Deep Learning for Analyzing Perception of Human Appearance

Original

Deep learning techniques can be used to extract facial imaging biomarkers of human health status and to track the effects of cosmetic interventions. At the Deep Learning in Healthcare Summit, Research Scientist, Anastasia Georgievskaya from Beauty.AI, will be presenting a set of tools for analysis of perception of human age and health status. 

She will also demonstrate that when certain population groups are under-represented in the training sets, these populations are left out or may be subject to higher error rates. This is why Youth Laboratories launched Diversity.AI, a think tank for anti-discrimination by the deep-learned systems. The presentation describes the strategies for evaluating human appearance for machine-human interaction and reveals the risks and dangers of deep-learned biomarkers.

What motivated you to start your work in deep learning?

My academic background is in bioengineering and bioinformatics. My greatest passion in life is ageing research and I was inspired by a lecture, where Dr. Zhavoronkov presented the roadmap for extending human longevity. I have been tracking progress in deep learning since the 2013 ImageNet. But in early 2015 I participated in a bioinformatics hackathon co-organized by Dr. Zhavoronkov and his team and got some hands on experience with the deep neural nets. I was surprised to see that even in rather primitive tasks deep learning outperformed more traditional machine learning methods and decided to focus my career on deep learning. 

I then joined Youth Laboratories since it utilized deep learning to recognize the minute age-related changes on facial photographs and track the changes induced by the various interventions as well as emotions and other machine vision applications. I got to work with some of the best deep learning experts in the region and when we launched Beauty.AI, I got to work with the many creative people from all over the world. I have been in deep learning for about two years.

Tell us more about your current work.

Today I am the deputy director of Youth Laboratories and manage a team of scientists and programmers, and work with our collaborators in healthcare, finance and skincare industries. I am also managing the Beauty.AI contest and the development of RYNKL and other apps, and contribute to the Diversity.AI initiative. Most of our projects are in machine vision, where we are developing deep learning systems to predict the age and health status of the person, detect human emotions in videos, detect the various abnormalities in skin like wrinkles and pimples and predict the perception of a person’s appearance by a certain population group. 

My own big research project, which I am planning to use for PhD defence and series of publications is a cross-species imaging biomarker of aging. Our projects are very ambitious and ostentatious. I draw inspiration from the biographies of Emma Walmsley, because of her confidence and transformative visions, but I will try to make everything as open as possible and will be publishing in peer-reviewed literature.

What are the key factors that have enabled recent advancements in deep learning?

The three main factors are: the availability of big data, the ability to use GPU for deep learning applications thanks to the work of Ruslan Slakhutdinov, and “ice breaking” proof of concept results that demonstrate the power of deep learning to the world in image recognition. 

Now the field of deep learning is quickly commoditized with thousands of people taking courses and participating in literature review sessions. While many concepts in deep learning are rather old, there are many fresh waves like Generative Adversarial Networks (GANs) lead by very young people like Ian Goodfellow and Metalearning lead by Nando Freitas. While many of these latest advances are still driven by the pioneers in the field, commoditization of deep learning enabled scientists like me to start developing applied deep learning systems in niches, where expertise in computer science and math is not enough and the domain expertise is extremely important. 

Domain expertise in ensuring biological relevance of deep learning systems gives me confidence that Youth Laboratories and many other young companies will also produce many advances.

What present or potential future applications of Deep Learning in Healthcare excite you most?

Despite the major growth in deep learning, there are still very few teams that use deep learning for healthcare applications. There is still a major gap between advances in IT, which is exploding and biomedicine, which is very conservative. I think that the most exciting application of deep learning in healthcare in 2016 was the first proof of concept showing that deep generative adversarial autoencoders (AAEs) can be used to generate new cancer drugs.

This paper was published by Insilico Medicine in Oncotarget, where the editors struggled for two months to find reviewers for the paper, because of the gap between deep learning and biomedicine. Another major advance, also by the pharmaceutical artificial intelligence group at Insilico Medicine was the application of multiple deep learning predictors trained on structural data, transcriptional response data and metadata for a huge number of molecules to predict the Phase I and Phase II outcomes of clinical trials. Both papers were published in December 2016 and will have major implications on the pharmaceutical industry. In theory, these two discoveries may be used for personalized drug discovery and predicting the effects of gene therapy.

