/* ---- Google Analytics Code Below */

Thursday, September 12, 2019

N Shot Learning

Remember we kind of discovered this accidentally during the early days of using neural networks.  Its a natural thing, having a simple example.   An having N shots is next.   But is often limited in context.

Artificial Intelligence is the new electricity - Andrew NG

If AI is the new electricity, then data is the new coal.

Unfortunately, just as we’ve seen a hazardous depletion in the amount of available coal, many AI applications have little or no data accessible to them.
New technology has made up for a lack of physical resources; likewise, new techniques are needed to allow applications with little data to perform satisfactorily. This is the issue at the heart of what is becoming a very popular field: N-shot Learning.

N-Shot Learning

You may be asking, what the heck is a shot, anyway? Fair question.A shot is nothing more than a single example available for training, so in N-shot learning, we have N examples for training. With the term “few-shot learning”, the “few” usually lies between zero and five, meaning that training a model with zero examples is known as zero-shot learning,  one example is one-shot learning, and so on. All of these variants are trying to solve the same problem with differing levels of training material.

Why N-Shot?
Why do we need this when we are already getting less than a 4% error in ImageNet?

To start, ImageNet’s dataset contains a multitude of examples for machine learning, which is not always the case in fields like medical imaging, drug discovery and many others where AI could be crucially important. Typical deep learning architecture relies on substantial data for sufficient outcomes- ImageNet, for example, would need to train on hundreds of hotdog images before accurately assessing new images as hotdogs. And some datasets, much like a fridge after a 4th of July celebration, are greatly lacking in hotdogs.

There are many use cases for machine learning where data is scarce, and that is where this technology comes in. We need to train a deep learning model which has millions or even billions of parameters, all randomly initialized, to learn to classify an unseen image using no more than 5 images. To put it succinctly, our model has to train using a very limited number of hotdog images. .... "

No comments: