Eliasmith, C. (2015). on the Eve of Artificial Minds. in T

On the Eve of Artificial Minds Chris Eliasmith

I review recent technological, empirical, and theoretical developments related to Author building sophisticated cognitive machines. I suggest that rapid growth in robotics, brain-like computing, new theories of large-scale functional modeling, and finan- Chris Eliasmith cial resources directed at this goal means that there will soon be a significant in- celiasmith@ uwaterloo.ca crease in the abilities of artificial minds. I propose a specific timeline for this development over the next fifty years and argue for its plausibility. I highlight some University of Waterloo barriers to the development of this kind of technology, and discuss the ethical and Waterloo, ON, Canada philosophical consequences of such a development. I conclude that researchers in this field, governments, and corporations must take care to be aware of, and will- Commentator ing to discuss, both the costs and benefits of pursuing the construction of artificial minds. Daniela Hill daniela.hill@ gmx.net Keywords Johannes Gutenberg-Universität Artificial cognition | Artificial intelligence | Brain modelling | Machine learning | Mainz, Germany Neuromorphic computing | Robotics | Singularity Editors

Thomas Metzinger metzinger @uni-mainz.de Johannes Gutenberg-Universität Mainz, Germany

Prediction is difficult, especially about the future Jennifer M. Windt – Danish Proverb [email protected] Monash University Melbourne, Australia

1 Introduction

The prediction game is a dangerous one, but ing this kind of prognostication are justified be- that, of course, is what makes it fun. The pit- cause of the enormous potential impact of a falls are many: some technologies change expo- new kind of technology that lies just around the nentially but some don’t; completely new inven- corner. It is a technology we have been dream- tions, or fundamental limits, might appear at ing about—and dreading—for hundreds of any time; and it can be difficult to say some- years. I believe we are on the eve of artificial thing informative without simply stating the minds. obvious. In short, it’s easy to be wrong if you’re In 1958 Herbert Simon & Allen Newell specific. (Although, it is easy to be right if claimed that “there are now in the world ma- you’re Nostradamus.) Regardless, the purpose chines that think” and predicted that it would of this essay is to play this game. As a con- take ten years for a computer to become world sequence, I won’t be pursuing technical discus- chess champion and write beautiful music sion on the finer points of what a mind is, or (1958, p. 8). Becoming world chess champion how to build one, but rather attempting to took longer, and we still don’t have a digital paint an abstract portrait of the state of re- Debussy. More importantly, even when a com- search in fields related to machine intelligence puter became world chess champion it was not broadly construed. I think the risks of undertak- generally seen as the success that Simon and

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 1 | 17 www.open-mind.net

Newell had expected. This is because the way in style computation—has been rapidly scaling up, which Deep Blue beat Gary Kasparov did not with several current projects expected to hit strike many as advancing our understanding of millions (Choudhary et al. 2012) and billions cognition. Instead, it showed that brute force (Khan et al. 2008) of neurons running in real computation, and a lot of careful tweaking by time within the next three to four years. These expert chess players, could surpass human per- hardware advances are critical for performing formance in a specific, highly circumscribed en- efficient computation capable of realizing brain- vironment. like functions embedded in and controlling Excitement about AI grew again in the physical, robotic bodies. 1980s, but was followed by funding cuts and Finally, unprecedented financial resources general skepticism in the “AI winter” of the have been allocated by both public and private 1990s (Newquist 1994). Maybe we are just stuck groups focusing on basic science and industrial in a thirty-year cycle of excitement followed by applications. For instance, in February 2013 the disappointment, and I am simply expressing the European Union announced one billion euros in beginning of the next temporary uptick. How- funding for the Human Brain Project, which fo- ever, I don’t think this is the case. Instead, I cuses on developing a large scale brain model as believe that there are qualitative changes in well as neuromorphic and robotic platforms. A methods, computational platforms, and finan- month later, the Obama BRAIN initiative was cial resources that place us in a historically announced in the United States. This initiative unique position to develop artificial minds. I devotes the same level of funding to experi- will discuss each of these in more detail in sub- mental, technological, and theoretical advances sequent sections, but here is a brief overview. in neuroscience. More recently, there has been a Statistical and brain-like modeling meth- huge amount of private investment: ods are far more mature than they have ever Google purchased eight robotics and AI been before. Systems with millions (Garis et al. companies between Dec 2013 and Jan 2014, in- 2010; Eliasmith et al. 2012) and even tens of cluding industry leader Boston Dynamics Stunt millions (Fox 2009) of simulated neurons are (2014). suddenly becoming common, and the scale of Qualcomm has introduced the Zeroth pro- models is increasing at a rapid rate. In addition, cessor, which is modeled after how a human the challenges of controlling a sophisticated, brain works (Kumar 2013). They demonstrated nonlinear body are being met by recent ad- an Field-Programmable Gate Array (FPGA) vances in robotics (Cheah et al. 2006; Schaal et mock-up of the chip performing a reinforcement al. 2007). These kinds of methodological ad- learning task on a robot. vances represent a significant shift away from Amazon has recently expressed a desire to classical approaches to AI (which were largely provide the Amazon Prime Air service, which responsible for the previously unfullfilled prom- will use robotic quadcopters to deliver goods ises of AI) to more neurally inspired, and brain- within thirty minutes of their having been like ones. I believe this change in focus will al- ordered (Amazon 2013). low us to succeed where we haven’t before. In IBM has launched a product based on Wat- short, the conceptual tools and technical meth- son, which famously beat the best human Jeop- ods being developed for studying what I call ardy players (http://ibm.com/innovation/us/wats “biological cognition” (Eliasmith 2013), will on/). The product will provide confidence based make a fundamental difference to our likelihood responses to natural language queries. It has been of success. opened up to allow developers to use it in a wide Second, there have been closely allied and variety of applications. They are also developing a important advances in the kinds of computa- neuromorphic platform (Esser et al. 2013). tional platforms that can be exploited to run In addition, there are a growing number of these models. So-called “neuromorphic” com- startups that work on brain-inspired computing puting—hardware platforms that perform brain- including Numenta, the Brain Corporation, Vi-

