Mike_Kiev - Fotolia
While there are many forms of machine learning, they generally group into one of three areas: supervised learning, which focuses on learning through example; unsupervised learning, which focuses on learning by observation; and reinforcement learning, which focuses on learning through trial and error. One of the challenges of supervised learning is that you need many examples of data to train a system adequately enough for a task. Contrarily, unsupervised learning doesn't need structured training data but does need mass information to find patterns.
Increasingly, developers are trying to combine different aspects of these learning approaches to augment the training process. By training supervised learning approaches using less data, developers can try to make use of reinforcement learning approaches to enable a hybrid semi-supervised machine learning approach, thus speeding up training time and handling more ambiguity. These ideas are being tested in recent developments around one-shot learning and semi-supervised learning.
The desired end goal for those pursuing the vision of artificial general intelligence (AGI) where systems can learn in environments of ambiguity or apply knowledge from one domain to another. The lack of a need for any prior knowledge is known as zero-shot learning; however, we haven't quite advanced to the point of zero-shot learning.
The intermediate step is to reduce the total amount of training data to provide accurate results. One-shot learning requires only a single instance of training data to build an accurate model. It's hard to generalize from one data point, so instead, this "golden standard" single data is used to identify an instance of circumstances in the process of training data. The one-shot can be used to refine a model that has already been trained on a much larger data set.
One practical example of one-shot learning is facial recognition. Instead of needing hundreds of examples of a face, we can provide one example -- given all the rest of the training data that has already been used to recognize faces in the abstract – and this one example can be used as a way to identify that one particular person. Images can be compared to one another based upon an image as a variable function. Instead of explicitly classifying every single image of an individual subject, like a person, images can be compared one-to-one.
In recent years, there continues to be advancements in one-shot learning processes. Companies such as IBM are working on minimizing data sets to yield the same quality results. It takes time and resources to teach an AI program, so cutting down on that timer will only speed up future advancements in the field.
Samsung developed an AI that can produce a fully animated three-dimensional head model from a two-dimensional image in one-shot learning. In another example, researchers at OpenAI built an AI system that was able to stack colored blocks in the correct order, despite the blocks starting from different positions every time. The computer was not simply stacking the blocks in an order but adapting to a new set of circumstances every time. There are 720 possible combinations of those six blocks, so it was an impressive feat that using a singular comparative data set (the correct block combination), an AI was able to learn how to problem-solve from numerous possible starting combinations.
Semi-supervised machine learning is a type of machine learning where an algorithm is taught through a hybrid of labeled and unlabeled data. Using unsupervised learning to help inform the supervised learning process makes better models and can speed up the training process.
A supervised learning algorithm assigns labels to data based on a set of labels that a human designed during the learning process. A simple example might be a human teaching an AI to look for an image on webpages. If the image is present, the site would be labeled under one category, while if the image is not present, the site would be labeled under a different category. Naturally, machine learning can involve far more complicated variables. The important thing to remember is that supervised learning entails a human presenting an AI with the method of categorization.
In unsupervised learning, the machine learning model might find inherent similarities and differences between those webpages and categorize them in a manner it thinks is most accurate. The categories may seem logical to a human, or a human may not see the same pattern. Because the learning process is unsupervised, the machine learning model will not have any direct human influence in its categorization. By combining these two approaches into semi-supervised learning, we do not restrict potential categories for whatever we're learning to the human-specified labels, but it begins from a set of human suggestions and categories. Algorithms produced by this type of machine learning approach enable the freedom of defining labels for the data, while still being directed by a human perspective.
Google's said semi-supervised machine learning may be easier with unsupervised data augmentation. If you were trying to perform supervised machine learning, the data sets would be practically incomplete and missing crucial labels. However, for the purposes of semi-supervised learning, unlabeled data can be incorporated. To deal with the difference between labeled and unlabeled data, Google taught an AI to produce two different types of predictions at once. The AI sorts labeled and unlabeled data and follows a different set of criteria for unlabeled data before comparing it. This ensures that consistency is relatively maintained, while still empowering the AI to learn more information through semi-supervised learning methods.
Currently, huge sets of data are needed for machine learning projects -- Google's facial recognition software utilizes 260 million images. The tedious nature of learning one thing at a time is a current restriction to the process. Regardless of what type of AI learning is performed, a system will need to be instructed on new information one piece at a time using current approaches. These latest updates to machine learning approaches with one-shot learning and semi-supervised machine learning aims to achieve better overall outcomes, while reducing the continuous need for large, well-labeled training data sets.