Impact driven AI applications — Fast.ai students

A collaboration of mission-driven deep learning practitioners.

Co-written by Janardhan Shetty

“Hungry to learn and Driven by impact community”

It was not just another Friday evening here at the University of San Francisco. But it is one where “Hungry to learn and Driven by impact community” of Deep learning practitioners came together to discuss various domains, what are some high impact problems solvable through AI and how do we solve them. I only peripherally mentioned in my last post about why I think Deep learning could be an excellent tool for impact. Now, I am even more convinced that this tool can very well be the greatest digital tool of this generation for impact.

In this article, we give you a glimpse of what deep learning student community of fast.ai is passionate about and the talks are classified into domain areas. Here is the link for the recorded talk. Please watch this space as we will be going deeper into some of these applications in the future blogs.

Medical diagnosis

Janardhan Shetty spoke about early detection tool for Autism. Early detection of autism in kids saves ⅔ of the lifetime costs for their care. He explained how Deep Learning can be helpful to identify facial features of autistic kids, characteristic voice patterns through audio clips, behavioral video patterns and movements through video clips.

Jeremy Howard candidly shared how his professional journey led to an impact journey and how he started Enlitic, a company which uses Deep learning to augment doctors in medical diagnosis. Ljubomir Buturovic shared how his company ClinicalPersona is working on detection of breast cancer. In both these cases, medical imagery is used to detect and classify cancerous tissues.

Agritech

Sahil Singla spoke about how FarmGuide is digitizing and solving the information asymmetry problem in Agriculture, the biggest yet loss-making industry in India. It was fascinating to learn they actually scraped Google Earth and segmented individual farms which were essential to help farmers get loans by increasing transparency. It is amazing to see the use of advanced technology in one of the most neglected yet important areas.

Deforestation

My friend Sara Hooker, spoke about Delta Analytics, a non-profit she co-founded, which provided 15000 pro-bono hours to nonprofits all over the world over a period of just three years. She is now co-leading a project with Rain Forest Connection, to cut Deforestation, a leading cause of climate change. They are building models to detect the chainsaw noise captured from a continuous audio stream sent by recycled solar powered cell phone tied to trees. She also shared how they are working around with so little ground truth data and mentioned other logistic challenges as being able to detect the direction of the noise.

Disaster response and resilience

I, Sravya Tirukkovalur, have been following the work done by Digital Humanitarians during past natural disasters around the world and was very inspired by the power of crowd-sourcing and use of technology to tackle immediate problems. I spoke about my experience researching this space and proposed some paths where Deep learning can really move the needle including quickly assessing structural damage using satellite imagery (which currently takes a few days on average) and intent classification of social media and text help-line content to fast track critical information to on-field aid workers.

Education and accessibility to under-served populations

Samar Haider shared how he created first embeddings for 175K Urdu words, which is key in many NLP applications. He explained how he started solving the chicken and egg problem of language Corpus <-> knowledge by doing unsupervised training on 150 million tokens from Wiki and Urdu corpora.

Rachel Thomas shared her story on how uncomfortable she felt with disparities in access to technology and knowledge which led her to co-found fast.ai to make Deep learning really accessible to all populations. The people who are intimately familiar with the problems are often best equipped to solve it. And the key contribution of fast.ai, which I strongly believe is to give those populations the powers to use this incredible technology to solve their problems even faster!

Where do we go from here?

Above projects fall in various stages of development. Some of the above are already producing state of the art results and are validated to have a huge impact and there are some which are just getting started with ambitious goals.

As with any successful software project, it is important for us to start small and iterate to make sure we are asking the right questions, solving actual problems and having a real impact. And these are hard problems which can benefit from an army of passionate individuals, especially domain and tech experts. So collaborations are very welcome!

It is important to note that many of us had little machine learning background when we started the course. Deep learning, what started as a black box and a tough task to conquer, turned out to be not that hard. The results achieved by students are phenomenal. It’s just an inspiring journey and a great validation that anyone can get started irrespective of their backgrounds as long as they are passionate to keep at it.