At its annual developer conference (I/O) on May 10, Google unveiled Duplex, a new technology that allowed its virtual assistant to have a conversation with a human over the phone, completing real-world tasks such as fixing a salon appointment and making a dinner reservation. You could tell the assistant to make an appointment or reservation and it would then make the call, talk to the person respond to questions, negotiate timing and even thank the person at the end. While Duplex is bringing natural language processing, deep learning and text-to-speech technology together, what is remarkable is that the conversation was so natural that the other person didn’t realise that he or she was talking to a bot because it changed annotation according to the flow of the conversation. While computers have been able to understand and generate natural speech, automated phone conversations are still stilted, and Duplex is the fulfilment of a long-standing goal of enabling humans to have a natural conversation with computers.
At the core of Duplex is a recurring neural network which mimics a human brain, internalising past experiences and drawing from them during training. In fact, Google’s artificial intelligence (AI) arm, DeepMind’s latest program was able to beat humans in a maze game after it learnt to find its way like a human. It was faster and was even able to find shortcuts when doors opened up suddenly in the maze. Navigating spaces is second nature to us because of neurons called grid cells in our brains that help us find our way. The program was trained with data on how rodents search for food and it was able to mimic the process of navigating from one place to another. The team, which also trained the network to navigate through unfamiliar mazes through reinforcement learning, found that the neural network developed something similar to grid cells formed in human brains.
While deep-learning algorithms have always outperformed humans in areas such as diagnosis, and future prediction, they take a long time to learn. What takes us few hours to learn