Computational Intelligence - February 2015 - 22

Our second challenge required us to design a fusion technique that can work in classifier output space so that we could
integrate heterogeneous learners. FCUBE allows a learner's
classifier to either produce continuous values or produce a
label. Continuous results are then standardized to labels by
applying a decision rule. We implemented and evaluated a
number of decision level fusion techniques and finally decided
upon logistic regression.
Our third challenge involved management of data when a
large number of learners are deployed in a data-parallel manner.
It is a challenge to efficiently supply all learning algorithms
with different subsets of the training data, i.e. factor the data. To
support factoring, FCUBE implements a distributed data factoring service that is efficient but hidden from both FCUBE's
contributors and users.
III. Related Work

Ensemble-based learning strategies have been accomplished
previously in [6]-[8]. In [6], authors generate multiple models
from the same learner by providing different subsets of data.
These solutions are customized for a specific learner that the
researchers use. The approaches are also developed to enable the
learning of an ensemble on a specific compute cluster. However,
frameworks capable to learn 100's of models on the cloud from
a pool of heterogeneous learners are not readily available.
There have been efforts to enable evolutionary computation on parallel compute infrastructure. One such effort is EAsy
Specification of Evolutionary Algorithms EASEA [9], where
the authors claim that the user only needs to write some problem-related code. The framework is known to run on clusters
and GPUs but to our knowledge does not run on the cloud.
Another recent approach builds a framework called Distributed
Evolutionary Algorithms in Python (DEAP) [10]. The platform
is not as integrative as FCUBE because it was not designed for
distribution on the cloud.
Several machine learning systems are being developed to
run on the cloud and provide solutions for data science prob-

lems. Examples include [2], [3], and [4]. All these systems rely
on learners provided (or assembled) by the system developers
themselves and do not allow researchers to contribute to the
core learning algorithm repository. We embrace an extensible
approach that enables multiple researchers to contribute their
learners believing that the result will be unique and diverse,
ensemble learning systems.
Comparison frameworks are used to assess the performance
of different applications, and can speed up their development
and validation processes. One example is [11], a framework for
comparing optimization algorithms in the context of logistics.
However, most frameworks impose restrictions such as programming language or parallelization strategy among others.
Our insight is to provide, through virtualization, large-scale
resource access without the user needing to be concerned
about managing it. In FCUBE, contributions of external developers are totally decoupled from the framework software layer.
This enables the integration of stand-alone approaches in a
plug-and-play manner. Moreover, the hardware specifications
of the virtual machine where the framework will be executed
can be customized.
IV. Fcube Architecture

In this section, we give an overview of FCUBE's functionality
and introduce its different components. As depicted in Fig. 1,
domain users bring new problems and interface with an
FCUBE Server to generate massive data-parallel ensembles.
The FCUBE deployment server launches a number of cloud
nodes with a variety of learners executed with different algorithm parameters. Nodes sample data randomly from the data
server. Once learning is finished each learner provides the
model back to the FCUBE server. The server then learns a
meta-model (fused) using portion of data set aside for this, validates the model and outputs the final model to the domain
user. The repository of learners is composed of the algorithms
contributed by machine learning researchers.
A. FCUBE Components

Domain
User

New
Problem
Dataset

Data
Server
Dataset 1 Dataset 2 Dataset 3

Run

Model

FCUBE
Image

Factor
FCUBE
Server

Filter
Fuse

Cloud
FCUBE
Instances

Figure 1 FCUBE commons' typical use case.

IEEE ComputatIonal IntEllIgEnCE magazInE | fEbruary 2015

ML
New Researcher
Learner

FCUBE Learners

1) Learners: FCUBE deploys
stand-alone learners compliant with an input/output
specification.
2) FCUBE image: A cloud
image or snapshot contains
all the learner executables
and the logic that instantiates
a learner.
3) FCUBE Server: An FCUBE
server is responsible for
deploying FCUBE instances,
retrieving models, fusing models, and evaluating the resulting
meta-model.The server can be
instantiated from a publicly
available software repository.

Table of Contents for the Digital Edition of Computational Intelligence - February 2015