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 2 | 17 www.open-mind.net carious, DeepMind (recently purchased by velopments in neuromorphic hardware that I Google for $400 million) and Applied Brain Re- discuss below are inspired by some basic fea- search, among many others. In short, I believe tures of neural computation. For instance, all of there are more dollars being directed at the the neuromorphic approaches use spiking neural problem than ever before. networks SNNs to encode and process informa- It is primarily these three forces that I be- tion. In addition, there is unanimous agreement lieve will allow us to build convincing examples that biological computation is in orders of mag- of artificial minds in the next fifty years. And, I nitude more power efficient than digital compu- believe we can do this without necessarily defin- tation (Hasler & Marr 2013). Consequently, a ing what it is that makes a “mind”—even an ar- central motivation behind exploring these hard- tificial one. As with many subtle concepts— ware technologies is that they might allow for such as “game,” to use Wittgenstein’s example, sophisticated information processing using small or “pornography,” to use Supreme Court Justice amounts of power. This is critical for applica- Potter Stewart’s example—I suspect we will tions that require the processing to be near the avoid definitions and rely instead on our soph- data, such as in robotics and remote sensing. In isticated, but poorly understood, methods of what follows I begin by providing a sample of classifying the world around us. In the case of several major projects in neuromorphic comput- “minds,” these methods will be partly behavi- ing that span the space of current work in the oural, partly theoretical, and partly based on area. I then briefly discuss the current state of judgments of similarity to the familiar. In any high-performance computing and robotics, to case, I do not propose to provide a definition identify the roles of the most relevant technolo- here, but rather to point to reasons why the ar- gies for developing artificial minds. tifacts we continue to build will become more To complement its cognitively focused and more like the natural minds around us. In Watson project, IBM has been developing a doing so, I survey recent technological, theoret- neuromorphic architecture, a digital model of ical, and empirical developments that are im- individual neurons, and a method for program- portant for supporting our progress on this ming this architecture (Esser et al. 2013). The front. I then suggest a timeline over which I ex- architecture itself is called TrueNorth. They ar- pect these developments to take place. Finally, I gue that the “low-precision, synthetic, simultan- conclude with what I expect to be the major eous, pattern-based metaphor of TrueNorth is a philosophical and societal impacts on our being fitting complement to the high-precision, ana- able to build artificial minds. As a reminder, I lytical, sequential, logic-based metaphor of am adopting a somewhat high-level perspective today’s of von Neumann computers” (Esser et on the behavioural sciences and related techno- al. 2013, p. 1). TrueNorth has neurons organ- logies in order to make clear where my broad ized into 256 neuron blocks, in which each (and likely wrong) predictions are coming from. neuron can receive input from 256 axons. To as- In addition, if I’m not entirely wrong, I suspect sist with programming this hardware, IBM has that the practical implications of such develop- introduced the notion of a “corelet,” which is an ments will prove salient to a broad audience, abstraction that encapsulates local connectivity and so, as researchers in the area, we should in small networks. These act like small pro- consider the consequences of our research. grams that can be composed in order to build up more complex functions. To date the demon- 2 Technological developments strations of the approach have focused on simple, largely feed-forward, standard applica- Because I take it that brain-based approaches tions, though across a wide range of methods, provide the “difference that makes a difference” including Restricted Boltzmann Machines between current approaches and traditional AI, (RBMs), liquid state machines, Hidden Markov here I focus on developments in neuromorphic Model (HMMs), and so on. It should be noted and robotic technology. Notably, all of the de- that the proposed chip does not yet exist, and

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 3 | 17 www.open-mind.net current demonstrations are on detailed simula- There are also a number of neuromorphic tions of the architecture. However, because it is projects that use analog instead of digital im- a digital chip the simulations are highly accur- plementations of neurons. Analog approaches ate. tend to be several orders of magnitude more A direct competitor to IBM’s approach is power efficient (Hasler & Marr 2013), though the Zeroth neuromorphic chip from Qualcomm. also more noisy, unreliable, and subject to pro- Like IBM, Qualcomm believes that constructing cess variation (i.e., variations in the hardware brain-inspired hardware will provide a new due to variability in the size of components on paradigm for exploiting the efficiencies of neural the manufactured chip). These projects include computation, targeted at the kind of informa- work on the Neurogrid chip at Stanford Univer- tion processing at which brains excel, but which sity (Choudhary et al. 2012), and on a chip at is extremely challenging for von Neumann ap- ETH Zürich (Corradi, Eliasmith & Indiveri proaches. The main difference between these 2014). The Neurogrid chip has demonstrated two approaches is that Qualcomm has commit- larger numbers of simulated neurons—up to a ted to allowing online learning to take place on million—while the ETH Zürich chip allows for the hardware. Consequently, they announced online learning. More recently, the Neurogrid their processor by demonstrating its application chip has been used to control a nonlinear, six in a reinforcement learning paradigm on a real- degree of freedom robotic arm, exhibiting per- world robot. They have released videos of the haps the most sophisticated information pro- robot maneuvering in an environment and cessing from an analog chip to date. learning to only visit one kind of stimulus In addition to the above neuromorphic (white boxes: http://www.youtube.com/watch? projects, which are focused on cortical simula- v=8c1Noq2K96c). It should again be noted that tion, there have been several specialized neur- this is an FPGA simulation of a digital chip omorphic chips that mimic the information pro- that does not yet exist. However, the simula- cessing of different perceptual systems. For ex- tion, like IBM’s, is highly accurate. ample, the dynamic vision sensor (DVS) artifi- In the academic sphere, the Spinnaker cial retina developed at ETH Zürich performs project at Manchester University has not fo- real-time vision processing that results in a cused on designing new kinds of chips, but has stream of neuron-like spikes (Lichtsteiner et al. instead focused on using low-power ARM pro- 2008). Similarly, an artificial cochlea called cessors on a massive scale to allow large-scale AEREAR2 has been developed that generates brain simulations (Khan et al. 2008). As a res- spikes in response to auditory signals (Li et al. ult, the focus has been on designing ap- 2012). The development of these and other proaches to routing that allow for the high neuronal sensors makes it possible to build fully bandwidth communication, which underwrites embodied spiking neuromorphic systems (Gal- much of the brain’s information processing. luppi et al. 2014). Simulations on the Spinnaker hardware typic- There have also been developments in tra- ally employ spiking neurons, like IBM and ditional computing platforms that are import- Qualcomm, and occasionally allow for learning ant for supporting the construction of models (Davies et al. 2013), as with Qualcomm’s ap- that run on neuromorphic hardware. Testing proach. However, even with low power conven- and debugging large-scale neural models is often tional chips, the energy usage is projected to much easier with traditional computational be higher on the Spinnaker platform per platforms such as Graphics Processing Unit neuron. Nevertheless, Spinnaker boards have (GPUs) and supercomputers. In addition, the been used in a wider variety of larger-scale development of neuromorphic hardware often embodied and non-embodied applications. relies on simulation of the designs before manu- These include simulating place cells, path in- facture. For example, IBM has been testing tegration, simple sensory-guided movements, their TrueNorth architecture with very large- and item classification. scale simulations that have run up to 500 billion

