 Welcome to today's combined INCF and Society for Physiology seminar series. So today we are really happy to have Professor Hagail Bergman here, he's from, and he's a professor in physiology at the Hebrew University. And he's a really good representative both for physiology but also for neuroinformatics. He has a wide toolbox which he uses, he uses multi-electrode in vivo recordings and data analysis of these parallel spike planes and there's a lot of data, I'm afraid. And he's also doing simulations and modeling and he has also developed disease models for Parkinson's disease in my case. So earlier in life in the late 80s and beginning of the 90s, he was a popular course with Moshe Avelis and then Malon Delong which both are famous. And I think today Hagail is not well known for having figured out that he can sleep some ceramic nucleus that you can, if you inhibit that, you can ameliorate the Parkinson's disease symptoms. So I think this is sort of something that most people know. That was actually a science publication that has been cited. I don't know how many times but a lot of times. More than one thousand. Yeah, it's really amazing. We stopped counting the two one thousand. Good goal too. Yeah, so please go ahead. Thank you. Thank you, Jeanette. Thank you Stan and Gilly and everybody else for coming in the noon to hear me. I feel a little bit too equipped but I'll try to survive. I'll try to speak about the computational physiology of the basal ganglia. That is, what do the basal ganglia do in normal life? What is getting wrong in the basal ganglia when we are getting Parkinson's disease? So the first part, the normal basal ganglia and Parkinson's disease I will start with our experiment in the monkey. Then I will move a little bit to our physiological recording in the human patient undergoing deeper stimulation in order to treat their advanced Parkinson's disease. And in the end I will try to come back and use disinformation that we are learning from the abnormal basal ganglia in order to say what are the best methods for the future to treat the human patient with advanced disorder of the basal ganglia. I apologize, I will try to finish in time but as always when I'm coming to speak in front of Stan I'm trying to Jeanette and Gilly, all my friends, I'm trying to show them that we are working from time to time. So I will try to do it in time. Usually when we are speaking about the agent that is living in the world and the agent can be you, can be your dog and can be your computer or robot that you will send outside. We can make a long story short by saying that the agent is making an action on the world. For example now for me it is I am the agent, you are the world so I'm speaking to you and as a return the world is changing and I am the agent as I'm getting the new state and reward. And this reward can be positive if I see that you are happy and can be negative if I see that you are really bored. And the bottom line is that we summarize this that the goal of the agent is to optimize the behavioral policies that is the mapping between the state and the action such to maximize the cumulative reward that you will get from the environment. So we are looking for a behavioral policy that is mapping between current state and action that will maximize the cumulative reward and again reward can be either positive if you are happy or negative if you get punishment. And therefore you are trying to maximize your pleasure. And the major problem is that finding this optimal behavioral policy, at least probably you know from yourself, it is not an easy problem. There is no old teacher that tells us what we should do, what is the best behavioral policy. The world environment is changing, the rules are changing from time to time and even if we know the rules they are noisy. So it is not fixed rule that we have to deal with them. And finally the feedback that we are getting from the environment is not instructive. Again this is coming back to the teacher that is not an old teacher and nobody tells us go to Stockholm or go to Jerusalem We are just getting feedback which is color, we are happy or not happy and many times it is delayed. So now Janet know that she is done mistake in inviting me to come here but if she would like to go back to the origin, the decision to do it was done three, six months ago and therefore the feedback is delayed and it is not complete. So this problem is not only for us, it is also in the field of machine learning and usually we say that to make a long story short we are having today's solution for this common problem. And the first solution is that we should find a trade-off between exploitation and exploration. So we should not all the time exploit everything that we have done because it could be that real good neuroscience is done in Jerusalem and not in Europe. So if you never explore Jerusalem you will never find if there is a good neuroscience over there. And on the other side you cannot explore all the time because this is too risky. So the big question that we looking for optimal behavioral policy without a teacher that will tell us that know everything and will tell us what is the optimal behavioral policy is to find the optimal trade-off between exploration and exploitation. And the way that we can do it is by dividing our agent into two parts. So this is the agent that we have done over here and we can divide it into two parts. One part is the actor and the actor is a very, very simple part of the agent. This is where the mapping between the current state and the action is done. So this is the actor part, it could be deterministic mapping, it