How does Neuralink (actually) work?

[@konstantinarnold6754]

Hey dude, great video, i really appreciate you breaking it down so a lot of people can understand it! i wanted to ask if you created the graphs yourself or if not, where I coukd find those images. I am doing a presentation for my finals on Neuralink and those graphs would be super helpful to visualize the process.

[@ericinthewild]

Hey Konstatin, I did make those plots. They are of actual data from my own arm muscles during an EMG task. I may use them in again in a publication. So instead of using my plots, I wanted to show you an easy way to make your own! Here is a Jupyter Notebook with some code that will help you generate blobs! drive.google.com/file/d/1kn4ITlFBdtjCegT0_gZP3EWDMInRbl_x/view?usp=sharing

[@konstantinarnold6754]

@@ericinthewild Thanks alot! Is there also a file like this to create a 3 dimensional plot?

[@Dmadiroe]

Hey @eric james I really enjoyed your video. Do you think that the data the machine learning program has learned could be transmitted to another monkeys brain or it has to be trained with each new monkey? Greetings from Paris, France.

[@ericinthewild]

Hi @webigfamily.online, this is a really great question! It was a nice thought exercise to construct this answer for you. I tried to be thorough without going over the top ;). To answer this core question: "Could the model from Monkey A be used for Monkey B", we would have in practice have to better specify many things, but in theory: Yes, it could! To give you an idea of why I say this, let's take an 'extreme' example. In this case, Monkey B is a perfect clone of Monkey A. (I will label them MA and MB from now on). Since MA and MB have the EXACT same brain structure, the only thing that we would have to control now, is that the electrode is implanted in EXACTLY the same place. If this were the case, and if we could assume that no differences in brain activity or structures arose during the lives of each of the monkeys, then the model would transfer perfectly from one monkey to the other monkey. However, this is likely to never be a case in reality, and be therefore merely a thought exercise. Let's take the other extreme, two monkeys MC and MD, where for some reason MD is actually missing the brain region that the model was trained on for MC. In this case, no matter what we do, the patterns of activity will be different when each monkey does the task, and therefore the model will not generalize well and we will likely receive a chance level or even a worse-than-chance level accuracy. Now, let's take a more realistic example, two normal monkeys ME and MF. In reality, although we might get closer and closer to an "exact" placement of the implant in each monkey (as certain structures in the brain are rather identifiable and present throughout brains, search for 'Brodmann Areas'), the characteristics of the individual neurons within these regions will almost certainly be different. So, first we try to minimize the location error of the implant; but undoubtedly we will still have some error in model transfer due to this imprecision. With this minimized, we still have the difference in neuronal properties between the two monkeys. Since larger brain networks are roughly defined by anatomy from birth (i.e. think of the frame of a house under construction, at this point only with timber), there will be general consistency in the regions (i.e. there will be a "bedroom" over there). So, given this consistency, if for example, I wanted to make a model that detects simply whether or not there should be a movement at all, then this might generalize at an acceptable level. However, if we want to get into the details, and perhaps (as in the video) explore the case of several different classes of movements or directions, we have to imagine that the brain is quite flexible, and the responsible neurons may arrange themselves over time according to each monkey's own individual life experiences (i.e. it is very likely the intricate details of one monkey's "bedroom" are different than the other monkey's "bedroom"). With this said, the real question again comes down to minimizing error. A critical point to make here, is that we could use the model from ME to begin training MFs model. Since the other monkey's model is most likely closer to the truth than a random model, this will reduce the amount of time it takes to make a model for the new monkey! A concept like this could be searched with the term 'Transfer Learning'. So, in the end, if we just want a 'better than chance' accuracy, or we are asking more simplified questions (i.e. is there any activity in the left motor cortex? (yes or no)), then we could probably use the model from one monkey to generalize at an acceptable level to the other. However, if want a high level of accuracy, we would want to train a model for each individual monkey --- with the add-on point that we could use one monkey's model as a starting point to refine the other monkey's model! Hope that made sense and helped you with your understanding of the problem space. -Eric

[@Dmadiroe]

@@ericinthewild Thank you so much for the clear and detailed answer. I realy like the analogy with the house! :) It's a good thing that in most cases it doesn't have to be done from scratch!

[@lorenzobrunati8797]

I totally enjoyed the video! However, I have a question: how long does it take to make the monkey learn to play the game on his own? Is this an easy task?

[@ericinthewild]

Hey Lorenzo --- Glad you enjoyed the video! As to your question, though I do not work with monkeys myself, I can say that it can take many weeks (or months) of daily training to get a monkey to perform a task with high precision and repeat accuracy --- that said, it really depends on the demand and complexity of the game. Also, once they learn it, they are really really good, often even better than humans. Check these out: www.livescience.com/monkeys-outsmart-humans.html & kzbin.info/www/bejne/sKS7gWunmqt5bKM

[@ChechoColombia1]

Im a physic and mathematics student but i hope to end in neuroscience one day, thank you for the video.

[@ericinthewild]

Hi Sergio, physics and mathematics are very important for a lot of areas of neuroscience! I am sure they will be invaluable tools in your journey. Wish you the best!

