“Can machine think?”-Few-shot learning might give you the answer

As early as the last 1950s, Alan Turing, the father of computer science, has ever proposed, “Can machine think?”Carrying this infinite question, human beings have built up myriad milestones. In recent years, powerful computing devices GPU and large-scale dataset ImageNet as well as cutting-edge models and algorithms spring up like mushrooms. Traditional machine learning is hinged on oceans of labeled image support, which involves great amount of manpower and energy. Nevertheless, the algorithm capability might be hampered when it comes to just a few shots. The machine will usually be on a strike as a result of insufficient valid labeled images.

Bridging the gap between human and machine is the key that the scientific community gives its life for. As a human being, one of evident hallmarks is we can establish recognition of a new concept just by one or a few instances at a very quick speed. It is scientifically proved that kids are capable of realizing what a zebra or a rhino is simply by some pictures given from books as well. As a result, how to make machine learning system effectively learn and promote its recognition capability from a limited quantity of samples turns a blueprint that researchers desperately want to accomplish.

Few-shot learning is a subset of Meta Learning. As its name suggests, learn to learn. Few-shot learning differs from conventional supervised learning, the aim of which is not asking machines to recognize objects that they have never seen before, instead teaching it how to perceive the intentions and distinguish among different objects. That is to say, the machine is asked to determine whether things displayed in two given pictures are the same or not. For example, if you give machines two images of the same spices which are out of few-shot training set, it obviously cannot recognize what it is, but it will understand the intentions and tell you whether they are similar or distinct. Let’s get further, if the machine is provided with a picture of otter and meanwhile it will be given a support set containing pictures of fox, squirrel, rabbit, hamster, otter and beaver. Neutral network can understand and select the one which looks most similar to the query sample. Of course the actual application will be more complicated.

Applications

With the progress of few-shot learning, it is widely used in various applications, like computer vision, natural language processing (NLP), robotics, audio signal processing, and others.

-Machine vision:

• Character recognition
• Image classification
• Object recognition
• Image retrieval
• Object tracking
• Specific object counting in images
• Scene location recognition
• Gesture recognition
• Image generation
• Shape view reconstruction for 3D objects
• Image captioning
• Video classification
• Motion prediction

-Natural language processing

• Parsing
• Sentence completion
• Sentiment classification from short reviews
• Criminal charge prediction
• Word similarity tasks

-Robotics

• Learning a movement by imitating a single demonstration
• Visual navigation
• Learning manipulation actions from a few demonstrations

-Audio signal processing

• Voice cloneing
• Voice conversion

Three big meanings of Few-Shot Learning:

• Subhuman learning method: Few-shot learning can bridge the gap between human intelligence and machine intelligence. Human beings can see the difference among written characters after learning a few samples. However, it is not the same story for machines. Few-shot learning is a kind of test base method, and we hope machines can also accomplish what we humans are able to do.
• Study of rare cases: Few-shot learning can be applied to study of rare cases. For instance, machine models with few-shot learning can correctly classify pictures of rare animals after learning simply a small amount of samples.
• Lower the cost of data collection and calculation: The methodology of few-shot learning will not too much depend on large-scale data set in order to avoid high cost of getting prepared for data in some specific applications.