could be probabilistic mapping we will come to some kind of mapping that we are doing between state and action but this is what is done in the actor. But then actually we have the critic or the teacher of the system and what this critic is calculating all the time is the mismatch between prediction and reality. So you don't have a priori knowledge or you have a limited a priori knowledge. So let's say that you don't have any a priori knowledge so you start with prediction 50-50%, it is going to be a good seminar or a bad seminar, 50-50. And then two minutes later you are telling yourself this is going to be a bad seminar. So reality is worse than your prediction. Next time you will not come and so on and so on. So all the time and this is done on continuous temporal manner you compare your prediction to reality and using the temporal difference error that calculates the mismatch between your prediction and reality you change your behavioural policy. For example, if reality is better than your prediction you reinforce the doing that you have done up to now or if it is negative you don't do it and of course you update your kind of prediction. So this is the way that we can overcome, okay? Achieve in automatic way this goal of optimizing behavioural policy in spite of all this fact that we don't have the feedback that we are getting from the environment is incomplete and non-instructive and the environment is changing and noisy. So this is the solution in the machine learning field but eventually we can now look and see that this was not invented by the people from computer science it was invented by biology before and if we are looking at this agent we can replace it with the loop of the bubble ganglia network, okay? So now this is the agent, this is all cortical area and coding your current state. So this is your auditory system encoding the voice that you hear now it is your vision system encoding the vision that you are looking it is your frontal cortex or hippocampus encoding your memory encoding your motivation. The current state, your current state is encoded in this part of the, all part of the cortex. Then we'll go through the bubble ganglia starting in the input stage of the bubble ganglia, the sub-tylamic nucleus and the striatum and we'll come to it going to the external segment of the globospidus output structure of the bubble ganglia and then from the output structure of the bubble ganglia we are going up to the motor centre of the brain either in the frontal cortex or in the brainstem. So eventually we have this loop of action that is coming from the motor centre of the brain and the world is changing and the state and reward are giving back to the cortex. So this is the main axis or the actor of the bubble ganglia network and you can see that from my point of view the bubble ganglia network is all the brain almost, okay? So all cortical area including amygdala and the hippocampus are projecting to the bubble ganglia and therefore this is the most part of the brain that I'm putting in the bubble ganglia network. I'm putting them in the bubble ganglia network because there is another part that it is mainly in the bubble ganglia and this is the neuromodulator or the critic and the neuromodulator it is clearly the dopamine but also the serotonin and the cholinergic interneuron that are strongly located here inside the striatum and from our point of view these are the critic or the teacher of the bubble ganglia and therefore we see that to make a long story short we have exactly the same architecture as we have over here. A critic actor, we have the bubble ganglia main axis and we have the bubble ganglia neuromodulator who are the critic that enable us to find the optimal behavioral policy. So this is what we are having in mind about the model and the question is of course can we support this thing by physiological finding. So the way that we have done it we are recording electrophysiological activity mainly spikes but also local field potential from behaving monkeys. The task is quite simple now in this day for us is we have cues, fractal cues that are different from monkey to monkey and so on and so on. One of the fractal cues predict food for the monkey that the monkey like, the other fractal cues predict air puff to the eyes that the monkey don't like and in this case this fractal cue predict the neutral outcome so no food and no air puff. And clearly the monkey and these cues are given for two seconds and then the outcome is coming and then intertrial interval, long intertrial interval and you can see that the monkey clearly understand the meaning of the cues. So this is for example when we give the cues a predatory word you see that the monkey start licking even during the time of the cue and then when the food was coming the monkey was licking a lot. And when we give the cues that predict aversive outcome, air puff you see that the monkey has the monkey blink even before the air puff is coming and avoid the air puff to the eye and when we give the neutral cue the monkey neither lick, neither blink. So the monkey, it is very easy for the monkey to learn this task and what we've done over the last several years is that we played with this task and modulated it. So for example in one set of experiments we change the probability of the outcome. So we give different cues and tell the monkey the probability that you'll get food is one, two-third or third and so on for the other cue. In another set we change the outcome intensity. This is the human primate and this is the non-human primate that's participating in the study. So in one set we change the outcome