start	end	text
0	4700	Hello everybody, my name is Jack Twitty and today I'm going to be getting an introduction to the
4700	9940	world of neuroengineering, how it intersects with biohacking, and how you might one day
9940	16820	grow a brain in a jar yourself. First some real quick biotype information about myself. I'm
16820	21800	currently a doctoral student in the UNC NCSU Joint Department of Biomedical Engineering in
21800	26240	the Research Triangle part of North Carolina. You may or may not have heard of one or both
26240	29720	of those schools depending on whether or not you're a basketball fan, so I've included a
29720	35480	map just to be safe. So even though I've been working in STEM for a while now, I actually
35480	40360	started out as a philosophy and studio art major in undergrad at UNC. I always like to bring up
40360	44540	that disclaimer because I feel it's important to, you know, destroy any bit of credibility I might
44540	49640	have early on in the talk like this. After undergrad I took a hard right turn into engineering
49640	54540	and after teaching myself some electronics for fun and getting myself involved in 3D printing,
54540	58780	which is sort of the hacker gateway drug, I decided I liked that kind of thing and I wanted
58780	65000	to pursue it full-time. So I went on to get an engineering master's at NC State focusing on
65000	69880	microfabrication and then I moved into biomedical engineering specifically first at the City College
69880	75400	of New York and then back at a joint department run by both of my earlier schools. Another thing
75400	78840	I'm interested in is implantable hardware development and I'm actually in collaboration
78840	83720	with some of the other Biohacking Village speakers on a few projects in that area, so stay tuned next
83720	88580	year for that hopefully. And in the almost non-existent spare time I have outside of all
88580	94740	that, I also work as an EMT occasionally. So this talk is not going to be focused on my own research
94740	99320	but just to cover that briefly, my main interest is in the development of tools for neuroscience
99320	104180	and neurology research. I'm really interested in studies of consciousness in particular as sort of
104300	110000	a holdover from my past life as a philosophy major. Some specific areas I've worked in in the
110000	115200	include electronics and hardware for a variety of neurostimulation devices, circuits for energy
115200	120920	harvesting and self-powered systems, a variety of medical type sensing devices, primarily in the
120920	127280	form of wearables for health care. My current research focus involves using electrochemistry
127280	132980	to detect and monitor neurotransmitters both in vivo and in culture models. And for my doctoral
132980	137280	work I'm developing a cell culture system to try and explore some questions related to the
137280	144120	simultaneous release of two neurotransmitters in a specific region of the brain. So again, my goal
144120	148860	with this talk is not really to cover my own research in this area but rather to provide a
148860	153500	very, very high-level overview of neuroengineering and some interesting work that falls under that
153500	158540	umbrella. So neuroengineering is very loosely defined and even a very restrictive slice of
158540	163000	that kind of work would still be a tremendous amount of ground to cover. Since it's an impossible
163000	167080	task to try to do this field justice in such a short talk, I'm not really going to try to do that.
167280	170960	Instead I'll be providing a roadmap for the kind of work being done in neuroengineering
170960	174700	and I'll be highlighting some particular developments that I feel are important just to
174700	179680	give a sampling. This is intended merely as a starting point and in that spirit I'll conclude
179680	183700	by discussing what might be involved in getting started in neuroengineering in a garage,
183700	187560	basement workshop, or other improvised mad science lab if you're choosing.
189260	194220	So to start out, because neuroengineering is such an expansive and diverse field, I think it's useful
194220	198200	to start with some basic delineations of the kind of work being done in that space.
198600	203340	Many different groups define neuroengineering in different ways, sometimes vastly so,
203340	207360	and that can make categorizing this type of work fairly difficult. And since all these different
207360	211820	groups are putting forth their own definitions of the term, I figure I might as well join in.
