Neural Computing Engines
The NCEs project is focused on developing formal methods of massively parallel neural encoding/decoding, functional identification of linear receptive fields and biophysical spike generators, and encoding, decoding and identification of (non-linear) dendritic stimulus processors arising in sensory systems. This research is theoretical in nature. It builds on methods of communications/networking, control theory, information theory, machine learning, nonlinear dynamical systems, signal processing and system identification. A brief description of representative milestones with links to further reading is given below.
Massively Parallel Neural Encoding and Decoding
We pioneered formal methods of decoding of stimuli (signals) encoded with Time Encoding Machines (TEMs). TEMs are asynchronous signal processors that encode analog information in the time (spike) domain. TEMs play a key role in modeling the representation of the natural world by biological sensory systems and in building sensors for silicon-based information systems. There is also substantial amount of interest in TEMs as front ends of brain machine interfaces, i.e., as building blocks connecting biological and silicon-based information systems.
We analyzed the representation of stimuli in the spike domain by TEMs realized as single-input single-output, single-input multi-output, multi-input multi-output and space-time encoding circuits. The encoding neural circuits are built upon classical spiking neuron models including Integrate-and-Fire, Threshold-and-Fire, and Hodgkin-Huxley. Neural encoding circuits with feedback and random thresholds were also investigated. The encoding "silicon" circuits are built upon Asynchronous Sigma/Delta Modulators and FM modulators in cascade with zero-crossing detectors.
Time Decoding Machines (TDMs) recover the encoded stimulus from the time (spike) sequence. We discovered that arbitrarily precise stimulus recovery can be achieved under Nyquist-type rate conditions. Furthermore, we demonstrated for the first time the faithful recovery of natural video (movies, animation) and auditory scenes (speech, sounds) encoded with neural circuits consisting of canonical spiking neuron models (including Hodgkin-Huxley). The derivation of these analytical results is based on frame theory, the theory of dynamical systems, statistics and machine learning.
- Aurel A. Lazar, Population Encoding with Hodgkin-Huxley Neurons , IEEE Transactions on Information Theory, Volume 56, Number 2, pp. 821-837, February, 2010, Special Issue on Molecular Biology and Neuroscience.
- Aurel A. Lazar and Yiyin Zhou, Massively Parallel Neural Encoding and Decoding of Visual Stimuli, Neural Networks, Special Issue: Selected papers from IJCNN11, Volume 32, August 2012, Pages 303–312.
- Aurel A. Lazar and Yiyin Zhou, Reconstructing Natural Visual Scenes from Spike Times, Proceedings of the IEEE, Vol. 102, Number 10, October 2014.
A visual demonstration of decoding a short video stimulus encoded with a Video Time Encoding Machine is shown below.
The original video (upper left corner above) has a resolution of 640x360 pixels (HD format). The RGB channels of the original video were encoded separately. For each channel, the Video Time Encoder consists of 96,544 receptive fields in cascade with a population of Hodgkin-Huxley neurons. About 100 million spikes were generated per channel during the 10 seconds of the video. The reconstruction of the video stimulus is shown in the upper right corner. The visual error and its 2D spectrum of the R channel are shown on the bottom left and bottom right corner, respectively. The simulation was performed on a cluster with 55 Tesla M2050 GPU's.
Functional Identification of Linear Receptive Fields and Biological Spike Generators
We pioneered a novel class of algorithms for the functional identification of spiking neural circuits. Called Channel Identification Machines (CIMs) , these algorithms identify receptive fields in circuit models that incorporate biophysical spike-generating mechanisms (e.g., the Hodgkin-Huxley neuron) and admit both continuous sensory signals and multidimensional spike trains as input stimuli. Our neural circuit models take explicitly into account the highly nonlinear nature of spike generation that has been shown to result in significant interactions between various stimulus features and to fundamentally affect the estimation of receptive fields. Furthermore, and in contrast to many existing methods, our approach estimates receptive fields directly from spike times produced by a neuron, thereby obviating the need to repeat experiments in order to compute the neuron’s instantaneous rate of response (e.g., PSTH). The employed test signals belong to spaces of bandlimited functions and bridge the gap between identification using synthetic and naturalistic stimuli. This makes our methodology particularly attractive in those sensory modalities (most notably olfaction), where it is difficult to produce stimuli that are white and/or have particular distribution/attributes.