intensity. That is how much food or what is the intensity of the air puff and in another set that we are now doing we change the outcome delay. That is if the outcome will arrive immediately after the cue or later. And the recording method is quite simple. Let's put it this way. What we do is that we are putting eight microelectrons that we can manipulate each one of them separately in order to find as many as possible neurons in our recording target. We are quantifying the isolation quality so we are very proud that we are not saying only it is an isolated cell but we have objective quantification of the isolation quality and our database is composed of a unit that we can give you the number. This is only a unit with isolation quality usually bigger than 0.8 on a scale of 0 to 1. So the good unit but you don't have to trust us you can go back to the original data. And this is for example of 6-electrode recording of tonically active neuron in the striatum that usually fires something like 5 to 10 spikes per second. This is example of 8-electrode in the external segment of the globus pallidus much faster area. So each line is the original spike that we are seeing in this area and this you may see it is much higher frequency 50 to 60 spikes per second. And again the target of the recording so we've done all the record and we verify using MRI and again details that you can find in the paper that you publish. And the major idea is that we have again the same task. So one trial that was given in every monkey and we recorded the activity of the dopaminergic neuron of the tonically active neuron or the cholinergic interneuron of the striatum in the posterior putamen, GPI internal segment of the globus pallidus, suspension nigraticulata in this monkey. Over here we recorded activity in the anterior singulate cortex, anterior putamen coated ventral striatum of medium spiny tonically active neuron and fast spiking interneuron and GPI. So we always have GPI because GPI is so nice. So for us we record all the time GPI and our assumption that if the GPI is the same we can merge everything else in the data and now we are recording from the sub-talamic nucleus and the GPI and striatum in the current project. And then we look at the activity of neuron and the bottom line is let's start looking at this discrimination of critical teacher and main axis. So we recorded the activity of the teacher and the teacher in our end are the dopaminergic neuron in the substantia nigra pars compacta or the cholinergic interneuron in the striatum, the tonically active neuron. And over here you see the average PSTH, this time to the Q, so this is 0.8 second and you see the response of the dopaminergic neuron to the Q that predict reward, to the Q that predict air puff in red and in green to the Q that predict a neutral outcome. And you may see that this is almost similar to as predicted, dopaminergic neuron encode better the future rewarding event than the aversive event. When the outcome arrived, okay, when the monkey got the actual reward there was much bigger response and then there was also response to the aversive event. But you see over here that when this was the probabilistic task so when we tell the monkey there is probability of one-third or two-third that you will get reward or air puff in one-third or two-third of the cases, the reward of the air puff were omitted. And you see over here the response of the dopaminergic neuron to the case that air puff was omitted in red to the case that reward was omitted in blue and to the neutral outcome, so again nothing happened. And you may see that dopaminergic neuron do not encode the reward or the air puff emission. So they do not encode, they are very similar again, this is the average of one-third of the dopaminergic neuron so there is no encoding over here, okay? So you may see this is not a good teacher, okay? A good teacher should tell me if reward a positive outcome was omitted or aversive outcome was omitted, okay? And what you may see over here is that luckily enough the buzzer ganglia have more than one teacher, okay? And if we are looking at the other teacher, the cholinergic teacher, and again coming to the classical buzzer ganglia literature who is dopaminergic cholinergic balance, okay? So the cholinergic teacher is clearly playing a major role in the buzzer ganglia circuity. You see that over here when the queue was coming the cholinergic teacher was very inefficient, okay? In our end, okay? So you see the red, the blue and the green curves are completely overlapping telling us that the cholinergic interneuron don't discriminate between the queues that predict outcome or predict positive outcome or negative outcome unlike the dopaminergic. But when you are looking for, when you are going to the outcome you see the response of the cholinergic interneuron for the actual air puff and for the reward you look over here for the omission, okay? So the cholinergic interneuron this is for the neutral outcome in green this is for the omission of the air puff in red and this is for the omission of the reward. So the clear-take-home message is that if we are looking at the outcome omission the times are doing much better than the dopaminergic neuron. If you are looking over here the cholinergic teacher is doing much better than the cholinergic teacher. So luckily enough we don't have only one teacher in the Basel Ganglia Web as a teacher. Most of you that know the Basel Ganglia literature should be surprised, okay? Because most of you probably are very aware of this classical description of dopaminergic neuron by Wolfram Schulz, okay? That tell us that