211820	217380	So I'm going to share my own totally arbitrary ad hoc taxonomy of neuroengineering as it makes
217380	222700	sense to me. Your mileage may vary. I like to divide up neuroengineering projects based on
222700	228300	where the system is being used and specifically what it's being used for. The first main branch
228300	232320	under this scheme is in vivo neuroengineering and that's probably the part of neuroengineering
232320	236520	people in general are most familiar with and what you might think of when I say the term
236520	242020	neuroengineering. I define this branch to include systems that involve creating an interface with
242020	247080	the nervous system of a living human or other animal if applicable. This includes a huge range
247080	251100	of equipment encompassing both invasive and non-invasive devices which can be used for
251100	256460	sensing stimulation or in some cases both. This family includes things like prosthetic devices
256460	260960	that interact with peripheral nerves, stimulators found anywhere in the body like pacemakers,
260960	266620	deep brain stimulation leads, EEG arrays, electrocorticogram probes, and many other
266620	270720	devices depending on where exactly you want to draw the line between nervous system and not
270720	276520	nervous system. As far as recent development goes, this is the category into which virtually all brain
276520	281720	interfaces would fall which have been a huge area of active research for the past few decades.
283000	288260	The second main branch of neuroengineering is the sort of the odd one out and I think a lot
288260	292840	of people might not actually categorize it as neuroengineering proper. But that being said, I
292840	297760	think its potential overlap with the other two branches justifies its inclusion. If we think of
297760	303460	in vivo neuroengineering as using human-designed electronic systems to work in a biological context,
303460	307620	neuromorphic engineering could be posed as the exact opposite of this process.
307720	311980	So in neuromorphic engineering, researchers take inspiration from the behavior of the nervous system
311980	316780	and apply those behaviors and mechanisms to problems of computation, often in a context that
316780	321600	doesn't directly involve biology at all. This works because the nervous system of humans and other
321600	326980	animals exists primarily as a powerful processing and control system, one which has evolved over
326980	331840	millions of years to be highly robust in the face of interference and edge cases. These are things
331840	337000	that computing systems, we typically envision them, are notoriously bad at handling. So while
337280	341520	a nervous system can't compete with digital circuits in terms of processing volume and speed,
341520	346360	all its being equal, many real-world environments would benefit tremendously from the kind of
346360	352020	flexibility offered by neuro-inspired processing. In particular, this combined analog and digital
352020	356220	nature of the sum and fire model of neural processing, where a bunch of digital inputs
356220	361340	are weighted in an analog fashion and determine a single digital output. This allows for advanced
361340	366800	categorization and even quasi-learning behaviors. As testament to the applicability of this idea,
366800	370920	just look at the explosion of the use of neural networks as part of the machine learning tool set
370920	375180	in recent years. While these are implemented almost exclusively in software, the same
375180	380580	neuro-inspired principles apply in a loose way to vastly improve performance in specific use cases.
380800	385660	And under this totally arbitrary taxonomy, neural networks would also fall under the loose umbrella
385660	390160	of neuroengineering. However, more interesting, in my opinion, are recent attempts to apply the
390160	395860	same strategy to hardware directly. So in this approach, electronic stand-ins for neurons, which
395860	400320	vary in terms of how closely they mimic actual neural behavior, are assembled together to create
400480	404800	a processing unit as needed. So this often involves either the use of microelectronic
404800	408680	circuits consisting of multiple conventional components, which together generate neuron-like
408680	413720	behavior, or in more recent examples, the use of newer exotic components to directly create this
413720	418500	desired behavior. So by designing integrated circuits to intrinsically operate in a neuro-inspired
418500	423080	fashion rather than implementing this behavior in software using a conventional digital processor,
423080	427120	this process can be massively paralyzed and sped up, greatly increasing the benefits of this
427120	432400	approach. The final branch in this taxonomy is one that's probably less well-known than the other two.
432400	436420	In in vitro neuroengineering, neural cells themselves are used in the construction of other
436420	442040	systems engineered towards some purpose other than their intrinsic biological function. So much like
442040	446920	the previous branch, this often involves leveraging the inherent processing capabilities of neuron-like
446920	451640	units to perform some computational task which may itself be totally unrelated to biology.
