References

Active Learning Surveys & Benchmarks

• Burr Settles’ Survey gives a great introduction to the classical active learning framework.

• This active learning benchmark gives a very thorough comparison of many classic active learning methods using the logistic regression classifier on many datasets. They too find that uncertainty sampling is surprisingly effective across a wide range of datasets.

Methods We’ve Covered

• Bayesian Unceratainty Sampling

• The Core Set Approach - Silvio Savarese’s paper and Ran El-Yaniv’s paper.

• The Adversarial Approach

• The EGL paper by the Baidu team

• Coordinate Matching

• DAL - Our paper, submitted to ICLR 2019.

Other Interesting Methods

Ensemble Methods

As we already hinted at, since different active learning methods sometimes put emphasis on different examples, it may be beneficial to want to combine these methods together in some way. This hasn’t been attempted in a successful way in the batch framework, but there were a few papers that attempted this in the classic framework by formulating the ensemble as different arms in a multi armed bandit scenario, which is a field that has been studied for many years.

If you’re interested, this paper and this paper are good places to start.

Meta Learning

The field of meta learning is picking up a lot of speed in recent years, with the promise of being able to stop using algorithms based on heuristics in many fields and replacing them with algorithms that are learned. In the end of the day, an algorithm is a function which takes in examples and outputs a function. Because it’s a function, we can approximate it using some neural network architecture and try to learn the weights that will give us a great algorithm.

Meta learning has been used to try and tackle many types of problems, and active learning is one of them. If you’re interested, take a look at this paper, this paper and this paper.

Membership Query Synthesis Using GANs

We introduced membership query synthesis as a different framework for active learning, where the learner synthesizes new examples to be labeled by the annotator. Since the learner needs to have a strong model of the data distribution in order to synthesize examples which can be labeled by an annotator (and which aren’t pure noise), this field slowly faded away.

Thanks to huge improvements in the field of generative models brought by GANs, this is no longer such a problem and the field of membership query synthesis has been revisited with modern approaches. The results aren’t mind blowing yet, but there is reason to be hopeful. You’re welcome to read the paper.

Applications for Optimizing Neural Networks

Finally, there is a very interesting application of active learning in the rather surprising field of optimization. Neural networks are almost exclusively trained using mini batch SGD, where the mini batch is uniformly drawn from the training set.

By now you can probably guess how this process can be improved, since random sampling is something we’ve spent the whole blog trying to beat. Instead of trying to get a diverse and representative labeled sample from our data distribution, we can use active learning here to get a mini batch that is most informative for training the network and will lead to faster convergence of the network (and maybe even to a better optimum).

If you want to see many of the concepts we’ve covered applied to active mini batch selection, you can read this paper, this paper, this paper and this paper (among others).