dopaminergic neuron encode a positive outcome they encode the prediction of this positive outcome so if we train the monkey that this condition is stimuli-predictory reward you get a very strong response over here and they encode the omission of the positive outcome by depression of their activity. So this is the classical description of Wolfram Schulz that everybody is very, very happy because they tell us that dopaminergic neuron encode both the positive and the negative mismatch. And clearly in our end we don't see it, okay? So you don't see this very nice decrease of activity when we omitted the food outcome, okay? You see, it is the same. We are getting very, very similar response when we omit the reward outcome and the air puff outcome and this is actually going over here, okay? So we tell the monkey you are going to get food or air puff to the eye, okay? Clearly the monkey don't like to the air puff to the eye. So looking at Schulz's classical data one would assume that when I tell the monkey that you are going to get air puff to the eye there will be a decrease in the activity, okay? But what we have learned is that the response of the battle ganglia teacher is not dependent on the single trial, it's depending on the context, okay? And I would like to show you over here in our recording of the activity of the cholinergic interneural, okay? So this recording of the cholinergic interneural was done in the probabilistic task, okay? So you see it over here, for the queue and for the outcome, okay? So you see this and this over here for the two monkeys that were engaged in the probabilistic task. But over here you see the response only to the queue is probability one, okay? So we pick up only the trial that we give exactly the same trials in the second task with the dollar context. That is when we play with the intensity but not with the probability. So you see the response of the cholinergic neuron to exactly the same task but a different context. And you may see that now, okay, if you are looking at the response it is switch over, okay? The bigger responses of the cholinergic interneural is in the dollar context, in the intensity context is now to the queue and not to the outcome. So this mismatch between Wolfram Schuler result and our result over here in the dopaminergic and over here is a result of the complex teaching message that the Baselgangian or modulator are giving us which is depend on many, many things. It is not only single trial, it is the context that the monkey is living inside and it reflect many, many other things. So we cannot only look on the response to say what do these neuron are encoding. And nevertheless we have been interested to know if one can look at other aspect and then to say, yes, these are critic, these are teacher and these are part of the main axis. So over here we just look at the temporal pattern so this is now for the two seconds of the queue these are the very short response of the Baselgangian teacher and indeed as we expect from a teacher that encode the prediction error. Prediction error is a temporary, is a derivative, okay? Jeanette knew at the beginning that she was done mistake when she invited me but now, okay, this is over, okay? So now she is over here, okay? So there is no, no anymore error, okay? So this is for the, the short, the short, the, the short response are okay for the neuron modulator and you may see that when we are looking at the main axis of the Baselgangian, external segment of the globus pyridus, internal segment, a sub function nigareticulator we see very nice sustained response over two seconds. We see it also when we are looking at the other neuron in the main axis so this is from the anterior singulate cortex this is the striatum, in the striatum that the medium, the medium spine in neuron and this is again from another monkey in the external segment of the globus pyridus. Again, I don't have time to, to go into it but this was published, this sustained activity in the striatum is not that every neuron in the striatum have a long sustained activity but we have different cluster, a neuron with very sharp response with medium and very long response but as overall we are getting this very nice sustained response. Also when we are looking at the GP, at the globus pyridus and now with the task that we have delayed the class only for two seconds and delay the last for eight seconds you see that the globus pyridus neuron encode the, all over the delay duration for eight seconds as we are waiting for neuron as part of the active, active part of the basal ganglia. And finally we look for another issue that we have predicted that will be different between the critic and the main axis of the basal ganglia and here we look for the question how much the critic are synchronized. So we would like the teacher to be synchronized. You don't like one of your teacher tell you run away or run to the right and another teacher tell you run to the left. You would like a teacher to give you the same message. The way that we can do it to look for it is by looking at the signal correlation that is we have the recording of more than one neuron let's say of pair of neuron and we have the vector of the response and we can assess the correlation between the response vector of the two neuron. And if like over here, you see it for the basal ganglia teacher the dopaminergic and the homeinergic interneuron if the distribution of the correlation the signal correlation is shifted to the right you may say that the teacher are synchronized as you may see also in the spike to spike cross correlation matrix. But if you are looking at the