451820	455580	However, unlike neuromorphic computing, where we're using an engineered electronic device as our
455580	459880	building block, under the in vitro approach we're utilizing actual neurons and interfacing these to
459880	464560	external inputs and outputs. So rather than trying to recreate the behavior of a natural system, we're
464560	468820	cutting out the middleman and harnessing the real thing. While plenty of research has been done on
468820	473220	neuronal behavior in vitro over the past century, using neurons ex vivo as part of a larger
473220	478140	non-biological system is a much newer idea. One of the challenges inherent in this approach is that
478140	482600	unlike most circuits, neurons are living biological entities with constant metabolic requirements and
482700	487020	a propensity to change in real time, sometimes in undesired ways. Short of freezing in sub-zero
487020	490160	temperatures, there's no way to turn off a cell culture the way you can turn off an electronic
490160	495060	device. That said, the biological nature of these systems can also be beneficial. These same inherent
495060	499580	behaviors can offer the ability to self-organize, change in response to inputs, and theoretically
499580	504320	heal to an extent impossible in electronic circuits. Finally, and of particular interest to
504320	509300	the biohacker-turned-neuroengineer, cells can, with care, be grown and manipulated in a variety of
509300	513700	settings, including those available to the enterprising biohacker. Certainly less stringent
513700	517220	environments than the type of clean room required for microelectronic fabrication.
518320	521640	So hopefully at this point, my method of delineating these different branches of research
521640	525880	makes more sense. Between any two branches, there are a number of parallels that can be drawn, with
525880	530200	variation only in either the end goal of the system or the exact strategy and methods used to
530200	534940	accomplish that goal. It's also worth noting that, as defined, projects that fall into one of these
534940	539400	branches have tended to remain within only one branch of neuroengineering. However, I would argue
539400	544200	that as these technologies mature in these use cases to which they're applied grow in number,
544200	548060	we'll see more efforts to join these approaches and leverage the unique strengths of each to allow
548060	552280	us to do things we can only dream of at this point. This kind of synergistic approach is something
552280	557000	that I'm personally interested in exploring in my own future research. So speaking of doing things,
557000	560860	what are some actual examples of each of these types of neuroengineering? And as a reminder,
560860	565280	this is intended to be a highlight reel and not a comprehensive overview. My goal is just to point
565280	569840	out some specific efforts I think are of particular importance or interest, and I encourage you to use
569840	572760	them as a starting point if you're interested in learning more about this kind of thing.
574000	578140	So here's an example of a state-of-the-art neural interface array, which is the culmination of
578140	582560	several years of development work by multiple collaborating labs. The Shepard and Viventi
582560	586420	groups are both big players in the world of electrode arrays, and John Rogers is practically
586420	590740	the founder of the field of soft electronics. I really encourage you to look in the work done by
590740	594980	these labs if that kind of thing interests you. This device is an example of an electrode array
594980	599240	that's similar in operating principle to an EEG array, but with some key differences that dramatically
599240	603800	improve sensitivity and resolution. So this type of device is called an electrocorticogram, or
603800	607540	ECoG array, and it's designed to be placed directly on the surface of the brain.
607700	612060	Since neural activity recording resolution depends on both surface coverage and distance
612060	616700	from the neurons being monitored, an array like this offers tremendous advantages. So first,
616700	621220	the huge number of recording sites allow for high inherent spatial precision. Second, the thinness
621220	624920	and flexibility of this array allows it to be molded to the curvature of the brain, improving
624920	629740	coverage and minimizing separation distance. This is particularly important in the case of primates,
629740	633480	and especially humans, where there's a large degree of surface contouring. So this design
633480	637440	allows the array to be placed around the gyri and into the cell side, which would be impossible
637440	641600	with a thicker array. Further, this array incorporates active electronics into the
641600	646120	electrodes themselves, allowing for signals to be amplified near their origin and allowing multiple
646120	650060	electrodes to be multiplexed between a smaller number of conductors running between the skull
650060	654480	and the outside instrumentation. Without this type of signal management, high density arrays quickly
654480	659520	become impossible to manage practically. And because it's sort of difficult to appreciate
659520	664360	that scale, here's that array in higher resolution. So this is 1,008 electrodes in
664360	670600	less than one square centimeter, which is pretty darn impressive. As mentioned previously,
670600	674200	neuromorphic engineering applies biologically inspired processing to other computational
674200	678300	problems, which might not have anything to do with biology at all. However, in this particular case,
678300	682520	artificial neurons were used to model the activity of simulated portions of the brain.
