Skip to content

rmgarnett/active_learning

Repository files navigation

Active Learning Toolbox for MATLAB

This software package provides a toolbox for testing pool-based active-learning algorithms in MATLAB.

Active Learning

Specifically, we consider the following scenario. There is a pool of datapoints X. We may successively select a set of points x in X to observe. Each observation reveals a discrete, integer-valued label y in L for x. This labeling process might be nondeterministic; we might choose the same point x twice and observe different labels each time. In active learning, we typically assume we have a budget B that limits the number of points we may observe.

Our goal is to iteratively build a set of observations

D = (X, Y)

that achieves some goal in an efficient manner. One typical goal is that this training set allows us to accurately predict the labels on the unobserved points. Assume we have a probabilistic model

p(y | x, D)

and let U = X \ X represent the set of unobserved points. We might with to minimize either the 0/1 loss on the unlabeled points

\sum_{x in U} (\hat{y} \neq y),

where \hat{y} = \argmax p(y | x, D), or the log loss:

\sum_{x in U} -\log p(y | x, D).

We could sample a random set of B points, but by careful consideration of our observation locations, we hope we can do significantly better than this. One common active learning strategy, known as uncertainty sampling, iteratively chooses to make an observation at the point with the largest marginal entropy given the current data:

x* = \argmax H(y | x, D),

with the hope that these queries can better map out the boundaries between classes.

Of course, there are countless goals besides minimizing generalization error and numerous other strategies besides the highly myopic uncertainty sampling. Indeed, many active learning scenerios might not involve probability models at all. Providing a highly adaptable and extensible toolbox for conducting arbitrary pool-based active learning experiments is the goal of this project.

Using this Toolbox

The most-important function is active_learning, which simulates an active learning experiment using the following procedure:

Given: initially labeled points X,
       corresponding labels Y,
	   budget B

for i = 1:B
  % find points available for labeling
  eligible_points = selector(x, y)

  % decide on point(s) to observe
  x_star = query_strategy(x, y, eligible_points)

  % observe point(s)
  y_star = label_oracle(x_star)

  % add observation(s) to training set
  X = [X, x_star]
  Y = [Y, y_star]
end

The implementation supports user-specified:

  • Selectors, which given the current training set, return a set of points currently eligible for labeling. See selectors.m for usage and available implementations.

  • Query strategies, which given a training set and the selected eligible points, decides which point(s) to observe next. Note that a query strategy can return multiple points, allowing for batch observations. See query_strategies.m for usage and available implementations.

  • Label oracles, which given a set of points, return a set of corresponding labels. Label oracles may optionally be nondeterministic (see, for example, bernoulli_oracle). See label_oracles.m for usage and available implementations.

Each of these are provided as function handles satisfying a desired API, described below.

This function also supports arbitrary user-specified callbacks called after each round of the experiment. This can be useful, for example, for plotting the progress of the algorithm and/or printing statistics such as test error online.

Selectors

A selector considers the current labeled dataset and indicates which of the unlabeled points should be considered for observation at this time.

Selectors must satisfy the following interface:

test_ind = selector(problem, train_ind, observed_labels)

Inputs:

  • problem: a struct describing the problem, containing fields:

    •  `points`: an ![(n x d)][13] data matrix for the available points
      
    • num_classes: the number of classes
    • num_queries: the number of queries to make
  • train_ind: a list of indices into problem.points indicating the thus-far observed points

  • observed_labels: a list of labels corresponding to the observations in train_ind

Output:

  • test_ind: a list of indices into problem.points indicating the points to consider for labeling

The following general-purpose selectors are provided in this toolbox:

  • fixed_test_set_selector: selects all points besides a given test set
  • graph_walk_selector: confines an experiment to follow a path on a graph
  • identity_selector: selects all points
  • random_selector: selects a random subset of points
  • unlabeled_selector: selects points not yet observed

In addition, the following "meta" selectors are provided, which combine or modify the outputs of other selectors:

  • complement_selector: takes the complement of a selector's output
  • intersection_selector: takes the intersection of the outputs of selectors
  •    `union_selector`: takes the union of the outputs of selectors
    

Query Strategies

Query strategies select which of the points currently eligible for labeling (returned by a selector) should be observed next.

Query strategies must satisfy the following interface:

query_ind = query_strategy(problem, train_ind, observed_labels, test_ind)

Inputs:

  • problem: a struct describing the problem, containing fields:

    •  `points`: an ![(n x d)][13] data matrix for the available points
      
    • num_classes: the number of classes
    • num_queries: the number of queries to make
  • train_ind: a list of indices into problem.points indicating the thus-far observed points

  • observed_labels: a list of labels corresponding to the observations in train_ind

  • test_ind: a list of indices into problem.points indicating the points eligible for observation

Output:

  • query_ind: an index into problem.points indicating the point(s) to query next (every entry in query_ind will always be a member of the set of points in test_ind)

The following query strategies are provided in this toolbox:

  • argmax: samples the point(s) maximizing a given score function
  • argmin: samples the point(s) minimizing a given score function
  • expected_error_reduction: samples the point giving lowest expected loss on unlabeled points
  • margin_sampling: samples the point with the smallest margin
  • query_by_committee: samples the point with the highest disagreement between models
  • uncertainty_sampling: samples the most uncertain point

Label Oracles

Label oracles are functions that, given a set of points chosen to be queried, returns a set of corresponding labels. In general, they need not be deterministic, which is especially interesting when points can be queried multiple times.

Label oracles must satisfy the following interface:

label = label_oracle(problem, query_ind)

Inputs:

  • problem: a struct describing the problem, containing fields:

    •  `points`: an ![(n x d)][13] data matrix for the available points
      
    • num_classes: the number of classes
  • query_ind: an index into problem.points specifying the point(s) to be queried

Output:

  • label: a list of integers between 1 and problem.num_classes indicating the observed label(s)

The following general-purpose label oracles are provided in this toolbox:

  • lookup_oracle: a trivial lookup-table label oracle given a fixed list of ground-truth labels
  • bernoulli_oracle: a label oracle that, conditioned on the queried point(s), samples labels independently from a Bernoulli distribution with given success probability
  • multinomial_oracle: a label oracle that, conditioned on the queried point(s), samples labels independently from a multinomial distribution with given success probabilities

About

Active learning toolbox for MATLAB

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages