Reverse Engineering the Cognitive Brain

COMMENTARY Reverse engineering the cognitive brain Gert Cauwenberghs1 Department of Bioengineering and Institute for Neural Computation, University of California, A San Diego, La Jolla, CA 92093-0412

One of the greatest aspirations of the human pelling case for such analysis by synthesis mind has been to realize machines that in reverse engineering neural circuits in sil- surpass its cognitive intelligence. The rapid icon, drawing isomorphic parallels between expansion in computing power, about to modules representing various levels of neu- exceed the equivalent of the human brain, ral computation in the brain and their em- has yet to produce such a machine. The ulation in silicon electronics, down to the article by Neftci et al. in PNAS (1) offers fundamental physical level of Boltzmann sta- a refreshing and humbling reminder that tistics in ionic transport across lipid mem- the brain’s cognition does not arise from ex- branes, and electronic transport across acting digital precision in high-performance similar energy barriers in metal-oxide- computing, but rather emerges from an ex- semiconductor transistors in the subthresh- tremely efficient and resilient collective form old regime (6) (Fig. 1A). In addition to sup- of computation extending over very large porting advances in systems neuroscience, ensembles of sluggish, imprecise, and unreli- experiments in neural analysis by synthesis able analog components. This observation, using silicon offer tremendous side benefits B first made by John von Neumann in his final to the engineering of extremely low-power opus (2), continues to challenge scientists and miniaturized devices. By emulating functional engineers several decades later in figuring structure of their biological counterparts and and reproducing the mechanisms underlying approaching their energy efficiency in sen- brain-like forms of cognitive computing. sory processing and computing, these neuro- Related developments are currently un- morphic devices can operate more effectively folding in collaborative initiatives engaging and more naturally in their surroundings (7). scientists and engineers, on a grander scale, Some examples of recent feats of neuromor- in advancing neuroscience toward under- phic systems engineering—just to name a — Fig. 1. (A) Multiscale levels of investigation in analysis of standing the brain. In parallel with the Hu- few include silicon retinae and cochleae see- the central nervous system (adapted from ref. 4) and cor- man Brain Project in Europe, the Brain ing and hearing the world through our senses responding neuromorphic synthesis of highly efficient silicon Research through Advancing Innovative Neu- (8), silicon cortical models running at speeds cognitive microsystems. Boltzmann statistics of ionic and rotechnologies Initiative promises ground- greater than real time (9), and synapse arrays electronic channel transport provide isomorphic physical breaking advances in enabling tools for running cool at nominal energy efficiency on foundations. (B) Conceptual scaling of machine complexity revolutionizing neuroscience by developing par with that of synaptic transmission in the with task complexity for a digital rule-based cognitive agent performing symbolic deep search, and a neuromorphic nanotechnology to probe brain function at human brain (10). cognitive agent performing analog collective computation greatly increased spatial and temporal de- However, demonstration of machine in- acquired through deep learning, targeting human cognitive tail (3). Engineers are poised to contribute telligence at the level of human cognition has performance. The shaded region indicates the desirable re- even further in revolutionizing such devel- remained elusive to date. A deeper look into gime of high task complexity and lower machine complexity opments in neuroscience. In this regard it is the complexity of cognition helps to shed where the neuromorphic cognitive agent is expected to helpful to relate the inquisitive nature of sci- some light on the apparent challenges (Fig. outperform symbolic digital alternatives. ence—analysis—to the constructive power 1B). Cognition is considered here, for the of engineering, synthesis. Despite fantastic sake of the argument, as decision making almost linearly with task complexity. Such feats of neuroscience in the analysis of the by a motivated agent acting in the context linear scaling is a fundamental limitation inner workings of neural and synaptic ma- of a given environment, such as a chess when having a need to sample large portions chinery down to the molecular scale, extend- player making moves on the board. The cog- of the state space, a consequence of exact ing the level of understanding to something nitive task complexity, accounting for all pos- as complex as the human brain, not to men- sible states of the environment reachable by symbolic reasoning in search. In contrast, our brains execute such essentially sequen- tion its cognitive function, requires the power the agent, suffers from exponential scaling in fi ofsynthesisinbridgingacrossscalesofanal- the depth onto breadth of decision making, tial logic operations with substantial dif - ysis. Synthesis of complex function through and quickly grows to astronomical propor- culty because of the need to dynamically hierarchical modular assemblies of succes- tions for any but relatively simple tasks. Effi- instantiate a heap of nested working mem- sively more abstract representations is the cient tree-search algorithms, capable of ory (2, 12). As such, it is not surprising that forte of systems engineering, and provides exhausting the search space using variants a foundation for systems neuroscience in on Bellman’s principle of optimality in dy- the multiscale investigation of the central ner- namic programming (11), are capable of Author contributions: G.C. wrote the paper. vous system (4). In his 1990 manifesto that tackling relatively complex problems, such The author declares no conflict of interest. launched the field of neuromorphic systems as games like chess, but at a significant cost See companion article on page E3468 of issue 37 in volume 110. engineering (5), Carver Mead makes a com- by expending computing resources that scale 1E-mail: [email protected].