682800	687460	So due to the immense scale of connectivity in the brain, which is made up of about 86 billion
687460	691820	neurons and vastly more synaptic connections, and the difficulty of accessing the living brain
691820	696300	using instrumentation, simulation efforts like the Human Brain Project have played a vital role in
696300	700980	the study of cognitive processing. That said, even these massive joint efforts are limited due to the
700980	705820	sheer computational resources required. To help ameliorate this problem, researchers at the Human
705820	710240	Brain Project leveraged the fact that artificial neurons fabricated from metallic and semiconducting
710240	715080	materials have a huge intrinsic speed advantage over biological neurons due to differences in
715080	720260	conduction mechanisms. So as a result, by simulating real neurons using engineered ones in silicon,
720760	725000	limited neuronal processes were able to be simulated at a much faster rate than what's
725000	730460	observed in real physiology. The brain scale system shown here utilizes a large number of
730460	735080	artificial neurons fabricated on silicon wafers using conventional integrated circuit manufacturing
735080	740640	techniques. These are created in bulk across the wafer surface, and unlike most integrated circuits,
740640	745360	these are actually directly accessed from the intact wafer, so there's no dicing.
746080	750640	So here's a closer look at the individual neuron circuits. At the bottom, that's a layout of two
750640	756380	of the neural equivalent circuits that are fabricated in mass on that chip. They're
756380	762700	arranged into one of 384 clusters of 512 neurons each onto a single wafer, and the assembly into
762700	767540	which that wafer is loaded is shown on the right with supporting and interface electronics.
767540	776020	So 20 wafer modules consisting of 384 clusters of 114,688 programmable synapses and 512 neurons
776020	782820	each give a total of 880 million synapses and 3.9 million neurons for the entire system,
782820	788700	which is further capable of running 100,000 times faster than the same set of neurons would
788700	793780	in vivo. So these advantages and the ability to continuously monitor each simulated neuron
793780	799100	in real time have enabled considerable advances in the field of computational neuroscience.
799800	804680	So as I mentioned earlier, artificial hardware or software-based neurons aren't the only tools
804680	808820	that neuroengineers use to solve processing tasks. So here's an example of real neurons
808820	813820	being used to do the same thing. In 2005, researchers at the University of Florida were
813820	819360	able to extract about 25,000 neurons from rat cortical tissue, seed them onto a 60-electrode
819360	824400	microarray treated to function as a culture disk, and then observe the spontaneous formation
824400	828420	of interneuronal connections and eventually synchronized firing activity around 10 days
828420	834360	after initial seeding. Two electrodes were used for both stimulation and recording, and these
834360	838400	outputs were fed to a flight simulator program where neuronal activity of these electrodes was
838400	843580	used to establish pitch and yaw control. So by feeding back a stimulatory error signal based on
843580	848240	telemetry output by the flight simulator, firing activity and neurons near the electrode were
848240	853620	modified. And over the course of repeating the simulation multiple times, neuronal activity was
853620	858060	found to change in self-weight in order to produce stable flight in response to changes in telemetry
858060	864380	data over the course of about 15 minutes. So this feedback adjust and test mechanism is by itself
864380	868880	very similar to techniques used to train software neural networks, which isn't surprising given the
868880	873660	fact that these were in turn inspired by learning mechanisms seen in the nervous system. So it's
873660	879400	come full circle here. So now that we've seen some examples of different types of neuroengineering
879400	884020	research, I want to briefly discuss how biohackers might go about getting involved in this type of
884020	890120	work on their own. So growing cells in general is not a difficult process. To quote Jeff Goldblum,
890120	895240	life finds a way. The real difficulty in this process arises from the same principle, life
895240	900340	finds a way, but it's not always the way that you want it to. So in order to successfully culture
900480	905080	a specific cell type, you have to take careful consideration of the cell source, the conditions
905080	910400	in which the cells are kept, and the intended final use of those cells. So regardless of type,
910400	916020	though, a key principle of cell culture is aseptic technique. So care must be taken at all times in
916020	919920	the process to avoid introducing other microorganisms into the culture environment,
919920	925920	as these might outcompete or otherwise harm the cells of interest. So prior to culture,
925920	930420	the cells must first be acquired from some suitable source. The simplest mechanism for
930420	935040	obtaining neural cells is the harvest of primary cells from tissue samples, either acquired directly
935040	940320	from the donor animal or ordered through a supplier such as BrainBits or some other vendor.