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 4 | 17 www.open-mind.net neurons. These kinds of simulations allow for ping-pong champion Timo Boll (http://www.y- designs to be stress-tested and fine-tuned before outube.com/watch?v=_mbdtupCbc4). Despite costly production is undertaken. In short, the the somewhat disappointing outcome, this kind development of traditional hardware is also an of competition would not have been thought important technological advance that supports possible a mere five years ago (Ackerman 2014). the rapid development of more biologically- These three areas of technological develop- based approaches to constructing artificial cog- ment—neuromorphics, high-performance con- nitive systems. ventional computing, and robotics—are pro- A third area of rapid technological devel- gressing at an incredibly rapid pace. And, more opment that is critical for successfully realizing importantly, their convergence will allow a new artificial minds is the field of robotics. The suc- class of artificial agents to be built. That is, cess of recent methods in robotics have entered agents that can begin processing information at public awareness with the creation of the very similar speeds and support very similar Google car. This self-driving vehicle has success- skills to those we observe in the animal king- fully navigated hundreds of thousands of miles dom. It is perhaps important to emphasize that of urban and rural roadways. Many of the tech- my purpose here is predictive. I am not claim- nologies in the car were developed out of ing that current technologies are sufficient for DARPA’s Grand Challenge to build an building a new kind of artificial mind, but autonomous vehicle that would be tested in rather that they lay the foundations, and are both urban and rural settings. Due to the suc- progressing at a sufficient rate to make it reas- cess of the first three iterations of the Grand onable to expect that the sophistication, adapt- Challenge, DARPA is now funding a challenge ability, flexibility, and robustness of artificial to build robots that can be deployed in emer- minds will rapidly approach those of the human gency situations, such as a nuclear meltdown or mind. We might again worry that it will be dif- other disaster. ficult to measure such progress, but I would One of the most impressive humanoid ro- suggest that progress will be made along many bots to be built for this challenge is the Atlas, dimensions simultaneously, so picking nearly constructed by Boston Dynamics. It has twenty- any of dimensions will result in some measur- eight degrees of freedom, covering two arms, two able improvement. In general, multi-dimensional legs, a torso, and a head. The robot has been similarity judgements are likely to result in “I’ll demonstrated walking bipedally, even in ex- know it when I see it” kinds of reactions to clas- tremely challenging environments in which it sifying complicated examples. This may be de- must use its hands to help navigate and steady it- rided by some as “hand-wavy”, but it might self (http://www.youtube.com/watch?v=zkBn- also be a simple acknowledgement that FPBV3f0). Several teams in this most recent “mindedness” is complex. I would like to be Grand Challenge have been awarded a copy of clear that my claims about approaching human Atlas, and have been proceeding to competitively mindful behaviour are to be taken as applying design algorithms to improve its performance. to the vast majority of the many measures we In fact, there have been a wide variety of use for identifying minds. significant advances in robotic control algorithms, enabling robots—including quad- 3 Theoretical developments copters, wheeled platforms, and humanoid robots—to perform tasks more accurately and Along with these technological developments more quickly than had previously been possible. there have been a series of theoretical develop- This has resulted in one of the first human ments that are critical for building large-scale versus robot dexterity competitions being re- artificial agents. Some have argued that theoret- cently announced. Just as IBM pitted Watson ical developments are not that important: sug- against human Jeopardy champions, Kuka has gesting that standard back propagation at a pitted its high-speed arm against the human sufficiently large scale is enough to capture

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 5 | 17 www.open-mind.net complex perceptual processing (Krizhevsky et high dimensional control spaces (Cheah et al. al. 2012). That is, building brain-like models is 2006). more a matter of getting a sufficiently large Concurrently with these more abstract computer with enough parameters and neurons characterizations of brain function there have than it is of discovering some new principles been theoretical developments in neuroscience about how brains function. If this is true, then that have deepened our understanding of how the technological developments that I pointed to biological neural networks may perform sophist- in the previous section may be sufficient for icated information processing. Work using the scaling to sophisticated cognitive agents. How- Neural Engineering Framework (NEF) has res- ever, I am not convinced that this is the case. ulted in a wide variety of spiking neural models As a result, I think that theoretical devel- that mirror data recorded from biological sys- opments in deep learning, nonlinear adaptive tems (Eliasmith & Anderson 1999, 2003). In ad- control, high dimensional brain-like computing, dition, the closely related liquid computing and biological cognition combined will be im- (Maass et al. 2002) and FORCE learning portant to support continued advances in un- (Sussillo & Abbott 2009) paradigms have been derstanding how the mind works. For instance, successfully exploited by a number of research- deep networks continue to achieve state-of-the- ers to generate interesting dynamical systems art results in a wide variety of perception-like that often closely mirror biological data. To- processing challenges (http://rodrigob.git- gether these kinds of methods provide quantit- hub.io/are_we_there_yet/build/classification_ ative characterizations of the computational datasets_results.html#43494641522d3130). And power available in biologically plausible neural while deep networks have traditionally been networks. Such developments are crucial for ex- used for static processing, such as an image ploiting neuromorphic approaches to building classification or document classification, there brain-like hardware. And they suggest ways of has been a recent, concerted move to use them testing some of the more abstract perceptual to model more dynamic perceptual tasks as well and control ideas in real-world, brain-like imple- (Graves et al. 2013). In essence, deep networks mentations. are one among many techniques for modeling Interestingly, several authors have suggested the statistics of time varying signals, a skill that difficult perceptual and control problems are central to animal cognition. in fact mathematical duals of one another (To- However, animals are also incredibly ad- dorov 2009; Eliasmith 2013). This means that ept at controlling nonlinear dynamical sys- there are deep theoretical connections between tems, including their bodies. That is, biolo- perception and motor control. This realization gical brains can generate time varying signals points to a need to think hard about how diverse that allow successful and sophisticated inter- aspects of brain function can be integrated into actions with their environment through their single, large-scale models. This has been a major body. Critically, there have been a variety of focus of research in my lab recently. One result of important theoretical advances in nonlinear this focus is Spaun, currently the world’s largest and adaptive control theory as well. New functional brain model. This model incorporates methods for solving difficult optimal control deep networks, recent control methods, and the problems have been discovered through careful NEF to perform eight different perceptual, motor, study of biological motor control (Schaal et al. and cognitive tasks (Eliasmith et al. 2012). Im- 2007; Todorov 2008). In addition, advances in portantly, this is not a one-off model, but rather a hierarchical control allow for real-time compu- single example among many that employs a gen- tation of difficult inverse kinematics problems eral architecture intended to directly address in- on a laptop (Khatib 1987). And, finally, im- tegrated biological cognition (Eliasmith 2013). portant advances in adaptive control allow for Currently, the most challenging constraints for the automatic learning of both kinematic and running models like Spaun are technological— dynamic models even in highly nonlinear and computers are not fast enough. However, the

