an eyeball. Convolutional neural networks are really made feasable and possible by GPU computing so these are graphics cards that allow really fast computation also fed by large datasets or big data the convolutional neural Nets can perform to their Max. Convolved features use matrix multiplication for feature extraction. Transfer learning can expedite related tasks. For example many of the implementations which are used through care OS and tensor flow actually allow for you to use an image net free chain pre trained model directly out of the box which means that you have the benefit of a previously trained network and you can transfer that hard one knowledge to a similar problem such as in this case face identification period
- #11: why did we choose convolutional neural networks to approach this problem. Convolutional known neural networks CNN's can achieve state of the art results in image classification in fact imagenet the annual competition for computer vision in classifying images the winners these days our always using convolutional neural networks ever since about 2012 when adeep CNN was used to achieve a 10% increase over the the next competitor. Convolutional neural networks are modeled on the biological visual cortex of animals. CNN's feature multilayered feature extraction where the first layers are extracting simple features such as a diagonal line and higher level layers are extracting features with greater complexities such as an eyeball. Convolutional neural networks are really made feasable and possible by GPU computing so these are graphics cards that allow really fast computation also fed by large datasets or big data the convolutional neural Nets can perform to their Max. Convolved features use matrix multiplication for feature extraction. Transfer learning can expedite related tasks. For example many of the implementations which are used through care OS and tensor flow actually allow for you to use an image net free chain pre trained model directly out of the box which means that you have the benefit of a previously trained network and you can transfer that hard one knowledge to a similar problem such as in this case face identification period
- #12: So what are the performance metrics when talking about age prediction. The defacto performance metric for age prediction in the literature and in academia is Mae mean absolute error so we will use this as our main performance metric ma E is the mean absolute error so that is all the errors minus the actual values take the absolute value of that and find the mean. Ma E gives a sense of the overall correctness or error of the model. MAPE is mean absolute percentage error and this gives us a sense of relative error in terms of percentages how far are we off the mark. And additionally we will track test set classification accuracy and test set classification loss. In this case our loss function is categorical cross entropy
- #13: In developing the model we will set some constants to make it easier to compare different models with one another specifically we will set the number of epoch's to 200 the batch size to 128 the training set is 3525 images and the test set is 1512 images the training set to test set ratio is 70 to 30.
- #14: We went through many different iterations and model variations and permutations. In the end we focused on for models with greater and greater degrees of complexity. The first model is our baseline model which takes advantage of a VGG face architecture directly out of the box meaning uncustomized with an SGD optimizer stochastic gradient descent. The next model we used was a VGG face model with an atom optimizer the next is a VGG face with pre trained weights and finally for CNN four we used a VGG face model with adam optimizer with pre trained weights and data augmentation
- #28: Thank you for being here and taking the time to listen to my presentation my name is Prem and I'm a data scientist finishing up my data science career track with springboard data science. The name of this project is younger predicting age with deep learning.
- #31: Thank you for being here and taking the time to listen to my presentation my name is Prem and I'm a data scientist finishing up my data science career track with springboard data science. The name of this project is younger predicting age with deep learning.