[@LaMuflonaAssassina]

Great video :) was really clear and well explained! keep it going

[@ericinthewild]

Thanks! I am happy to hear this!

[@ConnoisseurOfExistence]

Nice&simple.

[@ericinthewild]

Thanks Nikolay!

[@lang1892]

MKe it longer sir

[@ericinthewild]

Haha, thanks Lang. For a little more insight, you can see my response to @webigfamily.online 's question!

[@billcichoke2534]

Some questions... Most animals do not process a 2D image as we do, moving or static. How is the monkey doing this? Also, any reflection of itself that it saw on the screen, would cause the monkey to attack the screen (they see their own reflection as another monkey). How was the monkey trained to see the screen as we do? Performing a set task is one thing; playing a video game or interacting with an input device to control another device is quite another. Where is the procedural describing how they did either thing? Or how they were able to get the monkey to perform a certain function on cue? Sorry, but it seems like a little flimflammery is going on with this demo...especially since no other experiments were documented showing the monkey's ability to do either one of these things WITHOUT any interface installed. The example used of the joystick controlling a robot arm was a THEORETICAL white paper, not a peer-reviewed assessment of an actual experiment.

[@ericinthewild]

Hi Bill - thanks for your thoughtful questions. While you're correct that most animals do not perceive images exactly how we do, primates have very similar visual systems, and therefore it could be thought that the monkey is roughly seeing an image that we would also be seeing. This is why they are the scientist's primary model in vision research. Primates playing games on screens, or controlling cursors are quite constant in research and can be found throughout the literature, this is no problem. Regarding the mirror image --- it is more than likely that the monkey is not seeing his reflection, as this would require a very particular lighting condition which the experimenters probably controlled for, i.e. you can also have a matte vs glossy screen ... but even if he was seeing his reflection, he would not necessary by default 'attack it'. While it's true that some animals attack mirror images when they feel it is an 'other' and that this 'other' poses a threat to them (like Robins would do) --- primates have been shown to be able to recognize the self (according to the mirror self-recognition test; and even perhaps Macaques fall into this group (see www.ncbi.nlm.nih.gov/pmc/articles/PMC2947497/ ) ) However, in the current setup, the monkey has been trained to sit in this chair, look this screen, and play the game --- all with positive reinforcement, so he would be likely relaxed and have no reason at all to feel threatened. As for whether the monkey can play a game or interact with a game --- there is no doubt in my mind that this is possible. There have been thousands of experiments where monkeys learn to use or play games, often on computer screens. In fact, sometimes they even far surpass what humans could achieve! Take a look at this memory game played by a chimpanzee (kzbin.info/www/bejne/sKS7gWunmqt5bKM ). How do they train the primate to perform? Again, almost always with 'rewards' in the form of food or drink. Your last comment about the referenced paper being only 'THEORETICAL' is simply incorrect, and I am not sure from where you are pulling this information. It is an actual experiment (see: journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.0000042 ), and I can even show you the actual demonstration itself in Nicolelis's (the senior author of the paper) TED talk from *2012*, where he says "...and that's exactly what we did 10 years ago" (www.ted.com/talks/miguel_nicolelis_a_monkey_that_controls_a_robot_with_its_thoughts_no_really ). You can start the video at 3:27 to get right into the good stuff. I hope that helped clarify some things for you!

[@billcichoke2534]

@@ericinthewild I'm afraid the links you provided for articles came up with 'file/page not found' results. As to the videos, it seems like the step of gradually training a primate in the recognition of symbology and any himan-directed task was either glossed over or not fully discussed. Obviously, this leaves the results in question, as real-time direction by an outside source would constitute bias and color the actual result.

[@ericinthewild]

@@billcichoke2534 You can try again now --- there was a problem with the ending parenthesis ")" being included in the URL... I'm not sure I understand what you mean by the second part of your comment. Essentially, the first time the monkey grabs the joystick and looks at the game, it has no idea what to do. However, it is reinforced slowly to learn to play by a series of training sessions: First steps would be to touch the joystick (monkey is rewarded), once they get that, then to move the joystick (rewarded), then move the joystick in the right direction (more reward), and eventually, to move the joystick toward a goal object (again rewarded).

[@billcichoke2534]

@@ericinthewild THAT is what I was talking about. The monkey learned a sequence of movements, not the abstract of a game, its rules or goals, or how to play it. The animal is following a set path instead of creating its own. This is experimental bias st its worst, where the subject is given one path to follow, thereby guaranteeing a specific result.

[@djfour]

What's your reaction to the video neuralink shared showing that the surgery robot punctures blood vessels routinely?

[@djfour]

also that the device disrupts the integrity of the skull permanently, leaving someone with a neuralink vulnerable if they were to have trauma to the area with the neuralink?

[@ericinthewild]

To which video are you referring to? In the demonstration of the robot, it is easy to see that it avoids veins ( kzbin.info/www/bejne/rqLKXpRmpNyiiKM ). Is there another video showing otherwise? Also --- I haven't seen any information about the skull integrity, could you link me to this study?