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 6 | 17 www.open-mind.net neuromorphic technologies mentioned previously real biological systems work that remain un- should soon remove these constraints. So, in some answered. Of course, complete knowledge of sense, theory currently outstrips application: we natural systems is not a prerequisite for build- have individually tested several critical assump- ing nearly functionally equivalent systems (see tions of the model and shown that they scale well e.g., flight). However, I believe our understand- (Crawford et al. 2013), but we are not yet able to ing of natural cognitive systems will continue to integrate full-scale versions of the components due play an important role in deciding what kinds to limitations in current computational resources. of algorithms are worth pursuing as we build Taken together, I believe that these recent more sophisticated artificial agents. theoretical developments demonstrate that we Fortunately, on this front there have been have a roadmap for how to approach the prob- two announcements of significant resources ded- lem of building sophisticated models of biolo- icated to improving our knowledge of the brain, gical cognition. No doubt not all of the methods which I mentioned in the introduction. One is we need are currently available, but it is not from the EU and the other from the US. Each evident that there are any major conceptual are investing over $1 billion in generating the roadblocks to building a cognitive system that kind of data needed to fill gaps in our under- rivals the flexibility, adaptability, and robust- standing of how brains function. The EU’s Hu- ness of those found in nature. I believe this is a man Brain Project (HBP) includes two central unique historical position. In the heyday of the subprojects aimed at gathering mouse and hu- symbolic approach to AI there were detractors man brain data to complement the large-scale who said that the perceptual problems solved models being built within the project. These easily by biological systems would be a chal- subprojects will focus on genetic, cellular, vas- lenge for the symbolic approach (Norman 1986; cular, and overall organizational data to com- Rumelhart 1989). They were correct. In the plement the large-scale projects of this type heyday of connectionism there were detractors already available (such as the Allen Brain Atlas, who said that standard approaches to artificial http://www.brain-map.org/). One central goal neural networks would not be able to solve diffi- of these subprojects is to clarify the relationship cult planning or syntactic processing problems between the mouse (which is highly experiment- (Pinker & Prince 1988; Fodor & Pylyshyn 1988; ally accessible) and human subjects. Jackendoff 2002). They were correct. In the The American “brain research through ad- heyday of statistical machine learning ap- vancing innovative neurotechnologies” (BRAIN) proaches (a heyday we are still in) there are de- initiative is even more directly focused on large- tractors who say that mountains of data are not scale gathering of neural data. Its purpose is to sufficient for solving the kinds of problems faced accelerate technologies to provide large-scale dy- by biological cognitive systems (Marcus 2013). namic information about the brain that demon- They are probably correct. However, as many of strates how both single runs and larger neural cir- the insights of these various approaches are cuits operate. Its explicit goal is to “fill major combined with control theory, integrated into gaps in our current knowledge” (http://www.ni- models able to do efficient syntactic and se- h.gov/science/brain/). It is a natural complement mantic processing with neural networks, and, in to the human connectome project, which has general, become conceptually unified (Eliasmith been mapping the structure of the human brain 2013), it is less and less obvious what might be on a large-scale (http://www.humanconnec- missing from our characterization of biological tomeproject.org/). Even though it is not yet clear cognition. exactly what information will be provided by the BRAIN intiative, it is clear that significant re- 4 Empirical developments sources are being put into developing technologies that draw on nanoscience, informatics, engineer- One thing that might be missing is, simply, ing, and other fields to measure the brain at a knowledge. We have many questions about how level of detail and scale not previously possible.

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 7 | 17 www.open-mind.net

Table 1

While both of these projects are just over a 5 A future timeline year old, they have both garnered international attention and been rewarded with sufficient fund- Until this point I have been mustering evidence ing to ensure a good measure of success. Con- that there will soon be significant improvements sequently, it is likely that as we build more soph- in our ability to construct artificial cognitive isticated models of brain function, and as we dis- agents. However, I have not been very specific cover where our greatest areas of ignorance lay, about timing. The purpose of this section is to we will be able to turn to the methods developed provide more quantification on the speed of de- by these projects to rapidly gain critical informa- velopment in the field. tion and continue improving our models. In short, In Table 1, the first column specifies the I believe that there is a confluence of technolo- timeframe, the second suggests the number of gical, theoretical, and empirical developments neurons that will be simulatable in real-time on that will allow for bootstrapping detailed func- standard hardware, the third suggests the num- tional models of the brain. It is precisely these ber of neurons that will be simulatable in real- kinds of models that I expect will lead to the time on neuromorphic hardware, and the last most convincing embodiments of artificial cogni- identifies relevant achievable behaviours within tion that we have ever seen—I am even willing to that timeframe. suggest that their sophistication will rival those of I believe that several of the computational natural cognitive systems. technologies I have mentioned, as well as empir-

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 8 | 17 www.open-mind.net ical methods for gathering evidence, are on an Being a philosopher, I am certain that, for exponential trajectory by relevant measures any contemporary problem we consider, at least (e.g., number of neurons per chip, number of some subset of those who have a committed neurons recorded Stevenson & Kording 2011). opinion about that problem will not admit that On the technological side, if we assume a doub- any amount of technical advance can “solve” it. ling every eighteen months, this is equivalent to I suspect, however, that their opinions may end an increase of about one order of magnitude up carrying about as much weight as a modern- every five years. I should also note that I am as- day vitalist. To take one easy example, let us suming that real-time simulation of neurons will think for a moment about contemporary dual- be embedded in an interactive, real-world envir- ism. Some contemporary dualists hold that even onment, and that the neuron count is for the if we had a complete understanding of how the whole system (not a single chip). For context, it brain functions, we would be no closer to solv- is worth remembering that the human brain has ing the “hard problem” of consciousness about 1011 neurons, though they are more com- (Chalmers 1996). The “hard problem” is the putationally sophisticated than those typically problem of explaining how subjective experience simulated in hardware. comes from neural activity. That is, how the Another caveat is that it is likely that phenomenal experiences we know from a first- large-scale simulations on a digital Von Neu- person perspective can be accounted for by mann architecture will hit a power barrier, third-person physicalist approaches to under- which makes it likely that the suggested scaling standing the mind. If indeed we have construc- could be achieved, but will be cost-prohibitive ted artificial agents that behave much like in fifty years. Consequently, a neuromorphic al- people, share a wide variety of internal states ternative is most likely to be the standard im- with people, are fully empirically accessible, and plementational substrate of artificial agents. report experiences like people, it is not obvious Finally, the behavioural characterizations I to what extent this problem will not have been am giving are with a view to functions neces- solved. Philosophers who are committed to the sary for creating a convincing artificial mind in notion that no amount of empirical knowledge an artificial body. Consequently, my comments will solve the problem will of course dismiss generally address perceptual, motor, and cognit- such an accomplishment on the strength of their ive skills relevant to reproducing human-like intuitions. I suspect, however, that when most abilities. people are actually confronted with such an agent—one they can interrogate to their heart’s 6 Consequences for philosophy content and one about which they can have complete knowledge of its functioning—it will So suppose that, fifty years hence, we have de- seem odd indeed to suppose that we cannot ex- veloped an understanding of cognitive systems plain how its subjective experience is generated. that allows us to build artificial systems that I suspect it will seem as odd as someone are on par with, or, if we see fit, surpass the nowadays claiming that we cannot expect to ex- abilities of an average person. Suppose, that is, plain how life is generated despite our current that we can build artificial agents that can understanding of biochemistry. Another way to move, react, adapt, and think much like human put this is that the “strong intuitions” of con- beings. What consequences, if any, would this temporary dualists will hold little plausibility in have for our theoretical questions about cogni- the face of actually existing, convincing artificial tion? I take these questions to largely be in the agents, and so, I suspect, they will become even domain of philosophy of mind. In this section I more of a rarity. consider several central issues in philosophy of I refer to this example as “easy” because mind and discuss what sorts of consequences I the central reasons for rejecting dualism are take building a human-like artificial agent to only strengthened, not generated, by the exist- have for them. ence of sophisticated artificial minds. That is,

