I. Machine Learning and Feature Selection Applications
In several works, some listed below, we develop and apply various feature selection and machine learning tools towards various projects.
Some of the feature selection tools include: “ReliefF”, and various “wrapper” approaches where we use the machine learning as a measure of the quality of features. Specifically, we have used various greedy mechanisms as well as more complex tools such as genetic algorithms to help optimise the choice of features.
Our machine learning tools arsenal includes advanced neural networks, support vector machines, trees, Bayesian and so on. We often use the success of the machine learning tools as a measure and guide for data selection and input.
In our works, we have developed “one-class” techniques as well as anomaly detection. We have been champions and innovators in this direction for many years.
Some recent research projects include
- Early detection of Parkinson’s Disease directly from the speech signal
- Ability to handle subtle human memory classification — recently we have been able to distinguish between “positive” memories and “negative” memories directly from fMRI signals. (joint with Japanese research group)
- Earlier we showed how different human declarative memory systems can be identified from machine learning and feature selection techniques (joint with Canadian research group)
Principles: Alex Frid, Hananel Hazan, Dan Hilu, Larry Manevitz, Ohad Mofasi, Haim Shalelashvili
II. Using Virtual Reality and Machine Learning Tools to Classify Brain Damage from Motion Data.
This ongoing work examines brain damaged patients in a virtual reality environment and establishes that various kinds of brain damage (e.g. Stroke, Traumatic Brain Injury, Neglect) can be classified by machine learning tools solely from motion data of patients. This data is captured by observing human subjects performing a task in a virtual reality environment.
These results opens up the possibility of “closing the loop” by developing models of the individual patients and making virtual reality therapy personalized so as to optimize the improvements of the subjects.
Principles: Uri Feintuch, Larry Manevitz, Eugene Mednikov, Natan Silnitsky
III. Human Memory and Cognitive Systems
Modeling and Identifying Human Memory Systems via Machine Learning Tools
Principles: Asaf Gilboa, Larry Manevitz, Ester Koilis, Hananel Hazan, Gal Star, Aviad Itzchovich
Dr. Gilboa has developed certain theories relating to kinds of memories formulated by humans (episodic, generic and semantic). Computational consequences of these theories are investigated in our laboratory. Using machine learning tools, we are showing that one can reliably identify the kinds of memory system being used by a subject directly from fMRI scans. This shows that the purported memory mechanisms do exist and can in fact be identified from physiological data.
In this work, it is being shown that one can both identify areas of the brain associated with different memory systems and that this is precise enough to identify which system is being used by subjects performing various memory tasks directly from fMRI data. The methodology is machine learning techniques including feature selection and classification techniques; some developed specifically for this in our laboratory
Neural Network Modeling and Testing of Theories of Human Reading
Principles: Larry Manevitz, Hananel Hazan, Rom Timor, Orna Peleg, Zohar Eviatar
In a series of works — we investigate the inter-relationship between architectural structure and how humans disambiguate words during reading. The psychological theories being tested are due to Dr. Orna Peleg and Prof. Zohar Eviatar; while the computational modeling ideas and testing of simulated architectures are performed in the laboratory.
M.Sc. Degrees :
- Hananel Hazan
- Rom Timor
IV. READING THE MIND:
One Class Identification of cognitive tasks using artificial neural networks and other techniques
In this work, we show that one can effectively identify what picture a subject is looking at, solely from fMRI images taken while he is performing this task. This work emphasizes the development of “one-class” machine learning techniques, which means that training can be done from sample data of subjects performing the task. As opposed to two-class training, this means that the system is in principle scalable.
Techniques involved include one-class artificial neural networks, genetic algorithms and methods developed for the artificial generation of data.
Principles: Larry Manevitz, David Hardoon, Omer Boehm, Moran Haliwa
IV. Developing Neurocomputation Tools for Spatio-Temporal Signal Processing
We investigate the abilities of “reservoir computing” a relatively recent model of computation that postulates the separation of mental signals into two aspects – separation (performed by complex interconnections of neural structure) and identification by a separate system.
Our studies have shown that this innovative work, championed by Maass of Austria and Jaeger of Germany, is in fact too brittle to serve as a model for biological mental processing. However, we have also shown that certain architectural constraints does allow it to reliably process spatio-temporal signals.
Principles: Larry Manevitz, Hananel Hazan
V. Data Driven Models of Voxels of fMRI studies
Current Models of Voxels, used extensively by neuroscientists assume assumptions regarding the arrival of oxygen to the voxels according to the “Balloon Model”. However, the assumptions are not known to be well founded. In our work, we show how a reasonable model can be derived solely from data without underlying assumptions
Principles: Larry Manevitz, Ester Koilis, Hananel Hazan, Paolo Avesani, Diego Sona
VI. User Modeling and Associative Memories
In this series of works, we were interested in developing a system to allow a individually adapted museum guide could input and adapt a user model from external data. The underlying theory of the user model was adapted from work developed in neural networks.
Principles: Shlomo Berkovsky, Ariel Gorfinkel, Tsvi Kuflik, Larry Manevitz
VII. Applications of Neural Networks to the Finite Element Method
In these works, we showed how various techniques of Neural Networks can be adapted to increase the efficiency of the numerical system of finite elements. Finite Elements are one of the major tools for both civil engineering and aviation engineering.
IX. Applications of Machine Learning Techniques to Biological Applications
In this work, we used neural network and other machine learning tools to classify gene networks.
Principles: Larry Manevitz, Malik Yousef, Mohamed Ketany
X. One Class Learning Techniques and Document Classification
In a series of papers, we developed and applied one-class learning techniques and showed how it can be used to do document classification.
Principles: Larry M. Manevitz, Malik Yousef