Learnware & Reuser#

Learnware is the most basic concept in the learnware paradigm. This section will introduce the concept and design of Learnware and its extension for Hetero Reuse. Then, we will introduce the Reuse Methods, which applies one or several Learnwares to solve the user’s task.

Concepts#

In the learnware paradigm, a learnware is a well-performed trained machine learning model with a specification that enables it to be adequately identified for reuse according to the requirement of future users who know nothing about the learnware in advance. Specifications are introduced in COMPONENTS: Specification.

In our implementation, the class Learnware has three important member variables:

  • id: The learnware id is generated by market.

  • model: The model in the learnware, can be a BaseModel or a dict including model name and path. When it is a dict, the function Learnware.instantiate_model is used to transform it to a BaseModel. The function Learnware.predict uses the model to predict for an input X. See more in COMPONENTS: Model.

  • specification: The specification includes the semantic specification and the statistical specification.

Learnware for Hetero Reuse#

In the Hetero Market (refer to COMPONENTS: Hetero Market for more details), HeteroSearcher identifies and recommends valuable learnwares from the entire market, returning learnwares with different feature and prediction spaces compared to the user’s task requirements, known as “heterogeneous learnwares”.

FeatureAlignLearnware and HeteroMapLearnware facilitate the deployment and reuse of heterogeneous learnwares. They extend the capabilities of standard Learnware by aligning the input and output domain of heterogeneous learnwares to match those of the user’s task. These feature-aligned learnwares can then be utilized with either data-free reusers or data-dependent reusers.

FeatureAlignLearnware#

FeatureAlignLearnware utilizes a neural network to align the feature space of the learnware to the user’s task. It is initialized with a Learnware and offers the following methods to extend the ability of this Learnware:

  • align: This method trains a neural network to align user_rkme``(the ``RKMETableSpecification of the user’s data) with the learnware’s statistical specification.

  • predict: Using the trained neural network and the original learnware’s model, this method predicts the output for the user’s data.

HeteroMapAlignLearnware#

If user data is heterogeneous not only in feature space but also in label space, HeteroMapAlignLearnware employs minor labeled data (x_train, y_train) from the user task to align heterogeneous learnwares with the user task. HeteroMapAlignLearnware provides two key interfaces:

  • HeteroMapAlignLearnware.align(self, user_rkme: RKMETableSpecification, x_train: np.ndarray, y_train: np.ndarray)

    • Input space alignment: Aligns the learnware’s feature space to the user task’s statistical specification user_rkme using FeatureAlignLearnware.

    • Output space alignment: Further aligns the label space of the aligned learnware to the user task through a simple model FeatureAugmentReuser, which conduct feature augmentation and is trained on (x_train, y_train).

  • HeteroMapAlignLearnware.predict(self, user_data)

    • If input space and output space alignment are performed, it uses FeatureAugmentReuser to predict the output for user_data.

All Reuse Methods#

In addition to directly applying Learnware, FeatureAlignLearnware or HeteroMapAlignLearnware objects by calling their predict interface, the learnware package also provides a set of baseline Reuse Methods for users to further customize single or multiple learnwares, with the hope of enabling learnwares to be helpful beyond their original purposes and reducing the need for users to build models from scratch.

There are two main categories of Reuse Methods: (1) data-free reusers which reuse learnwares directly and (2) data-dependent reusers which reuse learnwares with a small amount of labeled data.

Note

Combine HeteroMapAlignLearnware with the following reuse methods to reuse heterogeneous learnwares conveniently. See WORKFLOW: Hetero Reuse for details.

Data-Free Reusers#

Two methods for direct reuse of learnwares are provided: JobSelectorReuser and AveragingReuser.

JobSelectorReuser#

JobSelectorReuser trains a classifier job selector that identifies the most suitable learnware for each data point in user data. There are three member variables:

  • learnware_list: A list of Learnware objects for the JobSelectorReuser to choose from.

  • herding_num: An optional integer that specifies the number of items to herd, which defaults to 1000 if not provided.

  • use_herding: A boolean flag indicating whether to use kernel herding.

The most important methods of JobSelectorReuser are job_selector and predict:

  • job_selector: Train a job selector based on user’s data and the learnware_list. The approaches varies based on the use_herding setting:

    • If use_herding is False: Statistical specifications of learnwares in learnware_list, along with their respective indices, are used to train the job selector.

    • If use_herding is True:

      • The mixture weight is estimated based on user raw data and the statistical specifications of learnwares in learnware_list

      • The kernel herding method generates herding_num auxiliary data points to mimic the user task’s distribution using the mixture weight

      • The job selector is then trained on these auxiliary data points

  • predict: The job selector is essentially a multi-class classifier \(g(\boldsymbol{x}):\mathcal{X}\rightarrow \mathcal{I}\) with \(\mathcal{I}=\{1,\ldots, C\}\), where \(C\) is the size of learnware_list. Given a testing sample \(\boldsymbol{x}\), the JobSelectorReuser predicts it by using the \(g(\boldsymbol{x})\)-th learnware in learnware_list.

AveragingReuser#

AveragingReuser uses an ensemble method to make predictions. It is initialized with a list of Learnware objects and has a member variable mode which specifies the ensemble method(default is set to mean).

  • predict: The member variable mode provides different options for classification and regression tasks:

    • For regression tasks, mode should be set to mean. The prediction is the average of the learnwares’ outputs.

    • For classification tasks, mode has two available options. If mode is set to vote_by_label, the prediction is the majority vote label based on learnwares’ output labels. If mode is set to vote_by_prob, the prediction is the mean vector of all learnwares’ output label probabilities.

Data-Dependent Reusers:#

When users have a small amount of labeled data available, the learnware package provides two methods: EnsemblePruningReuser and FeatureAugmentReuser to help reuse learnwares. They are both initialized with a list of Learnware objects learnware_list and have different implementations of fit and predict methods.

EnsemblePruningReuser#

The EnsemblePruningReuser class implements a selective ensemble approach inspired by the MDEP algorithm 1. It selects a subset of learnwares from learnware_list using a multi-objective evolutionary algorithm and uses the AveragingReuser for average ensemble. This method effectively balances validation error, margin ratio, and ensemble size, leading to a robust selection of learnwares for specific user tasks.

  • fit: Effectively prunes the large set of learnwares learnware_list by evaluating and comparing the learnwares based on their performance on user’s labeled validation data (val_X, val_y). Returns the most suitable subset of learnwares.

  • predict: The mode member variable has two available options. Set mode to regression for regression tasks and classification for classification tasks. The prediction is the average of the selected learnwares’ outputs.

FeatureAugmentReuser#

FeatureAugmentReuser helps users reuse learnwares by augmenting features. In this method, outputs of the learnwares from learnware_list on the user’s validation data val_X are taken as augmented features and are concatenated with original features val_X. The augmented data (concatenated features combined with validation labels val_y) are then used to train a simple model augment_reuser, which gives the final prediction on user_data.

  • fit: Trains the augment_reuser using augmented user validation data. For classification tasks, mode should be set to classification, and augment_reuser is a LogisticRegression model. For regression tasks, the mode should be set to regression, and augment_reuser is a RidgeCV model.

References#

1

Yu-Chang Wu, Yi-Xiao He, Chao Qian, and Zhi-Hua Zhou. Multi-objective evolutionary ensemble pruning guided by margin distribution. In: Proceedings of the 17th International Conference on Parallel Problem Solving from Nature (PPSN’22), 2022, pp.427-441.