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 9 | 17 www.open-mind.net good arguments against the dualist view are pects of both statistical perceptual processing more or less independent of the current state of and syntactic manipulation. Even in simple constructing agents (although the existence of models, like Spaun, it is clear how the symbols such agents will likely sway intuitions). How- for digits that are syntactically manipulated are ever, other philosophical conundrums, like related to inputs coming from the external Searle’s famous Chinese room (1980), have re- world (Eliasmith 2013). And it is clear how sponses that depend fairly explicitly on our those same symbols can play a role in driving ability to construct artificial agents. In particu- the model’s body to express its knowledge lar, the “systems reply” suggests that a suffi- about those representations. As a result, the ciently complex system will have the same in- tasks that Spaun can undertake demonstrate tentional states as a biological cognitive system. both conceptual knowledge, through the sym- For those who think that this is a good rejec- bol-like relationships between numbers (e.g., in tion of Searle’s strong intentionalist views, hav- the counting task), and perceptual knowledge, ing systems that meet all the requirements of through categorization and the ability to drive their currently hypothetical agents would its motor system to reproduce visual properties provide strong empirical evidence consistent (e.g., in the copy-drawing task). with their position. Of course, the existence of In some cases, rather than resolving philo- such artificial agents is unlikely to convince sophical debates, the advent of sophisticated ar- those, like Searle, who believe that there is tificial agents is likely to make these debates far some fundamental property of biology that al- more empirically grounded. These include de- lows intentionality to gain a foothold. But it bates about the nature of concepts, conceptual does make such a position seem that much more change, and functionalism, among others. How- tenuous if every means of measuring intentional- ever these debates turn out, it seems clear that ity produces similar measurements across non- having an engineered, working system that can biological and biological agents. In any case, the generate behaviour as sophisticated as that that realization of the systems reply does ultimately gave rise to these theoretical ideas in the first depend on our ability to construct sufficiently place will allow a systematic investigation of sophisticated artificial agents. And I am sug- their appropriate application. After all, there gesting that such agents are likely to be avail- are few, if any, limits on the empirical informa- able in the next fifty years. tion we can garner from such constructed sys- More immediately, I suspect we will be tems. In addition, our having built the system able to make significant headway on several explicitly makes it unlikely that we would be problems that have been traditionally con- unaware of some “critical element” essential in sidered philosophical before we reach the fifty- generating the observed behaviours. year mark. For example, the frame problem— Even without such a working system, I be- i.e., the problem of knowing what representa- lieve that there are already hints as to how tional states to update in a dynamic environ- these debates are likely to be resolved, given the ment—is one that contemporary methods, like theoretical approaches I highlighted earlier. For control theory and machine learning, struggle instance, I suspect that we will find that con- with much less than classical methods. Because cepts are explained by a combination of vector the dynamics of the environment are explicitly space representations and a restricted class of included in the world-model being exploited by dynamic processes defined over those spaces such control theoretic and statistical methods, (Eliasmith 2013). Similarly, quantifying the ad- updating state representations naturally in- aptive nature of those representations and pro- cludes the kinds of expectations that caused cesses will indicate the nature of mechanisms of such problems for symbolic approaches. conceptual change in individuals (Thagard Similarly, explicit quantitative solutions 2014). In addition, functionalism will probably are suggested for the symbol-grounding problem seem too crude a hypothesis given a detailed through integrated models that incorporate as- understanding of how to build a wide variety of

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 10 | 17 www.open-mind.net artificial minds. Perhaps a kind of “functional- analysis. This is a unique quandary because ism with error bars” will take its place, provid- while it is currently possible for certain indi- ing a useful means of talking about degrees of viduals to claim to have such slave-aligned functional similarity and allowing a quantifica- goals, it is always possible to argue that they tion of functional characterizations of complex are simply mistaken in their personal psycholo- systems. Consequently, suggestions about which gical analysis. In the case of minds whose psy- functions are or are not necessary for “minded- chology is designed in a known manner, how- ness” can be empirically tested through explicit ever, the having of such goals will at least seem implementation and experimentation. This will much more genuine. This is only one among not solve the problem of mapping experimental many new kinds of challenges that ethical the- results to conceptual claims (a problem we cur- ory will face with the development of sophistic- rently face when considering non-human and ated artificial agents (Metzinger 2013). even some human subjects), but it will make I do not take this surely unreasonably functionalism as empirically accessible as seems brief discussion of any of these subtle philosoph- plausible. ical issues to do justice to them. My main pur- In addition to these philosophical issues pose here is to provide a few example instances that may undergo reconceptualization with the of how the technological developments discussed construction of artificial minds, there are others earlier are likely to affect our theoretical in- that are bound to become more vexing. For ex- quiry. On some occasions such developments ample, the breadth of application of ethical the- will lead to strengthening already common intu- ory may, for the first time, reach to engineered itions; on others they may provide deep empir- devices. If, after all, we have built artificial ical access to closely related issues; and on still minds capable of understanding their place in other occasions these developments will serve to the universe, it seems likely we will have to make complex issues even more so. worry about the possibility of their suffering (Metzinger 2013). It does not seem that under- 7 The good and the bad standing how such devices work, or having explicitly built them, will be sufficient for dismiss- As with the development of many technologies ing them as having no moral status. While cur- —cars, electricity, nuclear power—the construc- rent theories of non-human ethics have been de- tion of artificial minds is likely to have both veloped, it is not clear how much or little theor- negative and positive impacts. However, there is ies of non-biological ethics will be able to bor- a sense in which building minds is much more row from them. fraught than these other technologies. We may, I suspect that the complexities introduced after all, build agents that are themselves cap- to ethical theory will go beyond adding a new able of immorality. Presumably we would much category of potential application. Because artifi- prefer to build Commander Data than to build cial minds will be designed, they may be de- HAL or the Terminator. But how to do this is signed to make what have traditionally been by no means obvious. There have been several morally objectionable inter-mind relationships interesting suggestions as to how this might be seem less problematic. Consider, for instance, a accomplished, perhaps most notably from Isaac robot that is designed to gain maximal self-ful- Asimov in his entertaining and thought-provok- fillment out of providing service to people. That ing exploration of the three laws of robotics. For is, unlike any biological species of which we are my purposes, however, I will sidestep this issue aware, these robots place service to humans —not because it is not important, but because above all else. Is a slave-like relationship more immediate concerns arise from considering between humans and these minds still wrong in the development of these agents from a techno- such an instance? Whatever our analysis of why logical perspective. Let me then focus on the slavery is wrong, it seems likely that we will be more immediately pressing consequences of con- able to design artificial minds that bypass that structing intelligent machines.