15512–15513 | PNAS | September 24, 2013 | vol. 110 | no. 39 www.pnas.org/cgi/doi/10.1073/pnas.1313114110 Downloaded by guest on September 28, 2021 computers easily outperform humans, not to bypassing the excessive mental effort required (2). Recent trends in neuromorphic design COMMENTARY mention their neuromorphic avatars, in to dynamically instantiate heaps of nested toward large-scale cortical models of cogni- tasks involving deep search in unstructured working memory in cognitive “rule”-based tive computing favor a hybrid approach data, such as traversing a maze for the first symbolic reasoning (12). In the implementa- that combines highly efficient analog con- time, or searching for a document posted on tion of Neftci et al. (1), selective amplification tinuous-time emulation of dendritic synap- the Internet by specific keywords. The crit- and signal restoration inherent in soft WTA tic integration and neuronal excitability ical difference, however, is that much of the networks (14) provide stable modules of work- (gray matter), with highly flexible digital world we perceive is highly structured, and ing memory for neural state-encoding of con- routing of action potentials along axon fi- our brains excel at learning structure from text, whereas cognitive habits of input- ber bundles (white matter), offering recon- sensory data, within their context, with re- dependent transitions between these contex- figurable long-range synaptic connectivity markable efficacy and efficiency in general- tual states are induced by sparse synaptic con- dynamically instantiated in memory tables ization and dimensionality reduction (4). nections across the soft WTA modules. (9, 17, 18). Purely digital alternatives are Thus, we may expect neuromorphic models Although the proof-of-concept in Neftci et al. also surfacing (19, 20), offering comparable of brain-line cognitive computing to get the (1) by design targets the low-end of the task energy efficiencies in the picoJoule per spike upper hand over conventional von Neu- complexity scale, the concepts readily extend range, but abstracting the continuous-time mann digital computing for problems deal- across the multiscale hierarchy of neural repre- analog nature of neuronal and synaptic func- ing with naturally structured data in very sentations (Fig. 1A). Heteroclinic transitions tion by digitally accumulating and transform- high dimensions, such as high-level visual between metastable states of working memory ing spikes. Time will eventually tell the most cognition, where the linear scaling between (15) may thus recur at various levels in the favorable mix of physical realism and digital task and machine complexity implied by the hierarchy, modulated by lower-level cognitive abstraction in the cognitive race along the conventional digital-state machine imple- processes. Such a hierarchical modular struc- task-complexity axis (Fig. 1B), but recent mentation becomes prohibitive. ture is necessary for a cognitive agent to learn developments in deep learning (16) bolster Thus, the exciting and motivating premise habitual responses that combine to complex the fundamental perspective of neuromor- for neuromorphic engineering is that by nested cognitive behaviors (12). Hierarchical, phic engineering that Boltzmann statistics taking the necessary bold step of reaching into deep learning (16) is equally essential for such are an essential physical foundation to com- the vastly unexplored regime of high task cognitive behavior to be efficiently embedded putational intelligence. What better sub- complexity, not only may it become feasible to in compact neural code. Indeed, efficient strate for emulation does silicon offer than unravel the mysteries of the cognitive brain, learning across multiple scales is key to di- the innate Boltzmann statistics of electrons but do so with great impact at full scale, and mensionality reduction in minimizing ma- and holes moving across channels of field with efficiency unreachable by conventional chine complexity for a given task complexity effect transistors, akin to sodium, potassium, computing approaches. What does it take to (Fig. 1B). By virtue of Occam’s razor, such and other ions moving through ion channels get us there? For one, it is unlikely that a reduced representations also tend to offer across the cell membrane? As Carver Mead neuromorphic avatar can accomplish tasks superior generalization, further benefiting told us all along: “Listen to the technology of complexity at the human brain level cognitive performance. and find out what it’s telling you” (5). Be- without expending at least as many resources: Our brain offers an existence proof that fore long, and with a sufficient dose of the equivalent of roughly 1012 neurons and