Even though we still have a couple of more years to go until we have robots walking around and behaving like an actual human, Artificial Intelligence (AI) has already shown a significant impact in more subtle ways.
With big players developing products such as voice recognition Siri by Apple, Google’s search predictions, Facebook targeted advertising, etc. all of these technologies have machine-learning algorithms in common enabling them to react and respond in real time. Redefining Digital Experience with AI was one of the many topics that were covered during the EmTech Digital conference in San Francisco last month. Speaker, Quoc Le, is a pioneer in machine learning and was honored on the 2014 list of young innovators under the age of 35. He is now technical lead on Google Brain and chose to discuss robotics during the Emtech event.
According to Quoc, recognition is the key challenge to address in order to build a better robot. How can this function be improved? Inside object recognition there is a phase called feature engineering, meaning you take some data and transform it into some sort of feature representation and later you use machine learning to recognize the object. This cumbersome procedure requires experts to help design the features. For instance, if the machine is to recognize a physical object, an expert in defining the object would have to be hired and for voice recognition they would have to hire someone who focuses in speech recognition. How else can the problem of robots learning without the help of humans be solved? The answer is Deep Learning, the object of Quoc’s research for Google Brain.
How Google Brain teaches itself
Deep learning is another word for neuron networks and it builds layers of extraction, going from raw images, to edge detectors, to corner detectors to eventually a cat, a face or other objects detectors. This is a very high level concept and does not require a human to do the programming, the system will learn these feature representations automatically.
Even though there is a lot of positive feedback there are still some drawbacks as most of the system is slow working which affects the efficiency to solve large problems. In order to solve this, one could scale up deep learning, either through the neuron network and break it into several partitions or taking the data and divide it into multiple shots. The importance of scaling up is substantial in other words increase the volume of data. Deep learning may not have as many gains at first in amounts of data compared to traditional machine learning however deep learning eventually becomes much more accurate if the amount of data is increased.
The young innovator started to question himself how it was possible for humans to be the first animal to have been to the moon considering that humans aren’t the fastest or strongest nor are they flying animals. But still, humans made it. He came to the conclusion that it was because of our intelligence, which gave him the idea of building a machine with some sort of intelligence that could help humans.
Why couldn’t machines figure out the answer to very basic questions without the help of humans?
During the conference, Quoc showcased a video of the robot his team and himself had built, who is able to go grab a stapler in another room and bring it back to the person who asked for it.
One of the first experiment Quoc and his team did was giving the Google Brain, YouTube videos to watch, about 10 million frames and with about 16,000 connected machines and a network that has about 1 billion synapses feeding the neurons. At the end of the process one of the neuron on the network detected the concept of a cat, without being told ahead of time what object it was. It clearly has been proven that significant improvements have been made with the error rate of image recognition decreasing from 38% (2011) to 5% now and the error rate for speech recognition dropping down from 23% (2011) to 8%. Even though it is yet too early to say how long and how much of an impact this evolution will have on people and businesses, we know that obtaining data won’t be an obstacle and that deep learning will eventually outshine traditional machine learning helping robots to become self-taught.