Whether your work requires you to classify cucumber types, screen diabetics for debilitating eye conditions, or equip self-driving cars for the road—deep neural networks can help your organization in getting your work done in more intelligent and automated ways than has been possible so far.
A relatively new and exciting branch of machine learning, deep neural networks are loosely modeled after the human brain’s own complex wiring. Built as a series of deep layers with nodes that ‘light up’ the way neurons do when they process a causal connection, these networks are engineered to refine what machine learning customarily does. This is because deep neural networks enable far ‘deeper learning’ in ways that mimic the human brain. They not only assist in carrying out repetitive tasks that require a minimal understanding of a monotonous process, but also are specially designed to read, decipher, and connect a large volume of input—in the form of images, videos, words, or sounds—to intelligent outcomes.
Take the case of cucumber sorting. In 2015, a former systems designer from the Japanese automobile industry decided to help his parents streamline some processes of procurement at their cucumber farm in rural Japan. The nine different classes of cucumber grown on his parents’ plot of land used to take his mother eight hours of backbreaking work a day to sort in peak harvesting times. In Japan, the task of cucumber sorting is done not only according to size and thickness, but also by color, texture, prickliness, and whether the product is straight or bent. This is because each type bears a certain value and price. For instance, cucumbers that happen to meet a bunch of criteria simultaneously—including being ‘straight and thick with vivid color and lots of prickles’—are highly coveted because of these command higher prices in the market.
The former systems designer applied deep neural networks to cluster cucumbers into one of nine types for easy sorting and stacking. He did this by inputting thousands of images of the different varieties of cucumber into a simple cucumber sorter. The machine helped save his parents’ time and effort that they used to spend on manual work, and quickly delivered to them the kind of cucumbers that would fetch them the best price.
The ability of deep neural networks to intuitively and correctly recognize features after being fed with the widest possible range and volume of data on a subject is also a function that has helped researchers at Google’s Research Lab propose an unusual eye test for diabetics. They suggest that deep learning in combination with artificial intelligence can be applied to develop an eye-screening technique that can help ophthalmologists identify the signs of diabetic retinopathy. According to Google, such methods promise to help doctors correctly evaluate scores of patients quickly and refer those in urgent need of further help to a specialist. At Infinite we are integrating deep neural networks into our Zyter Health collaborative platform. Using neural networks to sift through a plethora of medical information can help physicians narrow down to only the most relevant patient information to help make timely and accurate decisions. This would not only save time per patient visit but also significantly improve the quality of care.
The capacity to ‘read’ tons of images and sounds by using this technology lends itself seamlessly to self-driving cars. Researchers at Berkeley say that speech, image recognition, translation, and robotics are functionalities that are crucial to making an autonomous vehicle run. So far, it has required different technologies to deliver these functions. However, since deep neural networks are equipped to simultaneously handle these capabilities, they can play a leading role in putting self-driven automobiles on the roads sooner than we imagine.
However, as with all new and evolving technologies, deep neural networks do pose some challenges. First, they require tons of data. This is not always feasible. Second, technology cannot be employed for tasks that need higher levels of strategic thinking, such as building a factory. Third, the technology itself is largely a mysterious ‘black box’—unable to explain why it gets to the results it does. And finally, it is not safe or suitable to fields where we require foolproof guarantees. For example, there shouldn’t be a minuscule chance that trains ‘be allowed’ to run on the same tracks at the same time.
Yet, businesses are waking up to the opportunities deep neural networks bring, especially in domains that rely on perceptive tasks such as sorting and ranking search results, interpreting x-ray images, or running autonomous cars. I believe that as technology builds a repertoire of successful applications, it is bound to find new avenues to mark its presence.
By Admin on 02 Apr 2020
By Rajesh Rao on 03 Oct 2019