
Brain Machine Interface
The CNEL interests in motor BMIs are very broad, ranging from the development of instrumentation, signal processing and new intelligent Brain Machine Paradigms based on co-adaptation. The ultimate goal is to develop clinical systems for rehabilitation. Presently we are addressing the issues of the design of subcutaneous electrode systems, portable instrumentation, distributed computation and remote data collection, intelligent co-adaptive architectures, novel signal processing algorithms and neurophysiology studies in conjunction with the Medical School. A technology transfer effort is also underway. We describe below the most salient projects.
- Sanchez J. and Principe J., Brain Machine Interface Engineering, Morgan and Claypool, 2007
- Neuroprosthetics Research Group (NRG)
- NeuroTechnology Workshop
Center for Innovative Brain Machine Interfaces
Source of Funding: NSF

This effort wraps the technology testbeds of the CNEL, NRG (Neuroprostetic Research Group), IML(Integrated Microelectronic Lab) and IEC (Integrated Electronic Center), into virtual companies to conduct technology transfer activities. Three virtual companies, with their own CEO, CTO, and CBO are presently underway, one developing Flexible Microelectrodes, the other Ultra low power devices for remote sensing and a rat pack to collect autonomously 64 channels of neural data and wirelessly transmit it to a base station in full resolution. Engineering students learn about entrepreneurship and business thru course work and by teaming with Business School students and faculty, while creating industrial prototypes related to their Ph.D. topics. The Center has an Industrial Board that meets once a year to pick new promising ideas and evaluating progress of the virtual companies.
Flexible Microelectrode Arrays
Sensing brain activity with microelectrode arrays is one of the enablers of BMIs. This virtual company product is a flexible microelectrode array that can be fabricated in batch mode using special combination of materials and micromachining techniques and can be easily integrated with silicon technology.
- Dr. Nishida Toshi
- Neuroprosthetics Research Group (NRG) lead by Dr. Justin C. Sanchez
Integrate and Fire for remote sensing
TBF
- David Cheney, Aik Goh, Jie Xu, Karl Gugel, John G. Harris, Justin C. Sanchez, Jose C. Principe, Wireless, In Vivo Neural Recordings Using a Custom Integrated Bioamplifier andthe Pico System ", Neural Engineering Informatics", /home/users/www/cnel_docroot/files/1485291325.pdf, pp. 19-22, 5 2007
FWIRE: Florida Wireless Implanted Recording Electrodes
Source of Funding: NIH
This projects seeks to develop an ultra-low power multielectrode array that will collect 16 channels of data, amplify it by 40 dB and wirelessly transmit it out at 500Kbits/sec on a small package (15x10x5 mm) using less than 2mWatts of power. The core of the technology is the Integrate-and-fire signal representation.
- Packaging and Electrodes
- Amplifiers with biphasic IF (Dr. Harris's lab)
- Ultra Low Power Wireless (Dr. Bashirullah's lab)
IF Analysis and Reconstruction
This project will characterize features of the IF sample representation such as data rates as a function of the input signal, and will develop on-line reconstruction algorithms in the back end.
Dynamic Data Driven BMIs
Source of Funding: NSF
This project seeks to develop new distributed models for motor control BMIs and open up the neurophysiology research laboratory to remote, online data collection and processing. The center piece is a cyberworkstation based on software virtualization technologies. We have already demonstrated a real time implementation of a closed-loop BMI by remote transfer of neural data thru the web, between the medical school and engineering where the computer cluster is located, to execute signal processing algorithms that control remotely a robotic arm in the neurophysiology laboratory.
Co-Adaptive BMIs

This project investigates closed-loop BMIs for reaching tasks in an animal model to develop new treatments for the disabled. Here, we focus on goal-based learning in motor BMIs to achieve control of a robotic arm. The approach is based on synergistic interaction between the user's neural modulations and an intelligent computational agent that are both seeking to maximize their own reward. For the development of BMIs, the RL framework provides a mechanism of learning that is very similar to operant conditioning of biological organisms because the learner (user) is not told what actions to take but must discover which actions yield the most reward by trying them.
- DiGiovanna J., Mahmoudi B., Fortes J., Principe J., Sanchez J., Co-adaptive Brain Machine Interface via Reinforcement Learning , Trans. on Biomedical Engineering, Vol. 54, No. 64, pp. 56-1, Jan 2009
- Neuroprosthetics Research Group (NRG) lead by Dr. Justin C. Sanchez
Cyberworkstation
TBF