940540	944640	The most common sources for these cells are rats, mice, and embryonic chicks.
945280	949500	After harvest, the cells of interest have to be purified from this sample, which will also likely
949500	954580	contain other cell types as well as other biomolecules and extracellular matrix proteins
954580	959020	and connected tissue. In order to enrich the primary culture, these extra components have to
959020	963540	be removed, typically with enzymes known as proteases, which break down matrix proteins,
963540	968560	allowing the cells to be dissociated. Additional sorting based on cell type may also be needed to
968560	973780	remove glial cells and other neuron phenotypes if necessary. So since mature neurons don't normally
973780	978140	undergo division, primary cultures need to be reestablished periodically as more cells are
978140	983340	required. To avoid this problem, another option is using immortalized secondary cultures,
983340	989140	and these are derivatives of primary cells which have been modified to exhibit tumor-like behavior,
989140	993820	including indefinite division. So these cells tend to take on very homogeneous characteristics,
993820	998200	which can be beneficial. However, this modification is the possible downside of modifying
998200	1002340	the cellular characteristics in a way that adversely impacts the intended application,
1002340	1006620	and it definitely can differ from the normal phenotype of the cell. So whether or not that's
1006740	1011060	a deal breaker depends on the application. So a more complex third option is the use of
1011060	1015360	progenitor cells, which are cells that are stem-like in that they're capable of undergoing
1015360	1020220	both division and differentiation, which allows for a baseline stock of cells to be maintained
1020220	1024260	while simultaneously producing an essentially indefinite supply of the final differentiated
1024260	1029240	cells needed. So this route's tricky because the cells rely on specific and sometimes subtle
1029240	1033300	external cues to guide their ultimate differentiation, which must be carefully
1033300	1039920	monitored during culturing because it can change pretty rapidly. So one major advantage of this
1039920	1045620	technique is that this represents one of the only ways of acquiring human neural cells. So mature
1045620	1050460	cells can undergo induced dedifferentiation, which reverts them back into stem-like cells
1051120	1055800	when appropriate genetic modifications are made. This procedure is moderately difficult,
1055800	1058960	but it's fairly well established at this point and remains one of the only ways that you can
1058960	1065020	get human neural cells as a culture, which is pretty cool. So in order to grow, cells have
1065020	1069360	to be stormed in a warm, humid environment, and the supply of gas must be suitable for the type
1069360	1074080	of cell being cultured. It's important for the environment to remain aseptic again, as it's
1074080	1079220	essential for preventing other opportunistic organisms such as yeasts, bacteria, and viruses
1079220	1084260	from colonizing the culture. Most abiding cell types will grow to the point where their density
1084260	1088840	in the culture begins to negatively impact cell survival. When the cells reach a certain coverage
1088840	1093920	of the culture chamber surface, termed confluence, some of the cells must be removed to prevent
1093920	1099120	adverse effects from damaging the remaining cells. So this removal of cells can also be
1099120	1104400	used to expand the culture back into additional flasks to build up stock if desired. This is a
1104400	1109100	technique called passaging. And in addition to the different temperatures and atmospheric
1109500	1113660	requirements, cells tend to differ even more with respect to their metabolic requirements.
1113660	1118660	So fresh media has to be changed regularly and supplied appropriately, and it has to contain
1118660	1123280	the required nutrients specific to that cell type. Further, neural cells are particularly
1123280	1127640	sensitive to the choice of substrate or the material on which the cells directly rest.
1127820	1131340	So frequently the culture chamber itself will be coated with an intermediate substance,
1131340	1135840	often a polymer or substances normally found in the extracellular matrix like collagen.