- Aurel A. Lazar and Yevgeniy B. Slutskiy, Channel Identification Machines, Computational Intelligence and Neuroscience, Volume 2012, pp. 1-20, July, 2012.
- Aurel A. Lazar and Yevgeniy B. Slutskiy, Functional Identification of Spike-Processing Neural Circuits, Neural Computation, Volume 26, Issue 2, pp. 264-305, February 2014.
- Aurel A. Lazar and Yevgeniy B. Slutskiy, Channel Identification Machines for Multidimensional Receptive Fields, Frontiers in Computational Neuroscience, Volume 8, Number 117, September 2014.
A visual demonstration of the identification of a non-separable spatio-temporal receptive field is shown below.
The original spatio-temporal kernel and the identified kernel are shown, respectively, in the first column. The Fourier amplitude spectrum of each of these kernels is respectively shown in the right column.
Dendritic Stimulus Processors: Encoding, Decoding and Identification
We investigated multi-input multi-output neural circuit architectures for nonlinear processing and encoding of stimuli consisting of a bank of dendritic stimulus processors (DSPs). DSPs execute nonlinear transformations of multiple temporal or spatiotemporal signals such as spike trains or auditory and visual stimuli in the analog domain. We demonstrated a fundamental duality between the identification of the dendritic stimulus processor of a single neuron and the decoding of stimuli encoded by a population of neurons with a bank of dendritic stimulus processors. We have also shown that identification algorithms can be effectively and intuitively evaluated in the stimulus space. In this space, a signal reconstructed from spike trains generated by the identified neural circuit can be compared to the original stimulus.
- Aurel A. Lazar and Yevgeniy B. Slutskiy, Spiking Neural Circuits with Dendritic Stimulus Processors: Encoding, Decoding, and Identification in Reproducing Kernel Hilbert Spaces, Journal of Computational Neuroscience, Volume 38, No. 1, pp. 1-24, February 2015.
- Aurel A. Lazar and Yiyin Zhou, Volterra Dendritic Stimulus Processors and Biophysical Spike Generators with Intrinsic Noise Sources, Frontiers in Computational Neuroscience, Volume 8, Number 95, September 2014.
- Aurel A. Lazar, Yevgeniy B. Slutskiy and Yiyin Zhou, Massively Parallel Neural Circuits for Stereoscopic Color Vision: Encoding, Decoding and Identification, Neural Networks, Volume 63, pp. 254-271, March 2015.
The videos above illustrate that the evaluation of the functional identification of a massively parallel neural circuit can be intuitively and in its entirety performed in the stimulus space. We quantitatively demonstrate how the quality of reconstruction of the encoded signal depends on the number of spikes used for identifying the circuit parameters.
The underlying massively parallel neural circuit for encoding color video consists of 30,000 IAF neurons with color sensitive receptive fields, covering a screen size of 160x90px. The IAF neurons all have the same parameters and are always assumed to be known. The underlying receptive fields are spatio-temporal non-separable in each of their color components and they resemble 2D Gabor filters spatially and rotate around their spatial center over time. However, the color receptive fields are assumed to be unknown and need to be identified. All the receptive fields are functionally identified using the same natural video, whose duration is up to 200 seconds. We identify the entire neural circuit based on 7 different settings. In one setting, each neuron's receptive field is identified using 1,000 spikes (measurements). In the other settings, each neuron's receptive field is identified using 2,000, 4,000, 6,000, 9,000, 13,000, 17,000 spikes (measurements), respectively.
A novel stimulus (the bee video shown here) encoded by the underlying circuit is then recovered. Instead of using the underlying circuit parameters, the identified circuit parameters are used in decoding. Note that the set of spikes used are still the ones generated by the same underlying circuit. The decoding quality then depends on how well the circuit is identified. The reconstructed video and the associated Signal-to-Noise Ratio (SNR) are shown in the above video for the circuit identified with different settings. As identification quality increases (more spikes are used), the quality of reconstruction converges to that of reconstruction using the underlying circuit parameters (known receptive fields/filters).