Brain encoding and decoding
Encoding models are used in neuroscience to predict the brain response given some representation of the stimulus presented to the subject. Decoding models start from the analysis of the brain signals to predict the stimulus properties. Encoding and enconding models are relevant not just to understand the brain but to evaluate the capacity of machine learning algorithms.
In Mei_et_al:2019, Soto_et_al:2020, we investigate word embeddings and computer vision models fitted with the image referents of the words as inputs of encoding models. The approach allows us to examine the properties of the corresponding brain representations.
The Mind reading from MEG - PASCAL Challenge consisted of predicting the type of video watched by a subject from the analysis of its MEG recordings. A number of teams proposed different ML approaches to this decoding problem. Along the years we have used this dataset to design and evaluate a variety of ML methods. In Santana_et_al:2011f, an ensemble of classifiers approach with multiple sources of information is used to solved the problem; in Carrera_et_al:2016, different strategies to construct vine-copula classifiers are evaluated for the problem. More recently, in Santana_et_al:2019, genetic-programming based classifiers are used as an alternative for domain adaptation.
A different decoding problem was addressed in Santana_et_al:2012. As a distinguished feature, the introduced approach proposed the use of network measures extracted from interaction graphs learned from MEG data.
The brain works as a connected computational mechanism where brain regions are coordinated to participate in solving a diverse set of tasks. Structural and functional brain networks have particular characteristics.
In Santana_et_al:2011a, we used a multi-objective evolutionary optimization approach to generate optimized artificial networks that have a number of topological features resembling brain networks. One of the topological features investigated were network motifs, small network building blocks that are defined by their size and interconnection patterns. In Santana_et_al:2012, we use similar features but within the framework of brain decoding. One of the benefits of network theory when applied in Neuroscience is that it can be used to address a variety of problems.
Spiking neural networks (SNNs) are a particular class of neural networks that resembles the spiking behavior of neurons, conduction delays between the neurons influence their behavior and the emergence of polychronous groups. In Santana_et_al:2012d, Santana_et_al:2013, and by means of an evolutionary algorithm, we investigated the effect that biasing the delays has on temporal spiking patterns in SNNs, in particular on polychronization.
Neuron classification, and sypnasis prediction
Neuron functionality are related to their morphology. Therefore, a fundamental question to understand the different roles played by neurons in the cortex is to conceive method able to automatically classify neurons from morphological or physiological features. In Santana_et_al:2013d, we propose the application of affinity propagation for classification of neocortical interneurons.
Neurons communicate by means of chemical and electrical contacts called synapses. Therefore, one important question to understand the neuronal interconnection patterns is to identify the synapses. In LaTorre_et_al:2011, we propose a differential evolution algorithm for the detection of synaptic vesicles from images. Another problem that involves synapses is how to predict the formation of synapses between any pair of neurons given the genes expression patterns. For simple organisms, such as C-elegans, this question can be posed as a combinatorial problem. In Santana_et_al:2012g, we apply estimation of distribution algorithms to deal with it. We found that affinity propagation correctly classified most of the neurons in a blind, non-supervised manner.
Brain computer interfaces and EEG analysis
Brain computer interfaces (BCIs) allow the direct human-computer interaction without the need of motor intervention. There exist different paradigms of BCIs but in all of them there is a decoding machine learning component that decodes the intent of the subject from the analysis of the brain signals. We have applied combinations of optimization and machine learning algorithms for BCIs mainly based on EEG.
In Santana_et_al:2011, we propose a direct optimization approach to the P300 classification problem. The P300 component of the brain event-related-potential is one of the most used signals in BCIs. In Santana_et_al:2012c, we introduce the use of the covariance matrix adaptation evolution strategy (CMA-ES) for BCI systems based on motor imagery while in Astigarraga_et_al:2014, a channel selection strategy from an the application an estimation of distribution algorithm is introduced. The method is applied to a BCI based on imaginary movements of the left hand, right hand, tongue, or foot during a specified time interval according to a cue.
Other applications of ML techniques to the analysis of EEG for purposes other than BCI have been also investigated. For example, in Santana:2013a, Santana:2015, different classification methods are investigated for for a three-class vowel speech imagery recognition problem. In Zheng_et_al:2015, different classification methods for emotion recognition from EEG are investigated. We show that the combination of classifiers using stacking can achieve higher average accuracies than that without stacking methods.
Psychiatric condition prediction
Psychiatric condition prediction consists of diagnosing the particular type of psychiatric disease a patient has. Early diagnosis of psychiatric conditions can be enhanced by taking into account eye movement behavior. However, the implementation of prediction algorithms which are able to assist physicians in the diagnostic is a difficult task.
In Santana_et_al:2015d, we propose a method for classifying 6 medical conditions: Alcoholism, Alzheimer's disease, opioid dependence, Parkinson's disease, and Schizophrenia. We introduce a multi-view model of saccades in which the feature representations capture characteristic temporal and amplitude patterns of saccades; and apply four different types of classification algorithms.