In this regard, foreign media conducted an interview with John Giannandrea, Google’s senior vice president of search operations. He served as Google’s machine intelligence business executive.
The following is the main content of the visit.
It is not the age of artificial intelligence but the age of machine learning.
John first clarified a problem that machine intelligence has three different levels: machine learning, machine intelligence, and artificial intelligence. Machine learning is our current stage of research and development. In a machine learning system, we can write an algorithm and enter some information to train the machine to run in a certain way.
The higher level of machine learning is that machines can absorb what they learn and adapt to new concepts. And true artificial intelligence can learn new concepts and achieve self-improvement, just like humans. The problem we have tackled so far is just writing artificial learning algorithms. Moreover, John said that it is too early for the system to absorb what it learns and rely on its own strength to adapt to the new environment.
Neural network and digital training field
The core of the machine to achieve the simplest level of intelligence is "training." Each machine must first receive "training" in order to process information in some way. For example, show a dog a photo of a dog and ask it to correctly mark the photo as a "dog." Google needs to provide thousands of pictures of dogs to a neural network. Essentially, a neural network is a multi-layer digital sensor that simulates the human brain.
Each of these layers has certain "ports" that are like neurons in the human brain and can be connected to the corresponding ports based on the stimuli they carry. Therefore, the researchers will input thousands of images containing only dogs into the neural network and see if the corresponding output information for all pictures is a "dog." Once an error occurs, the wrong information is fed back to the neural network so that the neural network can “learn†from the error and adjust the recognition mode. Google has made some significant progress, which can be reflected in the photo management application GooglePhotos - the application can sort photos based on content.
Users can enter "cats" in Google Photos, and the results will be found in the photo of all the cats in the gallery. This is a category of machine learning, but it is still quite limited. John pointed out that although this algorithm can retrieve all photos with cats, it cannot be divided into photos based on the cat's breed.
The true limits of machine learning
Machine learning at this stage has quite limited functionality. It can distinguish between cats and dogs, but there is no way to identify different breeds of cats. Machine learning can only operate within a very limited range of variables. Even if only one variable changes, it will not work properly. For example, if a cat is dressed as a dog, should the Photos application recognize it as a cat or a dog?
Google has been using machine learning technology to develop speech recognition software that can distinguish between speaker's voice and environmental noise. This software also recognizes different languages ​​but does not recognize intonation and emotional patterns (eg, satire) when speaking. It can only be run in a very limited set of parameters. To expand the range of parameters that can be run, researchers need to invest a lot of time and arrange thousands of trainings before the machine learning system can operate normally.