Modeling cognitive states using machine learning techniques

Mete Özay (Department of Computer Engineering, METU, Turkey), İlke Öztekin (Department of Psychology, Koç University, Istanbul, Turkey), Uygar Öztekin (Google Inc., USA), Fatoş T. Yarman Vural (Department of Computer Engineering, METU, Turkey)

Developing ways to study mental representations has been a long-standing and intriguing challenge to cognitive scientists. Recently, advances in neuroscience and machine learning techniques have provided researchers with means to approach this question via assessing neural data that reflects brain activity during deployment of certain cognitive operations. The use of multi-voxel pattern analysis (MVPA) has enabled a novel and complementary approach to study the human mind, and in particular questions regarding information representation in the brain.

In this study, we classify categories of information represented in the brain during memory encoding and retrieval using machine learning techniques. Specifically, we model the distributed patterns of brain activity acquired with functional magnetic resonance imaging (fMRI) during an item recognition task, and employ machine learning algorithms, such as Support Vector Machines (SVM) and K-Nearest Neighbor (k-nn), to classify categories of information represented in the brain during encoding and retrieval stages of memory processing. In this algorithm, four dimensional spatio-temporal fMRI data (time and x,y,z-dimensions) are initially represented by tensors. Then the tensors are modeled by a compact description, called Local Relational Linear Predictive Coding (LR-LPC). This model extracts the quasi-periodic spatio-temporal correlation among the voxels in the brain, as shown in Figure 1-a. Then the LR-LPC parameters are used as input to k-nn. The results indicate excellent correlation between the training and test data. The receiver operating curve (ROC) across the ten classes is illustrated in Figure 1-b. Our results and simulations indicate that the proposed algorithm can provide accurate and robust classification of cognitive processing based on the corresponding distributed neural activity patterns in the brain.

 Acknowledgements: Supported by a Google Research Award.


Modeling cognitive states using machine learning techniques
Figure 1: (a) LR-LPC Representation of fMRI data, (b) Receiver Operating Characteristic curve of k-nn algorithm.
Preferred presentation format: Poster
Topic: Computational neuroscience

Document Actions