assemblies of imprecise and unreliable analog providence and persistence, we may all im- 1015 synapses, which if not strictly needed circuit components are capable of producing merse in the beautiful sound of a silicon would have been pruned through evolution highly reliable and resilient, if not reproduc- symphony emerging through a concerted for greater fitness in size and metabolic use. ible, cognitive behavior. The open question is effort of scientists and engineers, all tuned Fortunately, the extreme density and energy to what extent such behavior could also to the dual goals of holistic understanding efficiency of the human brain, at a mere 0.002 emanate, perhaps more efficiently, from of what the brain is telling us, and leverag- m3 volume and 20 W power consumption, is carefully crafted functional abstractions ing this understanding to newly achievable within reach of current advances in neuromor- in more traditional computer architecture levels of health and intelligence. phic silicon nanotechnology—at 1011 synapses per square centimeter (13) and 1 fJ energy per synaptic operation (10)—for applications of 1 Neftci E, et al. (2013) Synthesizing cognition in neuromorphic 13 Kuzum D, Jeyasingh RG, Lee B, Wong HS (2012) Nanoelectronic electronic systems. Proc Natl Acad Sci USA 110(37):E3468–E3476. programmable synapses based on phase change materials for brain- embodied cognition on active mobile or 2 von Neumann J (1957). The Computer and the Brain. (Yale Univ inspired computing. Nano Lett 12(5):2179–2186. implanted platforms. Press, New Haven, CT). 14 Indiveri G, Chicca E, Douglas RJ (2009) Artificial cognitive Clearly, merely scaling up neuromorphic 3 Alivisatos AP, et al. (2013) Nanotools for neuroscience and brain systems: From VLSI networks of spiking neurons to neuromorphic activity mapping. ACS Nano 7(3):1850–1866. cognition. Cogn. Comput. 1:119–127. hardware to numbers of neurons and synapses 4 Churchland PS, Sejnowski TJ (1992) The Computational Brain (MIT 15 Rabinovich MI, Huerta R, Varona P, Afraimovich VS (2008) on par with the human brain is not sufficient Press, Cambridge, MA). Transient cognitive dynamics, metastability, and decision making. to warrant its cognitive intelligence. In their 5 Mead CA (1990) Neuromorphic electronic systems. Proc IEEE PLOS Comput Biol 4(5):e1000072. article, Neftci et al. (1) make a critical contri- 78(10):1629–1636. 16 Salakhutdinov R, Hinton G (2012) An efficient learning 6 Mead CA (1989) Analog VLSI and Neural Systems (Addison- procedure for deep Boltzmann machines. Neural Comput 24(8): bution in demonstrating that elements of con- Wesley, Reading, MA). 1967–2006. text-sensitive cognitive behavior arise in 7 Indiveri G, et al. (2011) Neuromorphic silicon neuron circuits. Front 17 Serrano-Gotarredona R, et al. (2009) CAVIAR: A 45k neuron, 5M neuromorphic systems that are modularly Neurosci 5:73. synapse, 12G connects/s AER hardware sensory-processing- learning- 8 Liu S-C, Delbruck T (2010) Neuromorphic sensory systems. Curr actuating system for high-speed visual object recognition and structured with reliable soft winner-take-all Opin Neurobiol 20(3):288–295. tracking. IEEE Trans Neural Netw 20(9):1417–1438. (WTA) dynamics. These neuromorphic “soft- 9 Schemmel J, et al. (2010) A wafer-scale neuromorphic hardware 18 Yu T, et al. (2012) 65k-neuron integrate-and-fire array transceiver state machines” are fundamentally different system for large-scale neural modeling. International Symposium on with address-event reconfigurable synaptic routing. Biomedical fi Circuits and Systems, ISCAS 2010 (Institute of Electrical and Electronic Circuits and Systems, BioCAS 2012 (Institute of Electrical and from digital nite-state machines with corre- Engineers), pp 1947–1950. Electronic Engineers), pp 21–24. sponding linear scaling in task complexity (Fig. 10 Karakiewicz R, et al. (2012) 1.1 TMACS/mW fine-grained stochastic 19 Jin X, et al. (2010) Modeling spiking neural networks on 1B). These soft state machines represent the resonant charge-recycling array processor. IEEE Sens J 12(4):785–792. SpiNNaker. Comput Sci Eng 12(5):91–97. equivalent of cognitive “habits” that are effort- 11 Bellman RE (1957) Dynamic Programming (Princeton Univ Press, 20 Merolla P, et al. (2011) A digital neurosynaptic core using Princeton, NJ). embedded crossbar memory with 45pJ per spike in 45nm. Custom lessly recalled through learned soft transitions 12 Dayan P (2008) Simple substrates for complex cognition. Front Integrated Circuits Conference, CICC 2011 (Institute of Electrical and between distributed states of working memory, Neurosci 2(2):255–263. Electronic Engineers) pp 1–4.

Cauwenberghs PNAS | September 24, 2013 | vol. 110 | no. 39 | 15513 Downloaded by guest on September 28, 2021