neuron at the main axis then you see that there is a very broad distribution of signal correlation and the average of the signal correlation is around zero and not shifted to the right as it is for the neuron over here. I don't have time to go to it this is part of the old world that we've done regarding temporal modulation of the signal correlation this is just now impressed in the journal of neuroscience but to say that the signal correlation is much bigger issue and we'll go to it sometime when I'll have time to it. So coming back to the model of the basal ganglia to start with this idea the basal ganglia we are having the main axis of the basal ganglia and we are having the world and I would say ten years ago we've been thinking only on one teacher of the basal ganglia this is the dopamine teacher that encode the positive or the negative prediction of pleasure but today we'd like to believe that we are having more than one axis of our emotion it is not that it is only pleasure that can be positive or negative this is with a smile I'm saying this is good for the American but we are trying to be European a little bit more complex so let's say that we have two axis and these two axis but again two axis is one too many so I'm not claiming that it is exactly two axis and we are having the cost and the gain and they are not the same so cost is not minus gain and we are trying to optimize our behavior that is we are trying to minimize the cost and to maximize the gain and again knowing from economics we know that there is the Pareto Frontier which is depending on the statistic of the environment and clearly if we are here we are paying this cost and getting only this gain we should go up to the Pareto Frontier so this is our optimal behavior is somewhere on this but where is the optimal behavior this is an open question and this is not clear and this is the issue of multi-objective optimization rather than single objective optimization as was done in previous reinforcement learning model and when and again the mathematics is over here so everybody can go over it but when we have done this kind of multi-objective optimization for these two parameters we found the solution that the probability of action is behaving as a softmax function and the idea with softmax action is the probability of action J is depending of the value of this action J or what we are expecting to get but also on the temperature and the temperature is let's say or the pseudo temperature in the Gibbs-Donnan distribution but this is exactly what you can think about it you go to south Italy you have high temperature you go to Milano you have low temperature you go to Sweden you have minus temperature and I'm speaking about the people so with a smile but the point is that the temperature is telling us where we'd like to be on this Pareto Frontier so let's say that we are now in Las Vegas we have high temperature so we don't care the probability of our action doesn't depend on the outcome because we just came to enjoy so this is the softmax with very very high temperature then we reduce the temperature very very strongly and we are getting to the situation that we can select only the greedy solution so all the time we will not explore over here we are just doing exploration this is one extreme this is the other extreme we just exploit we take the greedy we do only the actions that we believe that will maximize our outcome and somewhere in the middle temperature we will get, of course, this should be a kind of sigmoid but the point is that this will be something which is like probability matching so we will do more the action that should probably lead to higher value but from time to time we will explore so our behavior is really a function of our temperature and you can modulate it according to the state, according to your need but the point is that now with this single parameter or the softmax equation we can really see the diversity of our behavior from gambling to probability matching to a highly greedy solution and when we are coming now to the basal ganglia we can really have it now back to the basal ganglia so from our point of view the neuromodulator of the basal ganglia the dopamine, the acetylcholine, the serotonin do not only affect the corticostreatal plasticity they first affect the excitability of the striatal neuron and therefore when you inject apomorphin to a Parkinsonian patient you get this patient re-behaving in less than one minute so dopamine agonist, ultrafast dopamine agonist restore almost normal behavior in less than a minute so really what we need is that we have this neuromodulator of the basal ganglia acting not only on the efficacy of the sign-ups between the cortex and the stratum they are working on the pseudo-temperature which is now this is the excitability of the striatal neuron but clearly they are also working on the stratal excitability and they are having this state-to-action association and looking for example on the difference between dopamine and 5-HD so dopamine encode positive mismatch you enter to the room and you got chocolate so you increase the temperature because you would like to take the chocolate and also you increase the efficacy of this connection so next time you will enter the room 5-HD you enter the room and there was a lion in the room you run away, you need to increase the temperature you don't like to freeze, you have to run away so even if it is a prediction of a versus event you need to increase the temperature but next time better not enter to this room so this is the mapping of this multi-objective optimization top-down model into the basal ganglia