Skip to content

Latest commit

 

History

History
155 lines (123 loc) · 6.53 KB

how-it-works.md

File metadata and controls

155 lines (123 loc) · 6.53 KB
Raw

How does this work?

Deep learning is a phrase being thrown around everywhere in the world of machine learning. It seems to be helping make tremendous breakthroughs, but what is it? It's a methodology for learning high-level concepts about data, frequently through models that have multiple layers of non-linear transformations. Let's take a moment to analyze that last sentence. Learning high-level concepts about data means that deep learning models take data, for instance raw pixel values of an image, and learns abstract ideas like 'is animal' or 'is cat' about that data. OK, easy enough, but what does having 'multiple layers of non-linear transformations' mean. Conceptually, all this means is that you have a composition of simple non-linear functions, forming a complex non-linear function, which can map things as complex as raw pixel values to image category. This is what allows deep learning models to attain such amazing results.

The most common class of methods within the computer vision domain are Convolutional Neural Nets (CNNs). Typically, the challenge arises in how you have to choose how many layers your network has. And how to initialize the model parameter values (also known as weights). There’s a lot more, too. Basically, a deep learning model is a machine with many confusing knobs (called hyper-parameters, basically parameters that are not learned by the algorithm) and dials that will not work if set randomly. Yet, when good hyper-parameter settings come together, results are very strong.

The key goal of the image classifier toolkit is to reduce the complexity in creating a model that is suitable for a wide variety of datasets (small or large) with images of various categories.

Transfer Learning

It’s not uncommon for the task you want to solve to be related to something that has already been solved. Take, for example, the task of distinguishing cats from dogs. The famous ImageNet Challenge, for which CNN’s are the state-of-the-art, asks the trained model to categorize input into one of 1000 classes. Shouldn't features that distinguish between categories like lions and wolves should also be useful for discriminating between cats and dogs?

The answer is a definitive yes. It is accomplished by simply removing the output layer of the Deep Neural Network for 1000 categories, and taking the signals that would have been propagating to the output layer and feeding them as features to any classifier for our new cats vs dogs task.

So, when you run the Turi Create image classifier, it breaks things down into something like this:

  • Stage 1: Create a CNN classifier on a large, general dataset. A good example is ImageNet, with 1000 categories and 1.2 million images. The models are already trained by researchers and are available for us to use.

  • Stage 2: The outputs of each layer in the CNN can be viewed as a meaningful vector representation of each image. Extract these feature vectors from the layer prior to the output layer on each image of your task.

  • Stage 3: Create a new classifier with those features as input for your own task.

At first glance, this seems even more complicated than just training the deep learning model . However, Stage 1 is re-usable for many different problems, and once done, it doesn't have to be changed often.

In the end, this pipeline results in not needing to adjust hyper-parameters, faster training, and better performance even in cases where you don't have enough data to create a convention deep learning model. What's more, this technique is effective even if your Stage 3 classification task is relatively unrelated to the task Stage 1 is trained on. This idea was first explored by Donahue et al. (2013), and has since become one of the best ways to create image classifier models.

Pretrained Image Classifiers

The following shows the built-in state-of-the-art network architectures for image classification. We hope to be adding many more as the research in the field evolves:

Resnet

Detects the dominant objects present in an image from a set of 1000 categories such as trees, animals, food, vehicles, people, and more. The top-5 error from the original publication is 7.8%. The model is roughly 102.6 MBin size.

Squeezenet

Detects the dominant objects present in an image from a set of 1000 categories such as trees, animals, food, vehicles, people, and more. With an overall footprint of only 5 MB, SqueezeNet has a similar level of accuracy as AlexNet but with 50 times fewer parameters.

Performance

It’s always important to make sure any machine learning technique is consistent in its usefulness, and that its success is not a fluke. In order to do that, we've tested it on several datasets. For each dataset, the experiment was identical. We trained a model on a dataset of just a few images, and then evaluated the accuracy on a completely different dataset of roughly 1800 images per category.

Here is a short summary of results:

Resnet

Performance on 50 datasets with 5 categories with 25 training examples per category

  • Median accuracy: 0.994667
  • Max accuracy: 1.000000
  • Min accuracy: 0.978667

Performance on 50 datasets with 5 categories with 50 training examples per category

  • Median accuracy: 0.996000
  • Max accuracy: 1.000000
  • Min accuracy: 0.980000

Squeezenet

Performance on same 50 datasets with 5 categories with 25 training examples per category

  • Median accuracy: 0.961333
  • Max accuracy: 0.981333
  • Min accuracy: 0.912000

Performance on the same 5 categories with 50 training examples per category

  • Median accuracy: 0.970000
  • Max accuracy: 0.988000
  • Min accuracy: 0.876000

To give you some perspective, a user study was made on a dataset with 1000 classes and measured human performance is about 94%.