Tomorrow’s Forecast: Deep and Impactful

I was fortunate to have spent last week at the Google Deep Learning Indaba, an event bringing together 600 machine learning specialists from across the continent. The Indaba seeks to strengthen and celebrate machine learning in Africa, through teaching, sharing, networking and debate.

What struck me as the common thread across all lectures, poster sessions, discussions and debates was the enormous potential for deep learning (an aspect of AI that seeks to emulate the way humans acquire knowledge) to solve some of our most challenging problems. Below are some areas where deep learning is already having real-world impact.

Healthcare: Some medical diagnostic tasks are currently being solved using computer vision. A deep learning method has been successfully applied to three-dimensional optical scans from patients at a major eye hospital. Performance of the deep learning model in making a referral recommendation reached or exceeded that of human experts. 

Diabetic retinopathy, which affects approximately one third of diabetes mellitus sufferers and can progress to irreversible loss of vision, requires early detection for successful treatment. Researchers used a convolutional neural network on retinal images to detect diabetic retinopathy staging. 

Scans from the back of people’s eyes have also been used to work out a patient’s age, blood pressure and whether or not they smoke, which predicts their risk of suffering a major cardiac incident such as a heart attack. This application is useful because it may make it simpler, and faster, for doctors to assess a patient’s cardiovascular risk.

Financial advice: NextGenVest has built a retrieval-based neural network powered chatbot to deliver personalised student loan advice to GenZ customers over text message and at scale. The text message service is free and provides students with the help and advice they need to navigate the financial aid and student loan landscape. NextGenVest has already helped over 60 000 students in the US.

Agriculture: From crop and weed detection, disease diagnosis to inspection of fresh produce, the potential for deep learning to improve agricultural outcomes is significant. Uncontrolled weeds cause enormous reductions in crop yield and quality. Ecorobotix , a Swiss startup, has built a system to recognise and spray weeds in real time. The result is a highly targeted application leading to 20x less herbicide used.

Translation: Over the past two years, deep learning has fundamentally changed language translation systems. Older translation systems were built by coding complicated rules defined by linguists. Unfortunately languages don’t follow a rigid set of rules, and these systems didn’t fare well in real-world situations. Statistical translation introduced probabilities and gave more likely translations, which helped with real-world text. While results were better than rule-based systems, these systems were complicated to build and maintain and still required human experts to tweak. Recurrent neural networks can now take in parallel corpora and translate between two languages without the need for human experts.

Seismology: The US state of Oklahoma does not lie on any major faults and is a historically low earthquake area. However in 2015 the state experienced more than 900 earthquakes, thought to be linked to the state’s fracking industry. Seismologists in Oklahoma trained a convolutional neural network to recognise background noise in seismically “quiet” areas. Ambient sounds can now be removed from the data leaving behind small earthquakes that had previously been masked. It also allows the identification of the rough location of individual earthquakes. This method allows the detection of even the smallest of quakes and seismologists hope it will eventually lead to prediction of earthquakes across the state.

From these examples it’s evident that deep learning techniques are set to solve some of our biggest challenges across a range of sectors.