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 11 | 17 www.open-mind.net

The rapid development of technologies re- problematic societal shifts (Padbury et al. lated to artificial intelligence has not escaped 2014). the notice of governments around the world. Indeed, many of the benefits of automatiz- One of the primary concerns for governments is ation may help alleviate the potential down- the potentially massive changes in the nature of sides. Automatization has already had signific- the economy that may result from an increase ant impact on the growth of new technology, in automatization. It has recently been sugges- both speeding up the process of development ted that almost half of the jobs in the United and making new technology cheaper. The hu- States are likely to be computerized in the next man genome project was a success largely be- twenty years (Rutkin 2013). The US Bureau of cause of the automatization of the sequencing Labor and Statistics regularly publishes articles process. Similarly, many aspects of drug discov- on the significant consequence of automation for ery can be automatized by using advanced com- the labour force in their journal Monthly Labor putational techniques (Leung et al. 2013). Auto- Review (Goodman 1996; Plewes 1990). This matization of more intelligent behaviour than work suggests that greater automatization of simply generating and sifting through data is jobs may cause standard measures of productiv- likely to have an even greater impact on the ad- ity and output to increase, while still increasing vancement of science and engineering. This may unemployment. lead more quickly to cleaner and cheaper en- Similar interest in the economic and social ergy, advances in manufacturing, decreases in impacts of automatization is evident in many the cost and access to advanced technologies, other countries. For instance, Policy Horizons and other societal benefits. Canada is a think-tank that works for the Ca- As a consequence, manufacturing is likely nadian government, which has published work to become safer—a trend already seen in areas on the effects of increasing automatization and of manufacturing that employ large numbers the future of the economy (Arshad 2012). Soon of robots (Robertson et al. 2005). At the same after the publication of our recent work on time, additional safety considerations come Spaun, I was contacted by this group to discuss into play as robotic and human workspaces the impact of Spaun and related technologies. It themselves begin to interact. This concern has was clear from our discussion that machine resulted in a significant focus in robotics on learning, automated control, robotics, and so on compliant robots. Compliant robots are those are of great interest to those who have to plan that have “soft” environmental interactions, for the future, namely our governments and often implemented by including real or virtual policy makers (Padbury et al. 2014). springs on the robotic platform. As a result, This is not surprising. A recent McKinsey control becomes more difficult, but interac- report suggests that these highly disruptive tions become much safer, since the robotic technologies are likely to have an economic system does not rigidly go to a target position value of about $18 trillion by 2025 (Manyika et even if there is an unexpected obstacle (e.g., a al. 2013). It is also clear from the majority of person) in the way. analyses, that lower-paid jobs will be the first As the workplace continues to become affected, and that the benefits will accrue to one where human and automated systems co- those who can afford what will initially be ex- operate, additional concerns may arise as to pensive technologies. Every expectation, then, is what kinds of human-machine relationships that automatization will exacerbate the already employers should be permitted to demand. large and growing divide between rich and poor Will employees have the right not to work (Malone 2014; “The Future of Jobs: The On- with certain kinds of technology? Will employ- rushing Wave” 2014). Being armed with this ers still have to provide jobs to employees who knowledge now means that individuals, govern- refuse certain work situations? These ques- ments, and corporations can support progessive tions touch on many of the same subjects policies to mitigate these kinds of potentially highlighted in the previous section regarding

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 12 | 17 www.open-mind.net the ethical challenges that will be raised as we because of the robots. Challenges to societal develop more and more sophisticated artificial stability are nothing new: war, hunger, poverty, minds. weather are constant destabilizing forces. Artifi- Finally, much has been made of the pos- cial minds are likely to introduce another force, sibility that the automatization of technological but one that may be just as likely to be stabiliz- advancement could eventually result in ma- ing as problematic. chines designing themselves more effectively Unsurprisingly, like many other technolo- than humans can. This idea has captured the gical changes, the development of artificial public imagination, and the point in time where minds will bring with it both costs and benefits. this occurs is now broadly known as “The Sin- It may even be the case that deciding what is a gularity,” a term first introduced by von Neu- cost and what is a benefit is not straightfor- mann (Ulam 1958). Given the vast variety of ward. If indeed many jobs become automated, it functions that machines are built to perform, it would be unsurprising if the average working seems highly unlikely that there will be any- week becomes shorter. As a result, a large num- thing analogous to a mathematical singularity— ber of people may have much more recreational a clearly defined, discontinuous point—after time than has been typical in recent history. which machines will be superior to humans. As This may seem like a clear benefit, as many of with most things, such a shift, if it occurs, is us look forward to holidays and time off work. likely to be gradual. Indeed, the earlier timeline However, it has been argued that fulfilling work is one suggestion for how such a gradual shift is a central to human happiness (Thagard might occur. Machines are already used in 2010). Consequently, overly limited or unchal- many aspects of design, for performing optimiz- lenging work may end up being a significant ations that would not be possible without them. cost of automation. Machines are also already much better at many As good evidence for costs and benefits functions than people: most obviously mechan- becomes available, decision-makers will be faced ical functions, but more recently cognitive ones, with the challenge of determining what the ap- like playing chess and answering trivia questions propriate roles of artificial minds should be. in certain circumstances. These roles will no doubt evolve as technologies Because the advancement of intelligent change, but there is little reason to presume machines is likely to continue to be a smooth, that unmanageable upheavals or “inflection continuous one (even if exponential at times), points” will be the result of artificial minds be- we will likely remain in a position to make in- ing developed. While we, as a society, must be formed decisions about what they are permitted aware of, and prepared for, being faced with to do. As with members of a strictly human so- new kinds of ethical dilemmas, this has been a ciety, we do not tolerate arbitrary behaviour regular occurrence during the technological de- simply because such behaviour is possible. If velopments of the last several hundred years. anything, we will be in a better position to spe- Perhaps the greatest challenges will arise be- cify appropriate behaviour in machines than we cause of the significant wealth imbalances that are in the case of our human peers. Perhaps we may be exacerbated by limited access to more will need laws and other societal controls for de- intelligent machines. termining forbidden or tolerable behaviour. Per- haps some people and machines will choose to 8 Conclusion ignore those laws. But, as a society, it is likely that we will enforce these behavioural con- I have argued that we are at a unique point in straints the same way we do now—with public- the development of technologies that are critical ally sanctioned agencies that act on behalf of to the realization of artificial minds. I have even society. In short, the dystopian predictions we gone so far as to predict that human-level intel- often see that revolve around the development ligence and physical ability will be achieved in of intelligent robots seem no more or less likely about fifty years. I suspect that for many famil-

