Neural Network Simulations

The major activity in the previous month was to set up a simulation environment to simulate a neural network which captures the features of the network we plan to culture and train . This will be used to study the behavior of the network , explore its computational properties and develop training protocols for
obtaining desired network behavior .

   Such a network set up using the software 'NEURON' . 1000 LIF neurons are connected via synapses which model STDP and input frequency dependent characteristics . STDP is a mechanism by which the strength of synaptic connection between the two neurons is modulated by their relative firing times . This mechanism is hypothesized to help in learning.  We try to exploit this feature to develop training protocols to teach the network to do what we want. Various parameters of the neuron , network and noise levels were set to match the reported value in previous works .

  Using this setup we trained the network to perform some simple tasks.  The next aim is to train the network to perform some moderately complex tasks.

Here is a link to the report which details some of the above tasks . And here is a presentation 

General Steps in experiments using multi-electrode-arrays

Some details regarding the steps in experiments using multi - electrode arrays .

Here is a link to the presentation that I made on this topic . 

Record

It is argued that the activity of the neurons are encoded in 'Action Potential / Spikes ' in electrical signals generated by them . The electrical signals are hence recorded using electrodes . These electrodes pick up activity from multiple neurons at various distances . Those from far neurons appear as  'noise' . Other sources of noise also affect recordings . The shape of the recorded 'spikes' depend on the number of factors like property of neuron and distance from electrode. When an external stimuli is applied (order of 100's of mV)  , this is also picked up by electrodes along with spikes(order of uV's). A huge amout of raw data is generated from these recordings . For example , recording from 60 electrodes at 24kHz with 12 bit resolution generates 2.16 Mb data per second . It becomes necessory to handle and store these .

Process

   We need to identify spikes in the recordings . Each recording may contain spikes from many neurons . At times we desire to seperate them and isolate their source neurons .  To this end , the following steps are involved .

Pre-Process .

   The recording contains spikes , noise and other artefacts . In order to isolate spikes generally some transformation is done to the recording and a threshold is applied to the result to say spike or no-spike . To be able to effectivly apply the threshold , the features of the recording that signify spike need to be enhanced relative to other noise. Several techniquies are done towards this end . These include Band pass filtering , pre-whitening ,line-noise removal,removing stimulus aretefacts , wavelet denoising etc . Some transformations like NEO(non-linear energy operator) are also sometimes used to enhance spike features .

Detect

   This involves applying a suitable threshold on the pre-processed data . Determing a suitable threshold has been a topic of number of papers in these area . Several ad-hoc techniques and some rooted on a mathematical basis exist . Estimating threshold involves estimating noise . Wavelet decomposition , Optimal Linear filtering , blind deconvolution etc are some of the other techniques used .

Validate

   Steps are taken to validate the detected spikes to reduce the number of false detection by observing the nature of signal around the time of detection . This involves using criteria like minimum Inter spike intervals , expected peaks etc .

The following steps are done if identifying spikes from individual neurons is of importance. It becomes relevent when doing connectivity analysis and studying learning  . This is referred to as the problem of spike sorting . A segment of data around the detected spike time is taken as the shape of spike ( the collection of amplitude of  voltages is called the feature-vector for the spike) . It is required that we identify the neuron to which this belongs to . Following steps are taken to achive this goal .

Aligning

   An initial step is to align the detected spike to other spikes for classification . In order to classify it becomes necessary to compare spike with a template . A measure of distance between them is used . For this to be satisfactory , similar spikes need to be aligned at their dominant features . The spikes get mis aligned due to finite sampling frequency , noise etc (example : peak of two segments may differ ) . Interpolation (using splines , fourier reconstruction) , using center of mass around peaks , correlation etc are some of the methods adopted .

Feature Extraction

      This involves identifying the features from the feature vector which would help to distinguish between different classes of spikes . Properties of the spikes such as height ,width , number of peaks etc are used at times. These may not be always satisfactory . Methods such as PCA(pinciple component analysis) , segmented PCA , Integral Transfrom , Wavelet Transform , LDA(linear discriminant analysis) , Phase space reconstruction etc are used to transfrom the feature vector so that the dominant and distinguishable features can be easily identifed .

Feature Selection

     Since the number of features may be large and classifying using all of them may be computationally difficult , the important features are identified . Fitting a model to large number of features would result in over fitting . Methods like using few principle components , Genetic algorithms , selecting the best basis / local discriminat basis  etc are usd to reduce the number of features .

Classification

        This stage involves identifying the number of different classes (different neurons) and then classifying spikes accordingly. This may be template based or a model based . One of the problems is to identify appropriate boundaries between classes . Methods like basyian decision , kNN,K-means,nearest neighbour , SOM , SVM , ANN,EM etc are used . Hirearchical classifiers are also used in some cases .

Analyse

   Once we have the time of spiking(firing) various neurons , we would like to make some guesses on what the network is saying . Some say the language is firing rate , for some it is the sequence  or it may be the instant itself . Statistical  methods like Cross-correlogram, (J)PSTH, MANOVA and Information theoretic measures like mutual information are used . Techniques like conenctivity matrix are used to infer their interconnectivity . Some measures like center of activity , average firing rate etc are used for spatail information . Ensemble analysis using graphy theory are proposed for such analysis.

Act 

   This involves generating appropriate signals for the task at hand. A mapping between the activity and the actuator commands is required . It involves indentifying input output pathways and  selecting the right electrodes for control . The mapping problem is attacked using  ANN's , Fuzzy systems or at times a linear mapping itself.

