Neuroevolution
Neuroevolution refers to the use of evolutionary technique to optimize the architecture, hyperparameters,
and occasionally also the parameters of the neural network. While emergence of DNNs has highlighted the need and effectiveness
of neuroevolutionary methods, they have been around for a long time and substantial contributions , many of which yet unknown
to other fields of ML, were originally developed by the Evolutionary Computation community.
Most of the research on Neuroevolution has been focused on evolving DNNs with multiple components or strong design constraints,
such as UNET and generative adversarial networks (GANs). Application of GAs to evolve GANs were presented in
Garciarena_et_al:2018,
Garciarena_et_al:2020a,
Garciarena_et_al:2021c, and a library for evolving multi-components
DNNs was published in
Garciarena_et_al:2020.
Use of grammatical evolution (GE) for evolving UNETs (a particular architecture that comprises one convolutional and one deconvolutional network) was presented in
Lima_et_al:2020.
Also GE has been applied for optimizing the architectures of DNNs for computer vision tasks
Lima_et_al:2019a,
Lima_et_al:2021c,
and text classification
Lima_et_al:2022.
Adversarial examples
Adversarial examples are NN inputs perturbed in such a way that the perturbation is not noticeable to humans but it produces
an error in the prediction produced by the neural network. Adversarial perturbation methods are used to trick
or deceive the NN. The existence of adversarial examples is one of the main vulnerabilities of NNs and other ML models.
Universal adversarial perturbations are those that with a single perturbation try to simultaneously change the model
prediction for multiple examples.
In
Vadillo_and_Santana:2019,
we have introduced universal adversarial perturbations for speech command classification problems and proposed fine-grained
types of universal adversarial examples that target instances from only a subset of the classes. A critical assessment
of the metrics for evaluating the human perception of audio adversarial examples is presented in
Vadillo_and_Santana:2020,
Vadillo_and_Santana:2022, and an investigation of dominant classes as a result of the adversarial perturbations appears in
Vadillo_et_al:2020
Vadillo_et_al:2022.
The way in which adversarial examples are related to NN internal representation of the decision regions is not completely
understood. A number of hypotheses have been proposed to explain this phenomenon. In
Vadillo_et_al:2020b,
we explore "gaps" in the representation space of NNs attacked using Deepfool, a particular type of adversarial perturbation
method. In Vadillo_et_al:2020a,
we propose to extend adversarial attacks to produce adversarial class probability distributions.
More recently, we have proposed the use of adversarial perturbations within model-based evolutionary algorithms
Garciarena_et_al:2021d.
NN-based optimization
While PGMs have been the model of choice for model-based EAs such as EDAs, algorithms that learn a neural network
representing the relationship between the variables in the best solutions have been also proposed. A key question in
these approaches to find a fine balance between the capacity of the
neural network to represent the intricate relationships between the variables, and the computational time needed
to learn the neural networks and sample from them.
A review of early approaches to the use of NNs for modeling search distributions in EAs, is presented in
Santana:2017a.In
Garciarena_et_al:2018a,
we have proposed the use of variational autoencoders (VAEs) for modeling and sampling search distributions. We have introduced
two extensions of the VAE model to accommodate the specific ML problems that arise within the context of evolutionary search.
Another crucial issue for the use of NNs within evolutionary search is how to generate new solutions from a neural network. This
is a similar problem to that addressed with generative models. However, in the context of search distribution,
it would be beneficial that the generated
solutions that depart from the training solutions had a higher fitness or were biased to the regions of promising solutions. In
Garciarena_et_al:2020b,
we have investigated the use the back-drive technique, previously proposed by Baluja, for the sampling purpose. More
recently, in
Garciarena_et_al:2021d,
we redefine the problem of solution sampling by applying adversarial perturbation techniques.
Multi-task and explanatory neural networks
Multi-task NNs are models conceived for simultaneously solving multiple ML tasks (e.g., classification,
regression, and generation tasks). They can be a more efficient approach when a single module of the network
is involved in two or more task. However, this type of multiple-purpose networks are also harder to learn.
Explainable NNs are able to produce a human-understandable explanation of the their functioning. Designing
explanaible NNs usually implies a trade-off between the accuracy of the network to solve the ML tasks and
the clarity of the "reasoning" process used as conveyed to the human.
In
Garciarena_et_al:2019,
Garciarena_et_al:2021,
we have introduced the VALP model as a framework for NN multi-task learning. It has been extended in
Garciarena_et_al:2020c,
Garciarena_et_al:2021b,
where architecture search methods and operators specifically designed for multi-task optimization have been introduced.
An extensive analysis of explainable neural networks and their vulnerability to adversarial perturbations is presented in
Vadillo_et_al:2022.