9once elusive targets and demonstrating systems that really do surpass human ability when focused on specific tasks.Deep Learning, an approach originally inspired by the workings of the human brain, has produced a succession of breakthroughs. Deep learning is based on interconnected nodes of hidden layers between input and output layers. Deep learning was always thought to be a powerful “universal function approximator”—e.g., it can be used to model even complex non-linear problems like human vision or natural language processing. But breakthroughs were needed to build and train these systems. In the past couple decades, researchers methodically found clever solutions to make deep neural networks work.In more recent years, the pace of breakthroughs applying deep learning to traditional AI techniques has been breathtaking. Deep learning systems have demonstrated superhuman performance doing very human tasks such as image recognition. AI systems can already categorize photos of skin cancers better than dermatologists. Systems that translate from one language to another or that process speech are getting close to achieving human ability.In December 2017, Google’s UK-based subsidiary DeepMind, demonstrated a deep learning based program that when given only the rules of a board game such as chess or Go, could play itself over and over until, within days, it played at superhuman levels that could trounce any human or other computer. Google has used similar self-learning systems to significantly optimize and reduce costs in their server farms. Self-driving cars designed by Waymo, a Google subsidiary, can drive more than 5,000 miles without need for human intervention.A deep learning framework called a Generative Adversarial Network, or “GAN,” can even demonstrate an imagination. After learning from examples, say from pictures of movie stars, a GAN can generate pictures of imagined movie stars. Recently, a Silicon Valley firm, vue.ai, applied this technology to train a system to learn how to transfer photos of clothes hanging against a plain background so that they realistically appeared on computer imagined models.Applying to Your BusinessThese new advances are surprisingly accessible. Major companies such as Google and Facebook have made their software libraries publicly available. This means that your company can have direct use of the same application software used by market leaders that have spent billions developing their platforms. This, combined with a robust and open research community, is great news for newcomers looking to apply these technologies.The effort to apply AI is well justified. According to global consultancy firm McKinsey & Company, businesses that have proactively adopted AI have demonstrated significantly higher profit margins compared to their peers. Applying AI technology after its acquisition of robotics company Kiva, Amazon cut its operating costs by 20 percent by reducing inventories and cutting “click to ship” cycle time from 60 to 15 minutes. Netflix estimates that it saves $1 billion per annum from otherwise canceled subscriptions by implementing an AI algorithm to personalize recommendations.AI is actively applied in marketing for customer acquisition and retention, in production for planning and maintenance, and in finance for analytics, pricing, cost control, purchasing, and investment. Tech, auto, and financial services companies are leading the way but that does not mean other industries are not looking for a competitive edge. Big River Steel worked with Noodle.ai, a Silicon Valley firm, to build an entire steel mill in Arkansas with integrated AI solutions to, among other things, predict demand, reduce energy costs, enhance purchasing, manage inventory, and optimize production.Whether working with outside experts or building an internal capability, executives need to become familiar with these new technologies or risk falling behind competitors. Daniel BainDeep learning was always thought to be a powerful “universal function approximator”—e.g., it can be used to model even complex non-linear problems like human vision or natural language processing
< Page 8 | Page 10 >