Feedback and Training

   This involves applying stimulus to the culture . The appropriate spatio-temporal pattern of stimulus is to be determined to achive the target . Tentaic stimuli are given to induce LTP/LTD and thus train the network . Identifying the effective stimulus is a problem to be studied . Analysis of connectivity and modulating it is one of the approaches used .

(Note : None of the images are mine, they have been picked up from various sources in the internet (Thank you all) . I have made some block diagrams)

Intro

                Here I intend to document the work done by me as a part of my Masters Project at Indian Institute of Science , Bangalore . The project is titled 'Brain in a Box' . The aim of this project to exploit  the computational power of a neural network to do a task . We intend to culture neurons in a dish with a micro electrode array and then use this system for our studies . It is hoped that it would further our understanding about the behavior of an ensemble of neurons .

          The project kicks off with literature survey . Here is a summery of  what I could gather going through various papers on the topic .


Here is a link to the presentation that I made on this topic .

Motivation 

        The Brain . How it works ? This is a question that continues to fascinate a number of researchers working in  a number of fields . The answer is sought out for a number of purposes . To understand why we are the way we are . To help people suffering from various disorders  . To improve our capacity  . To engineer systems that can mimic arguably the best of them all .

           Where  does 'I' come from ? How do we see , hear , feel , think ? We would like to help people with disabilities . To make the blind see , the deaf hear , give the power to speak ,  provide limbs to those who don't have, cure epilepsy . We would like to play God . We demand a better way to communicate with our salves - the computers . We want our systems to have some of our capacities . To learn , to adapt , to be resilient , to execute complex tasks without detailed instructions,to comprehend.
           We have been able to map  various paths of action in the brain . We understand some mechanisms that help us see and hear , to send commands to our limbs . Yet this is at a basic level . We have made devices that connect to brain , artificial limbs that move at our command . We have been able to train monkeys to move levers by thought . These are still nowhere near human abilities . We have made very powerful computers. Yet these are not a millionth of our capacities and nowhere near our efficiency in doing so.
   
      These are some of the things that motivate people to look for the answer to  question . The Brain . How it works ?

What we do 

             We know(/think?) that the 'computational' part of the brain is made of cells called neurons . They are interconnected and play an orchestra that results in everything . The neuron (cell) has mainly three parts morphologically . The cell body or soma , the dendrites with synapses to which other neurons connect to and the axon which helps the neuron to connect to others. They work by moving a number of chemical ions / molecules / complex compounds in and out of the cell . Observing these movements give us some idea regarding what is happening . We have devised many ways to do this . Since we are very good with electricity , it is the major source of our information . Movement of ions cause electric currents . These currents can be detected and recorded using electrodes and appropriate electronics . We try to infer things based on these recordings .

Recording

            For electrical recordings , we  have chiefly two methods  intracellular and extracellular . Intracellular recordings allow us to record the activities of a single neuron in detail . Such recordings have helped in understanding the mechanisms of operation of single neurons in great detail .  Extracellular recordings are done by inserting the electrodes into a medium containing the neurons  . For example , inserting an electrode into the brain (of cats , monkeys , insects ) . These electrodes then pick up activities from number a of neurons . We also do recordings by placing electrodes on the scalp which pick up some of the underlying electrical activity . We also have multiple electrodes inserted into a region / over an area . This helps us pick up activity from a number of different sites and thus gain more information about the working of an ensemble of neurons .

          Recordings from a single neurons have helped us to understand the functioning of these basic units . However , we still lack ways to understand the properties that emerge when these from a networks . We hypothesize that it is these emergent properties that give us all our power . Computational models have helped in this regard . Several models have been tried out to explain various properties . Directly recording and analysing activities from the brain is a very challenging task .

          Towards this end , one approach has been to culture neurons in a dish and allow them to form networks . The dish has an array of electrodes .Properties of these networks are then studied by giving stimuli and recording responses. It has been argued that giving a body to this brain would help us understand more . A number of interesting projects have been done towards this end . People have connected these neurons to a robot with sensors . The neurons drive the robot and get signals from the sensor. The network is trained to move the robot in such  a way to avoid obstacles detected by sensors . There has been efforts to study the capability of these networks to learn a task (plasticity) and to retain .
       

General Steps in experiments using multi-electrode-arrays

       Here I briefly mention the techniques that are used . I hope to detail them as and when we use these techniques in our project

Record

  We record the activity from the culture to make some inferences

Process

   We need to identify spikes in the recorded data since we believe this contains the information about neural processing . Hence the next stage involves processing the data to identify spike from the neurons and time it occurred .

Analyse

   Next stage involves analysing the timing of spikes from various neurons to infer something about the activity of the network

Act

   This involves generating an output corresponding to the task we intend the network to perform . This involves a mapping from the inferred activity of the network. For example the task may be to drive the motors of a robot to avoid obstacles . Then the mapping would be from the activity to motor speeds .

Feedback

   For the culture to exhibit a 'behaviour' it must be able to sense the environment around . Hence the information from the outside world must be given as stimuli to network . This would be based of the changed in environment as a result of the action of network . For example , obstacle is nearing / going away .

Train

  This involves training the network towards a goal if the exhibited behaviour is not in agreement with the desired behaviour . We would like to train the network to move away from obstacle if it moves towards it .


    The next post adds some details to the above mentioned outline of steps .

(Note : None of the images are mine, they have been picked up from various sources in the Internet  Thank you all . I have made some block diagrams)