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 13 | 17 www.open-mind.net iar with the history of artificial intelligence such clear. Similarly, sensors on the scale and precision predictions will be easily dismissed. Did we not of those available from nature are not yet avail- have such predictions over fifty years ago? Some able. This is less true for vision and audition, but have suggested that the singularity will occur definitely the case for proprioception and touch. by 2030 (Vinge 1993), others by 2045 (Kurzweil The latter are essential for fluid, rapid motion 2005). There were suggestions and significant control. It also remains to be seen how well our financial speculation that AI would change the theoretical methods for integrating complex sys- world economy in the 1990s, but this never tems will scale. This will only become clear as we happened. Why would we expect anything to be attempt to construct more and more sophistic- different this time around? ated systems. This is perhaps the most fragile as- In short, my answer is encapsulated by the pect of my prediction: expecting to solve difficult specific technological, theoretical, and empirical algorithmic and integration problems. And, of developments I have described above. I believe course, there are myriad other possible ways in that they address the central limitations of previ- which I may have underestimated the complexity ous approaches to artificial cognition, and are sig- of biological cognition: maybe glial cells are per- nificantly more mature than is generally appreci- forming critical computations; maybe we need to ated. In addition, the limitations they address— describe genetic transcription processes in detail such as power consumption, computational scal- to capture learning; maybe we need to delve to ing, control of nonlinear dynamics, and integrat- the quantum level to get the explanations we ing large-scale neural systems—have been more need—but I am doubtful (Litt et al. 2006). central to prior failures than many have realized. Perhaps it goes without saying that, all Furthermore, the financial resources being direc- things considered, I believe the timeline I propose ted towards the challenge of building artificial is a plausible one.1 This, of course, is predicated on minds is unprecedented. High-tech companies, in- there being the societal and political will to allow cluding Google, IBM, and Qualcomm have inves- the development of artificial minds to proceed. No ted billions of dollars in machine intelligence. In doubt researchers in this field need to be respons- addition, funding agencies including DARPA (De- ive to public concerns about the specific uses to fense Advanced Research Projects Agency), EU- which such technology might be put. It will be im- IST (European Union—Information Society Tech- portant to remain open, self-critical, and self-regu- nologies), IARPA (Intelligence Advanced Re- lating as artificial minds become more and more search Projects Agency), ONR (Office of Naval capable. We must usher in these technologies with Research), and AFOSR (Air Force Office of Sci- care, fully cogniscent of, and willing to discuss, entific Research) have contributed a similar or both their costs and their benefits. greater amount of financial support across a wide range of projects focused on brain-inspired com- Acknowledgements puting. And the two special billion dollar initiat- ives from the US and EU will serve to further I wish to express special thanks to two anonym- deepen our understanding of biological cognition, ous reviewers for their helpful feedback. Many which has, and will continue, to inspire builders of of the ideas given here were developed in discus- artificial minds. sion with members of the CNRG Lab, parti- While I believe that the alignment of these cipants at the Telluride workshops, and my col- forces will serve to underpin unprecedented ad- laborators on ONR grant N000141310419 (PIs: vances in our understanding of biological cogni- Kwabena Boahen and Rajit Manohar). This tion, there are several challenges to achieving the work was also funded by AFOSR grant FA8655- timeline I suggest above. For one, robotic actuat- 13-1-3084, Canada Research Chairs, and ors are still far behind the efficiency and speeds NSERC Discovery grant 261453. found in nature. There will no doubt be advances in materials science that will help overcome these 1 It is quite different from that proposed by the HBP, for example. For further discussion of the differences in perspective between the HBP limitations, but how long that will take is not yet and my lab’s work, see Eliasmith & Trujillo (2013).

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 14 | 17 www.open-mind.net

References Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y. & Rasmussen, D. (2012). A large- Ackerman, E. (2014). Robots playing ping pong: What’s scale model of the functioning brain. Science, 338 real, what’s not? IEEE Spectrum (6111), 1202-1205. 10.1126/science.1225266 Amazon (2013). Amazon Prime Air. Eliasmith, C. & Trujillo, O. (2013). The use and abuse of Arshad, I. (2012). People and machines: Competitors or large-scale brain models. Current Opinion in Neurobio- collaborators in the emerging world of work? logy, 25, 1-6. 10.1016/j.conb.2013.09.009 Chalmers, D. (1996). The conscious mind: In search of a Esser, S. K., Andreopoulos, A., Appuswamy, R., Datta, fundamental theory. Oxford, UK: Oxford University P., Barch, D., Amir, A., Arthur, J., Cassidy, A., Flick- Press. ner, M., Merolla, P., Chandra, S., Basilico, N., Carpin, Cheah, C. C., Liu, C. & Slotine, J. J. E. (2006). Adaptive S., Zimmerman, T., Zee, F., Alvarez-Icaza, R., Kusnitz, tracking control for robots with unknown kinematic J. A., Wong, T. M., Risk, W. P., McQuinn, E., Nayak, and dynamic properties. The International Journal of T. K., Singh, R. & Modha, D. S. (2013). Cognitive Robotics Research, 25 (3), 283-296. SAGE Publications. computing systems: Algorithms and applications for Choudhary, S., Sloan, S., Fok, S., Neckar, A., Trautmann, networks of neurosynaptic cores. The 2013 Interna- E., Gao, P., Stewart, T., Eliasmith, C. & Boahen, K. tional Joint Conference on Neural Networks (IJCNN) (2012). Silicon neurons that compute. In A. E. P. Villa, (pp. 1-10). IEEE. 10.1109/IJCNN.2013.6706746 W. Duch, P. Érdi, F. Masulli and G. Palm (Eds.) Arti- Fodor, J. A. & Pylyshyn, Z. W. (1988). Connectionism ficial neural networks and machine learning - ICANN and cognitive architecture: A critical analysis. Cogni- 2012 (pp. 121-128). Berlin, GER: Springer. tion, 28 (1-2), 3-71. 10.1016/0010-0277(88)90031-5 10.1007/978-3-642-33269-2_16 Fox, D. (2009). IBM reveals the biggest artificial brain of Corradi, F., Eliasmith, C. & Indiveri, G. (2014). Mapping all time. Popular Mechanics arbitrary mathematical functions and dynamical sys- Galluppi, F., Denk, C., Meiner, M. C., Stewart, T., tems to neuromorphic VLSI circuits for spike-based Plana, L. A., Eliasmith, C., Furber, S. & Conradt, J. neural computation. IEEE International Symposium (2014). Event-based neural computing on an autonom- on Circuits and Systems (ISCAS) 2014 (pp. 269-272). ous mobile platform. Proceedings of IEEE international IEEE. 10.1109/ISCAS.2014.6865117 conference on robotics and automation (ICRA). IEEE. Crawford, E., Gingerich, M. & Eliasmith, C. (2013). Bio- De Garis, H., Shuo, C., Goertzel, B. & Ruiting, L. (2010). logically plausible, human-scale knowledge representa- A world survey of artificial brain projects, Part I: tion. In M. Knauff, M. Pauen, N. Sebanz and I. Wachs- Large-scale brain simulations. Neurocomputing, 74 (1- muth (Eds.) 35th Annual Conference of the Cognitive 3), 3-29. 10.1016/j.neucom.2010.08.004 Science Society (pp. 412-417). Austin, TX: Cognitive Goodman, W. C. (1996). Software and engineering indus- Science Society. tries: Threatened by technological change? Monthly Davies, S., Stewart, T., Eliasmith, C. & Furber, S. Labor Review, 119 (8), 37-45. Bureau of Labor Statist- (2013). Spike-based learning of transfer functions with ics, U.S. Department of Labor. 10.2307/41844604 the SpiNNaker neuromimetic simulator. The 2013 In- Graves, A., Mohamed, A. & Hinton, G. E. (2013). speech ternational Joint Conference on Neural Networks recognition with deep recurrent neural networks. IEEE (IJCNN) (pp. 1832-1839). IEEE. international conference on acoustic speech and signal 10.1109/IJCNN.2013.6706962 processing (ICASSP) (pp. 6645-6649). Vancouver, Eliasmith, C. (2013). How to build a brain: A neural ar- Canada: IEEE. chitecture for biological cognition. New York, NY: Ox- Hasler, J. & Marr, H. B. (2013). Finding a roadmap to ford University Press. achieve large neuromorphic hardware systems. Fronti- Eliasmith, C. & Anderson, C.H. (1999). Developing and ers in Neuroscience, 7 (118), 1-29. applying a toolkit from a general neurocomputational 10.3389/fnins.2013.00118 framework. Neurocomputing, 26-27 (0), 1013-1018. Jackendoff, R. (2002). Foundations of language: Brain, 10.1016/S0925-2312(99)00098-3 meaning, grammar, evolution. Oxford, UK: Oxford (2003). Neural engineering: Computation, repres- University Press. entation and dynamics in neurobiological systems. Khan, M. M., Lester, D. R., Plana, L. A., Rast, A., Jin, Cambridge, MA: MIT Press. X., Painkras, E. & Furber, S. B. (2008). SpiNNaker:

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 15 | 17 www.open-mind.net

Mapping neural networks onto a massively-parallel chip Metzinger, T. (2013). Two principles for robot ethics. In multiprocessor. IEEE. (pp. 2849-2856). E. Hilgendorf and J. P. Günther (Eds.) Robotik und 10.1109/IJCNN.2008.4634199 Gesetzgebung (pp. 263-302). Baden-Baden, GER: Khatib, O. (1987). A unified approach for motion and Nomos. force control of robot manipulators: The operational Newquist, H. P. (1994). The brain makers. Indianapolis, space formulation. IEEE Journal of Robotics and IN: Sams Publishing. Automation, 3 (1), 43-53. Norman, D. A. (1986). Reflection on cognition and paral- Krizhevsky, A., Sutskever, I. & Hinton, G. E. (2012). Im- lel distributed processing. In J. L. McClelland and D. ageNet classification with deep convolutional neural E. Rumelhart (Eds.) Parallel distributed processing: networks. In F. Pereira, C. J. C. Burges, L. Bottou and Exploration inthe microstructure of cognition (p. 531). K. Q. Weinberger (Eds.) Advances in Neural Informa- Cambridge, MA: MIT Press. tion Processing Systems 25 (p. 4). Padbury, P., Christensen, S., Wilburn, G., Kunz, J. & Kumar, S. (2013). Introducing qualcomm zeroth pro- Cass-Beggs, D. (2014). MetaScan 3: Emerging techno- cessors: Brain-inspired computing. http://www.qual- logies. MetaScan 3: Emerging technologies (p. 45). comm.com/media/blog/2013/10/10/introducing-qual- Cambridge, MA: Policy Horizons Canada. comm-zeroth-processors-brain-inspired-computing http://www.horizons.gc.ca/eng/content/metascan-3- Kurzweil, R. (2005). The singularity is near. New York, emerging-technologies-0 NY: Penguin Books. Pinker, S. & Prince, A. (1988). On language and connec- Leung, E. L., Cao, Z., Jiang, Z., Zhou, H. & Liu, L. tionism: Analysis of a parallel distributed processing (2013). Network-based drug discovery by integrating model of language acquisition. Cognition, 28 (1-2), 73- systems biology and computational technologies. Brief- 193. 10.1016/0010-0277(88)90032-7 ings in bioinformatics, 14 (4), 491-505. Oxford, UK: Plewes, T. J. (1990). Labor force data in the next cen- Oxford University Press. tury. Monthly Labor Review, 113 (4). Bureau of Labor Lichtsteiner, P., Posch, C. & Delbruck, T. (2008). Tem- Statistics, U.S. Department of Labor. poral contrast vision sensor. IEEE Journal of Solid- Print Edition (2014). The future of jobs: The onrushing State Circuits, 43 (2), 566-576. wave. The Economist. Li, C., Delbruck, T. & Liu, S. (2012). Real-time speaker Robertson, J., Sheppard, T. & Sarnes, S. (2005). Workers identification using the AEREAR2 event-based silicon compensation claim frequency continues to decline, cochlea. 2012 IEEE International Symposium on Cir- particularly for smaller claims. NCCI Research Brief, 2 cuits and Systems (pp. 1159-1162). Rumelhart, D. E. (1989). The architecture of mind: A 10.1109/ISCAS.2012.6271438 connectionist approach. In M. I. Poser (Ed.) The archi- Litt, A., Eliasmith, C., Kroon, F., Weinstein, S. & tecture of mind: A connectionist approach (pp. 133- Thagard, P. (2006). Is the brain a quantum computer? 159). Cambridge, MA: MIT Press. Cognitive Science, 30 (3), 593-603. Rutkin, A. H. (2013). Report suggests nearly half of U.S. Maass, W., Natschläger, T. & Markram, H. (2002). Real- jobs are vulnerable to computerization. MIT Techno- time computing without stable states: A new frame- logy Review. work for neural computation based on perturbations. Schaal, S., Mohajerian, P. & Ijspeert, A. (2007). Dynam- Neural computation, 14 (11), 2531-2560. Cambridge, ics systems vs. optimal control: A unifying view. Pro- MA: MIT Press. gress in brain research, 165 (1), 425-45. Malone, P. (2014). Wealthy need to share the spoils of 10.1016/S0079-6123(06)65027-9 automation. http://www.canberratimes.com.au/com- Searle, J. R. (1980). Minds, brains, and programs. Behavi- ment/wealthy-need-to-share-the-spoils-of-automation- oral and Brain Sciences, 3 (03), 417-424. Cambridge, UK: 20140301-33sp6.html Cambridge University Press. 10.1017/S0140525X00005756 Manyika, J., Chui, M., Bughin, J. & Dobbs, R. (2013). Simon, H. A. & Newell, A. (1958). Heuristic problem Disruptive technologies: Advances that will transform solving: The next advance in operations research. Op- life, business, and the global economy. McKinsey erations Research, 6 (1), 1-10. Global Institute. Stevenson, I. H. & Kording, K. P. (2011). How advances Marcus, G. (2013). Is “Deep Learning” a revolution in ar- in neural recording affect data analysis. Nature neuros- tificial intelligence? The New Yorker. cience, 14 (2), 139-42. 10.1038/nn.2731

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 16 | 17 www.open-mind.net

Stunt, V. (2014). Why Google is buying a seemingly crazy collection of companies. CBC News. Sussillo, D., Abbott, L. F. (2009). Generating coherent patterns of activity from chaotic neural networks. Neuron, 63 (4), 544-57. 10.1016/j.neuron.2009.07.018 Thagard, P. (2010). The brain and the meaning of life. Princeton, NJ: Princeton University Press. (2014). Explanatory identities and conceptual change. Science & Education, 23 (7), 1531-1548. 10.1007/s11191-014-9682-1 Todorov, E. (2008). Optimal control theory. In K. Doya (Ed.) Bayesian brain: Probabilistic approaches to neural coding (pp. 269-298). Cambridge, MA: MIT Press. (2009). Parallels between sensory and motor information processing. In M. S. Gazzaniga (Ed.) The cognitive neurosciences (pp. 613-624). Cambridge, MA: MIT Press. Ulam, S. (1958). Tribute to John von Neumann. Bulletin of the American Mathematical Society, 64 (3), 5. Vinge, V. (1993). The coming technological singularity: How to survive in the post-human era. Vision 21: In- terdisciplinary Science and Engineering in the Era of Cyberspace (pp. 11-22). Westlake, OH: NASA Confer- ence Publication 10129.

Eliasmith, C. (2015). On the Eve of Artificial Minds. In T. Metzinger & J. M. Windt (Eds). Open MIND: 12(T). Frankfurt am Main: MIND Group. doi: 10.15502/9783958570252 17 | 17