circuity and again we can speak a lot about it and still open and we call it the reinforcement driven multi-objective optimization and there are many many open questions first of all there are too many open parameters it is not that this is the only solution but we believe that life are more complicated than one axis and therefore we should go into this model so I am done with the first part of my talk and I would like to go now to the second part and I apologize because I have to switch gear and I have to move to second third gear but the point is that for us it is interesting and it is important to understand the basal ganglia because in many types of human disease we are getting depletion or a malfunction of the critic of the basal ganglia and we believe for example this depletion of dopamine in the Parkinson disease is leading to malfunction of the neural activity in the main axis of the basal ganglia so what are the changes that we can see in the basal ganglia clearly we can see changes in the firing rate decreasing the firing rate of the external segment of the globus pallidus increasing the sub-talamic nucleus and the GPI and therefore we inactivate the sub-talamic nucleus or the GPI in order to cure Parkinson disease so this is old story we see oscillations that appear in the main axis of the basal ganglia we see huge amount of synchronization so this is 7 electrode recording and you see a very strong, very long effect to second of synchronized activity and finally we can see synchronous oscillation so this is 2 and 3 so this is the flat cross correlation matrix in the normal monkey differ the auto correlation and this is the auto correlation in the MPTP monkey after dopamine depletion MPTP monkey monkey that we gave him the neurotoxin MPTP is Parkinsonia now so this are the auto correlation you see the oscillatory activity over here and this are the cross correlation, so synchronous oscillation so this is in the monkey but we can also look at the human patient and we are lucky in having a very active program for Dibran stimulation in our center and when we put our electrode trying to locate mainly the subthalamic nucleus for human patient with Parkinson disease we are doing electrophysiological recording in order to define the border to tell us what is the optimal place to look for to put our electrode and when we are looking at the activity that we can find in the subthalamic nucleus so this is 3 neurons so we are getting very nice recording of course these are best typical example but never the less ok we are getting very nice recording this is 3 seconds so you may see that this is over here we are having 1, 2, 3, 4, 5 so this is 5 hertz and when we are looking at the power spectrum you see over here this is a peak at 5 hertz this is at about 10 and this is at the better range 22, 30 hertz what we can do, what we are doing in the operating room is that we are looking at the spectrogram so we are usually starting when we are going to the subthalamic nucleus so these are different 5 patients that I randomly collected a few days ago so we start 10 mm above where the subthalamic nucleus should be we are going all the way in the internal capsule then we are entered to the subthalamic nucleus and we can look at the spectrogram so over here it is again the distance 10 to minus 2 in these cases this is the frequency 3 to 100 and over here in red telling us that we did see a lot of oscillation in the tremor range this is more common we see in the better range this patient of red and the different subthalamic this is a patient with both beta and tremor frequency so this is all from the subthalamic nucleus this is from the globus pallidus GPE, GPI, GPUC oscillation mainly in the GPI in our trajectory I will not speak about the GPGP I will speak only about the subthalamic nucleus so to make a long story short when we calculate something like more over than 300 penetration into the subthalamic nucleus and we put, we normalize it but the point that we enter to the subthalamic nucleus the point that we have been outside of the subthalamic nucleus subthalamic nucleus is somewhere between let's say 4 to 7 mm in our hand so we just normalize it all over the patient so this is distancing the subthalamic nucleus and this is again frequency 3 to 100 and red mean high energy you see that in the dorsolateral because we are coming this way in the dorsolateral part of the subthalamic nucleus there are tremor frequency oscillation and beta but not in the ventromedial part of the subthalamic nucleus and this is me inside the deviation I'm mad but we don't care now too much and the point that we ask ourselves is if this gradual decline is what we see in every patient or it is result of different span of the oscillatory area in different patient and the bottom line was that no, it is in some patient we see very short oscillatory region in some patient in medium is longer and we have done automatic system using the hidden mark of model that detect it so we are getting it now in real time so we can see that in each patient we can divide the subthalamic nucleus into the dorsolateral oscillatory region and the ventromedial non-oscillatory region and indeed we see now that if we are looking at the spectrum in the dorsolateral oscillatory region versus the spectrum in the ventromedial there are completely different this is the difference between them and if you are looking at the intensity of the oscillation so we have a score for weak oscillatory activity strong and even stronger oscillatory activity so you may see that when we are looking at really big number almost 4000 recording from the