The Indian Express (Delhi Edition)
Google Machine Learning smart, not intelligent (yet)
ARTIFICIAL INTELLIGENCE (AI) has been the holy grail of computer science for over a hundred years, and we are finally beginning to scratch the first layer of this incredibly complex system. Despite all major players in the technology business investing heavily in the R&D of AI systems, the development of true AI would still seem very far away.
In Machine Learning era
According to John Giannandrea, former Head of Machine Learning and currently the Chief of Search at Google, there are three distinct levels of Machine Intelligence: Machine Learning, Machine Intelligence and Artificial Intelligence. Machine Learning is what we have just started to get right — it’s a system where an algorithm can be written to train a machine to behave in a certain way, given certain kinds of inputs. Machine Learning, a higher version, would be where the machine is able to take what it has learnt and adapt it to a new concept. And a true AI would be the kind which is able to teach itself new concepts and evolve, just like humans. We’ve just started to be able to get really good at generating Machine Learning algorithms, but we’re still very far from having a system that can take what it has learnt, and adapt it to a new situation, says Giannandrea.
Digital training grounds
At the core of any machine resembling the simplest levels of intelligence is “training”. Every machine has to be first trained to process information in a certain way. For example, if you show a machine a photo of a dog, it should be able to label it as a dog. To be able to get that result, Google runs massive amounts of training material through a “neural network”, which is essentially multiple layers of digital filters that mimic the human brain. Each layer has ‘ports’, which connect with corresponding ports just like the neurons in our brains, depending on the stimulus they carry. So, they feed the neural network hundreds of thousands of images of dogs (and only dogs) and check that the output is “dog” for all images. Every time there is an error, it is sent backwards into the neural network so it can “learn” from the mistake and adjust the recognition pattern. Google has managed to get some really great results from this, and the proof lies in the Photos app, which is able to segregate photos based on their content. You can type “cat” in the search bar in the Photos app, and it will show you all the photos in your library with cats in them. That is Machine Learning — and it is fairly limited, because while you will get all photos of cats, the “machine” would still not be able to segregate them based on breed.
Limits of Machine Learning
While it may seem really ‘intelligent’ for a piece of software to be able to separate your photos into albums based on their content, or suggest when you should leave for work based on traffic conditions (and the time by when you need to clock into work), Machine Learning remains, at this stage, extremely limited — and will fail to execute well if even a single variable changes. For example, if a cat were to be dressed up as a dog, would the Photos app consider it a dog or a cat? Google has been using Machine Learning to develop its voice recognition software as well, being able to identify and separate the voice of the speaker from ambient noise. It can detect various languages as well, however, what it cannot do is detect intonations, emotional patterns evident in speech or even something as “simple” as sarcasm.
Intelligent affairs currently
According to Giannandrea, Google’s Machine Learning API are in their nascent stages, but developing rapidly. Google is using Machine Learning to augment Search (auto complete), Youtube (suggested videos), Inbox and Allo, to name a few. Inbox has a feature where it generates automatic responses for emails based on its contents and, according to Giannandrea, 10% of mails that are sent out using Inbox use auto-responses. Allo takes this a step further, where the machine learns the way you communicate, and then makes suggestions for responses based on what it has learnt. The pinnacle of this technology, however, is the Google Assistant, which is able to detect language and even separate commanding voice from ambient noise. Google Now uses Machine Learning to generate relevant information for you, based on usage patterns.
The privacy issue
Giannandrea says all data that is used for training is aggregated into a pool and hence, anonymised. However, once the API is trained and implemented in a service, it is able to read the information you have agreed to share with Google and make suggestions based on that. The information-sharing is twofold: one, to train the API itself, wherein your data is anonymised, and then, once the service is ready, to make suggestions based on your activity. This is how Google is able to give traffic information on Maps — it collects data from thousands of commuters and displays it on the app, but you cannot identify which pixel on that red line corresponds to your car.
What the future may hold
The medical potential is significant. For example, if a voice assistant is able to identify extreme stress or depression in the voice of the speaker, it may be able to help by either automatically connecting the user with a loved one or suggesting counsellors in the area. The next step, which would be Machine Intelligence, is where the phone itself is able to offer suggestions for things even before you think of doing them. For example, if you’ve just landed a new job, the phone should be able to suggest that you get a new wardrobe. Google’s Machine Learning resources are available for free under the name of Tensor Flow, and anyone can start using the tool to train machines for specific tasks.