Neural decoding - Wikipedia
Neural signals for control of voluntary movement and design of neuromotor prosthetic devices
Excerpt from "Control Signals" p.1160 - Link
Decoding algorithms for extraction of signals from the brain - central/peripheral nervous system to deliver them to control devices for prosthetics
4. Decoding Algorithms: Principles
The decoding methods for use in neuromotor prostheses are the culmination of many years of basic research on the motor system.
4.1. Two learning machines
4.2. Discrete and continuous signals
4.3. Mapped variables
4.4. Control and meta-control
4.5. Human calibration
5. Decoding Algorithms: Examples
5.1. Population vector
5.2. Principle component analysis
5.3. Maximum likelihood estimation: discrete control
5.4. Linear filters
5.5. Adaptive neural networks
5.6. Feedback-driven models
Notes based on reference articles for neural decoding for prosthetics
The intent neural code for "I wish to move my arm" (e.g. to grasp an object) has been deciphered and is being read from the brain/nerves and used to activate prosthetics.
The sensory code for "I am currently grasping an object e.g. a glass of water" ("as I realize that there is contact with the object") is necessary for prosthetics.
In the absence of the contact stimuli one would have to constantly look at their arm to verify that they are actually grasping the object.
This has also been deciphered and is being applied or written on nerves/brain as input from the prosthetic device. We are therefore talking about "sensory neuroprostheses", as somatosensory feedback has recently been integrated.
"Continuous-time level-crossing sampling for efficient data representation and subsequent spike processing.
(1) We first compare signals (synthetic neural datasets) encoded with this technique against conventional sampling.
(2) We then show how such a representation can be directly exploited by extracting simple time domain features from the bitstream to perform neural spike sorting.
(3) The proposed method is implemented in a low power FPGA platform to demonstrate its hardware viability."
Digital Neural Circuits: From Ions to Networks
Thesis by Junwen Luo
"(...) If two such electric fields are applied at high frequencies that differ by a small amount, which corresponds to a low frequency that neurons can follow, neurons in the brain may be able to demodulate and follow the envelope modulation that results from the temporal interference between these two applied fields, and which oscillates at the difference frequency."
"We found that interferential stimulation with two sinusoids at 2.01 kHz and 2 kHz, resulting in a ΔEq envelope frequency of 10 Hz, was able to recruit neurons to fire at 10 Hz (Figure 1D), as efficaciously as direct 10 Hz stimulation (Figure 1E) that would be expected to broadly affect neural activity (Miranda et al., 2013)."
"If the amplitude of the envelope modulation reaches a maximum at a site deep in the brain, it might be possible to drive deep-lying neurons without recruiting overlying ones. We here test this concept, which we call temporal interference (TI) stimulation, by using computational modeling and phantom measurements, as well as electrophysiological measurements in vivo.
"We demonstrate the ability of temporal interference (TI) stimulation to mediate activation of hippocampal neurons without recruiting overlying cortical neurons and steerably probe motor cortex functionality without physically moving electrodes by altering the current magnitudes delivered to a fixed set of electrodes."
Researchers have discovered a way to remove specific fears from the brain, using a combination of artificial intelligence and brain scanning technology.
Induction of "incorrect decision" via cocaine and restauration by imposing "correct decision" neural code
An example provided by the following study will be presented briefly:
Hampson RE, Gerhardt GA, Marmarelis V, Song D, Opris I, Santos L, Berger TW, Deadwyler SA.
J Neural Eng. 2012 Oct;9(5):056012. Epub 2012 Sep 13.
PMID: 22976769 Free PMC Article
Let us consider the use case of restauring faulty decision making related to a visual stimulus.
Choose this experimental setting: Monkey that matches image to correct image gets sweet reward.
Which neurons do decision making from visual representations? This information is known: L2 neurons.
We would need to decipher the neural code they generate.
What are they connected to? L5 neurons.
Then proceed by deciphering the neural code of L2 and L5.
Start by measuring their neural signals. They are 1350 μm apart so you need to have two electrodes at this distance. You build an electrode array that fits that.
You obtain recordings for correct decision and incorrect decision. How can you induce an incorrect decision? By using a drug like cocaine. You then analyse the recordings and you extract patterns using a non-linear dynamic model (MIMO). You determine the code for correct decision and code for incorrect decision. Having identified these, when you detect incorrect decision code you apply the code associated with correct decision making.
In the image below, we have a neural column (blue neuron on top of red). We measure the blue (L2) and we detect that it has a weak code. We therefore proceed by imposing strong code on the red (L5). In a way we are modifying the output of L2.
Figure 1: Excerpt from figure 8 of the publication demonstrating detection using MIMO model of weak L2/3 code following cocaine exposure. Output of MIMO model is then utilized to stimulate L5 with a “strong code” L5 pattern associated with correct decision.
Just to have an idea of what neural code looks like, here is an image which shows L2 above (weak) and L5 below (strong) after imposing the model prediction.
Figure 2: Excerpt of figure 4 of the publication demonstrating recording of L2 representing "weak code" according to analysis with MIMO model and imposition of "strong code" to L5.
In general, let us assume that you record neural signals with electrodes. You will then need a pattern extractor, and therefore a neural network model for this. It is mentioned in the literature that “A number of neural network (NN) models have been developed for pattern recognition and classification since mid-1980s. They include the backpropagation NN, counterpropagation NN, adaptive conjugate gradient neural network algorithm, recurrent NN, radial basis function NN (RBNN), wavelet NN, spiking NN, neural dynamics models, neuro-genetic models, and the probabilistic neural network (PNN) (Specht, 1990). PNN is often used for classification purposes (Adeli and Panakkat, 2009).”
In this case, the authors used a "Multi-input/multi-output (MIMO) nonlinear dynamic model."
Scientists from the same group as above are performing clinical trials of their technology on humans.
Introduction to Neuronal Networks
Here is an interesting link that provides an introduction to Neurons and Neuronal Networks
A feature of the biological brain is that while communication uses digital techniques for fast signalling over all but the shortest distances, all of the processing in neurons
and synapses uses much more efficient analogue chemical techniques .