subthalamic nucleus of our patient you see that the score for strong oscillation in the dorsolateral are much bigger than in the ventromedial and the most important issue is that this oscillation can tell us also about other things so when we are putting our two electrodes in the oscillatory region they are synchronized so this is the coherence function this is the intensity so if the two electrodes are in the dorsolateral oscillatory region we see a lot of coherent oscillation but if one electrode is over there one electrode is over there the coherence is flat and if the two electrodes are in the ventromedial it is still flat the intensity is different I will not go into it we can show that the synchronization is not an epic phenomena of oscillation at similar frequency by doing shuffling and again I will skip it but the most important TECO message is that we found that there is a correlation between the length of the dorsolateral oscillatory region in our surgery and the improvement that we are getting of the patient after surgery so today if we see such a recording in the operating room and condition allow we will try to find a better place to put our electrodes so this tell us now that we are better looking for this signature in the human patient and not to this signature and finally just to say that what is the different role we believe that the dorsolateral oscillatory region is the motor part of the subtalamic nucleus but the ventromedial at least does on the right side is the one which is related to the limbic part of the subtalamic nucleus so we give our patient emotional voice and ask them to recognize if this is a positive emotion or negative emotion you see there is no response in the left or right dorsolateral oscillatory region and there is a very very strong response to the emotional voice in the right ventromedial subtalamic nucleus the left we believe is for speech but again this is still under reconstruction so having done all this finding the bad guys of Parkinson disease can we use it and today when we are doing Dibran stimulation surgery the way that we are doing it is the pacemaker in the patient's chest and the patient is coming every 3-6 months to the clinic the neurologist is checking his neurological symptoms and adjust the parameter of the stimulation in order to obtain optimal therapeutic effect but doing it every 3-6 months it does not make sense it is like changing the eating or the cooling system over here only 3-6 months clearly neural activity is much more dynamic clearly the symptoms are much more dynamic so the way that we have done it we say if the basal ganglia are clever enough to use machine learning tricks in order to find the optimal behavioral policy let's see if we can do the same trick in order to cure the basal ganglia we started by using a closed loop a closed loop system that detects the abnormal oscillatory so you see this is recording for the monkey this is what we call normal recording and over here oscillation starts so we can detect disoscillation and whenever there is oscillatory activity we stimulate and again to make a long story short there are many many many stories over there it is that if we are looking at the amount of movement that the monkey was doing before any DBS this is with standard DBS and this is with closed loop DBS and these are many many controls that we've done and the most important is that we have shown over here that the basal ganglia can be observed and can be controlled so now it is only a matter for if we have this characteristic of observability and controllability we can now look for the optimal way in order to give the therapeutic to our patient I will not go into the detail but again part of it and everything is in this neuron paper we can show that again one can speak who is the bad guy the change in the firing rate the change in the oscillatory in the firing pattern or the synchronization we can show that it is the synchronous oscillation that are the bad guys in one of the parameters of the closed loop so this is my final my final slide what I was trying to tell you that we can look today the basal ganglia interaction with the world as an actor-critic system we believe that a dopamine depletion and the other disease of the basal ganglia like schizophrenia, like depression are caused by malfunction of the critic leading to abnormal activity in the main axis but the major take home message that we are trying to say that up to now we've been working when we've tried to help our patient we've been working on this axis we've tried dopamine replacement therapy stem cell therapy we've been working on this axis I do believe that the future is going over here looking at the spiking activity in the main axis of the basal ganglia doing closed loop very delicate manipulation of the spiking activity in the basal ganglia and getting optimal therapy for the different basal ganglia disorder starting with Parkinson's but maybe in the far future also the negative symptoms of schizophrenia that we believe that are related to disorder of this and I'd just like to say that if you're interested to hear more about the basal ganglia this merch will have the international basal ganglia 11 meetings in a lot the temperature in a lot in merch is 25 degree and above the sea is blue as you may see all the time there are five days of rain all over the year so clearly it is different this is a lot this is the red sea and this is the princess hotel you see just over the sea a lot is located over here this is the mediterran and I promise you a lot of dopaminergic spritz if you will come to this meeting thank you very much