Advanced Algorithms for Brain Activity
Sources of Funding: NSF, NIH
Monte Carlo Sequential Estimation for neural data analysis
This project seeks to evaluate the performance of the most general sequential estimation methodology for spike train analysis. We are using Monte Carlo techniques to lift the assumptions of Gaussianity and linearity in Bayesian filtering. The method requires accurate models of neural tuning functions, and indeed performs better than the rate models and Kalman filters in point processes.
- Y. Wang, A. Paiva, J. Principe, A Monte Carlo Sequential Estimation for Point Process Optimum Filtering ", Intl. Joint Conf. on Neural Networks", /home/users/www/cnel_docroot/files/1485291325.pdf, pp. 1846-1850, 7 2006
- Neuroprosthetics Research Group (NRG) lead by Dr. Justin C. Sanchez
Markov Chain Modeling for BMIs
The goal of this project is to use the theory of Markov models to create unsupervised algorithms that segment the multielectrode array data collected from motor cortex neurons during behavioral tasks. We are creating a graphical model with hidden dependencies to self-organize in space and in time the spike activity.
- * Darmanjian S., Principe J., Boosted and Linked Mixtures of HMMs for Brain Machine Interfaces, in European J. of Signal Processing, Special Issue on Machine Learning for Signal Processing, 2008
- Neuroprosthetics Research Group (NRG) lead by Dr. Justin C. Sanchez
Evaluation of ECoG for motor BMIs
TBF

- * Sanchez J., Gunduz A., Carney P., and J. Principe, “Extraction and localization of mesoscopic motor control signals for human ECoG neuroprosthetics,” in Journal of Neuroscience Methods – Special Issue on BCI, Volume 167, Issue 1, Pages 1-126, 2008
- Neuroprosthetics Research Group (NRG) lead by Dr. Justin C. Sanchez
Single Event ERP quantification
The electroencephalogram (EEG) is non-invasive and therefore it is the most likely candidate for the general use of brain activity to control external devices. The issue is the low signal to noise ratio and the nonspecific information carried by the EEG. Most often averaging is used to increase the SNR as done in event related potentials for BMIs. Here we develop a new single ERP algorithm that takes advantage of the massive number of electrodes of the current EEG systems to improve SNR, together with a generative model of the data to extract latencies and amplitudes of the ERP components (e.g. N100, P300). Preliminary results show that it is possible to quantify single ERP components in negative SNR (down to -15 dB) with reasonable accuracy. This project is conducted in cooperation with the NIH Center for Emotion and Attention at the UF.
- * Li R., Principe J., Bradley M., Ferrari V., “A novel spatiotemporal filtering methodology for single-trial ERP estimation”, accepted in IEEE Trans. Biomedical Eng., 2008
- CSEA (Center for the Study of Emotion and Attention)
Spike Based Computation
Funding Source: NSF
This project seeks to develop new computational models based on point processes. Biology uses spikes for communication amongst neurons and there are many advantages of using asynchronous events also in electronic analog implementations (low power, small real state, high dynamic range, low rates) for sensing, representation and computation. We have proposed an aperiodic sampling scheme (integrate and fire) and are studying signal processing methods of reconstructing and processing pulse train data in reproducing kernel Hilbert spaces (RKHS). Reservoir computing, specifically the Liquid State Machine, is a possible model of computation that is being investigated.
A RKHS for spike train analysis

One of the theoretical issues when working with spike data is the point process nature of the signal, which is very different from conventional time series, where the information is contained in the amplitude of the signals. Point processes carry the information in the time occurrence of the pulses. Normally point processes are studied as stochastic signals, but we are developing a functional analysis approach based on reproducing kernel Hilbert spaces (RKHS). We have defined bottom up a RKHS for the intensity function of PP, and have proposed kernels with and without memory that are applicable not only to Poisson but also renewal processes. The advantage of the RKHS formalism is that it creates an Hilbert space where conventional signal processing algorithms based on inner products can be readily developed and applied to point processes.
Paiva A., Park I., Principe J., "Reproducing Kernel Hilbert Spaces for Spike Train Analysis", in Proc. IEEE ICASSP 2008, Las Vegas.
Hybrid (biological-silicon) Liquid State Machines

Dissociated Cortical Tissue (DCTs) on MicroElectrode Arrays (MEAs) offer the possibility of studying the excitability of neural tissue in vitro. Our goal is to develop a hybrid (biological/silicon) computer based on the Liquid State machine (LSM) paradigm. Here the liquid is the DCT and the readout is a traditional computer. We have studied stimulation strategies of DCTs in MEAs to allow the coding of input signals into the firing of the DCT. We were able to show that the property of separation is indeed verified in DCTs and were able to create hybrid LSM classifiers. However, due to the stimulation artifacts and the periodic bursting present in DCTs, processing of time series is still not possible.