But I think that the main potential of deep learning is in multimodal learning with massive learned knowledge and starting from facial imaging data. Here I am biased, because that is what I am working on. 

To really commoditize deep learning in healthcare we need success stories like with ImageNet and the best place to start is facial imaging. The concept of linking the genome to the face is quite simple and Craig Venter’s team and others already demonstrated the proof of concept of extrapolating genetic information to predict the facial phenotype. We did the opposite and linked the faces to population groups with the various risk factors. If we had the ability to work with 23&Me or other companies in this area, we could make major progress. Now we are linking the facial data to the various syndromes and diseases that are clearly visible. Once we can use the face as the major biomarker of health status, both current and predicted, we can revolutionize many industries including nutrition and insurance. 

The face is also the best biomarker of your age. If you are trying the anti-aging intervention and it really works, you are likely to see it on your face. And if it does not work for a girl like me, what’s the point of the intervention? 

What are the main types of problems now being addressed in the field?

Currently the two problems that are addressed rather well are classification and prediction problems. When you have big data with certain tags, you can train the deep learning predictor that is likely to perform better than any other machine learning method. Imaging biomarkers working with the MRI and CT data are already pretty good. So are the deep learning systems working with data from the high-throughput experiments.

What developments can we expect to see in deep learning in the next 5 years?

I think the main development that we will see is that companies like Apple and Facebook will realize their transformative potential in healthcare and so will major banks. Currently we are working with one of the largest banks in the world, which already has the data we need to predict the health status of patients, but they do not realize the transformative potential yet and use us for other applications. 

Companies that have the most granular healthcare data will be transforming the field, especially in geographies, where governments did not yet impede this progress. I think the major progress will be made in drug discovery. Once artificially intelligent systems become better than chemists at the big pharmaceutical companies, which is a pretty easy task and Insilico Medicine already demonstrated the proof of concept, every small biotechnology company can start generating effective drugs that will not be covered by patents. This will be the single most important development.

How did Beauty.AI come about and what will you be talking about at the Summit? 

The big goal of Youth Laboratories is to use the facial imaging data to predict the health status of the patient and to keep the patient healthy and young as long as possible. We started by developing apps for tracking wrinkles and pimples, but quickly realized that we need to have large data sets of high-quality data to train the deep learning systems. We turned to the beauty industry and skin care companies for help. But we also realized that many of the beauty companies are more interested in analyzing the perception of how their products work rather than the actual hard numbers. We developed the first algorithms to analyze this perception focusing on perceived age and beauty and decided to run a contest called Beauty.AI. 

In addition, we wanted to contribute to a community of algorithm developers to generate new ideas on how to evaluate people using facial data and other metadata. That is how the contest came to be and it became more popular than all of our other projects with thousands of people entering the contest and a dozen of algorithm developers enrolling their judges into the competition. In this talk we will describe some of the approaches for evaluating the attractiveness of the human face. We will also highlight the importance of balanced and inclusive training sets to avoid racial, gender and age bias and present the Diversity.AI project dedicated to testing for and reducing such bias.

At the Deep Learning in Healthcare Summit, join Anastasia and many other experts including: Neil Lawrence, Professor of Machine Learning & Computational Biology at the University of Sheffield; Alex Matei, Digital Health Manager at Nuffield Health; Michael Kuo, Associate Professor at UCLA and more! 

Register here using discount code BLOG17 at checkout to save 20% off your ticket. 

Other Healthcare Articles 

Original

Deep Learning Healthcare AI Deep Learning in Healthcare Summit Women in Tech Deep Learning Algorithms


0 Comments

    As Featured In

    Original
    Original
    Original
    Original
    Original
    Original

    Partners & Attendees

    Intel.001
    Nvidia.001
    Graphcoreai.001
    Ibm watson health 3.001
    Facebook.001
    Acc1.001
    Rbc research.001
    Twentybn.001
    Forbes.001
    Maluuba 2017.001
    Mit tech review.001
    Kd nuggets.001