1136080	1140220	Different substrates can actually induce or prohibit different cell differentiation routes
1140220	1144840	in some cases, or it may even cause de-differentiation. It entirely depends on the cell
1144840	1150660	specific type. So certain substrates are able to quickly polymerize into 3D structures suitable
1150660	1155760	for seeding cells into to form volumetric cultures. So this is sometimes used with
1155760	1160760	neural progenitor cells to form 3D spatially differentiated structures called organoids.
1162120	1166020	And then finally, you need to be able to interact with your cultured cells in a way that's
1166020	1170320	appropriate for your specific application. So typically this is going to involve electrical
1170320	1175260	or chemical detection or stimulation, usually referred to as electrophysiology and neurochemistry
1175260	1180280	respectively. In some applications, both methodologies might be used at the same time.
1180280	1184760	So groups of cells can be stimulated and local potentials can be measured using electrode arrays,
1184760	1189820	which is those discussed earlier. If a neuron specific electrical activity needs to be measured,
1189820	1194080	patch clamps and other similar devices might be used. In this technique, you take a glass
1194080	1198720	capillary that's formed into a very small point that's filled with electrolyte. You then aspirate
1198720	1202560	that against the membrane of a target cell. And what that does is it creates an electrical
1202560	1207040	connection with the interior of the cell, allowing for the cell membrane potential to be measured or
1207040	1212400	even driven with respect to the extracellular space. So neurochemical measurements,
1212400	1216580	which are a little trickier, can be made using electrochemical techniques where specific
1216580	1220740	analytes are oxidized or reduced against the surface of an electrode in response to a varying
1220740	1225800	electrode potential. This works with only a subset of analytes, however, but there are additional
1225800	1230880	tricks that can be used to increase selectivity or expand applicability to other molecules through
1230880	1236260	the use of enzymes or other functional coatings. And then when the cells need to be arranged in
1236260	1242240	specific ways, microfluidic devices can be created using soft lithography and seeded with cells,
1242240	1247340	sort of like this one here. It's also entirely possible to create larger millifluidic devices
1247340	1252440	that can be fabricated using fairly basic tools. I've seen 3D printed ones, for example.
1252840	1257720	Generally, a positive pattern is made and an elastomer like PDMS or polydimethylsiloxane
1257720	1262220	is poured onto the mold and then removed. And then that resulting cast is then bonded to a
1262220	1267760	microscope slide using ozone to create a really tight seal. And then this can have fluid basically
1268200	1274320	flowed through it with a watertight joint between the cast and the glass. And you can put cells in
1274320	1278040	there and basically they're allowed to grow freely and then interact based on the geometry that's
1278040	1284940	specified by the design. So to finish up, I have some useful links here. One of the benefits of
1284940	1288900	doing this virtually is you all get access to the stuff after the fact. So if you're interested in
1288900	1293920	suppliers of neural culture reagents and whatnot, BrainBits is a great resource for that.
1294040	1299020	A bunch of the other vendors like Thermo Fisher have reagent and media guides and also tend to
1299020	1304040	publish protocols for de-differentiation and differentiation of neural cells. So if you want
1304040	1307760	to try getting started on any of this stuff at home, that would be a resource I would recommend.
1307960	1311620	If you're interested in the Human Brain Project, which has really taken a lot of these
1313120	1317020	methodologies and some of the systems that have been developed and applied them to problems of
1317020	1321880	neuroscience, as I mentioned earlier, it sort of extends beyond neuroengineering, but I encourage
1321880	1326060	you to check out the Human Brain Project in that case, and specifically their neuromorphic
1326060	1330920	computing research platform, which I discussed earlier, of which the BrainScale system is just
1330920	1336740	one facet. There's a bunch of other systems like that and other examples like IBM is coming out
1336740	1342660	with some models of that on dedicated ICs currently. So there's a huge variety of different
1342660	1347000	projects in all of these areas of neuroengineering. Like I said, this is just a very brief
1347620	1352700	skim of some of the projects that I think are interesting and representative of each area.
1353200	1355840	And so I definitely encourage you to check it out more if you're interested.
1356380	1359620	With that, here's some other references as well.
1359900	1364020	And thank you very much for the talk. I'd be happy to take any questions during the Q&A.