Brain-inspired Distributed Cognitive Architecture
BBrain-inspired Distributed Cognitive Architecture
Leendert A Remmelzwaal , Amit K Mishra , George F R Ellis Department of Electrical Engineering, University of Cape Town, Rondebosch, Cape Town, South Africa7700. Department of Mathematics and Applied Mathematics, University of Cape Town, Rondebosch, Cape Town,South Africa 7700.* Corresponding author: [email protected]
Abstract
In this paper we present a brain-inspired cognitive architecture that incorporates sensory processing,classification, contextual prediction, and emotional tagging. The cognitive architecture is implemented asthree modular web-servers, meaning that it can be deployed centrally or across a network for servers. Theexperiments reveal two distinct operations of behaviour, namely high- and low-salience modes of operations,which closely model attention in the brain. In addition to modelling the cortex, we have demonstrated thata bio-inspired architecture introduced processing efficiencies. The software has been published as an opensource platform, and can be easily extended by future research teams. This research lays the foundations forbio-realistic attention direction and sensory selection, and we believe that it is a key step towards achievinga bio-realistic artificial intelligent system.Keywords: distributed cognitive architecture, affect, cortex, prediction, corticothalamic connections
In this paper we attempt to answer the question: Can we design a cognitive architecture based on the modularstructure and functionality of the human brain? Cortical modules such as the corticothalamic circuitry [1], theascending arousal systems [2] and vision [3] are the result of of millions of years of Darwinian natural selectionand survival-driven optimisation [4] [5]; in this research we allow these structures to guide the design of ourcognitive architecture. We present a proof-of-concept of a distributed, computationally efficient, brain-inspiredcognitive architecture that has the ability to tag events with salience and direct cortical attention to the mostsalient current event. In addition, this framework is open-source and can be easily extended by future researchteams.
Why did we embark on this project? It is common knowledge that bio-realistic artificial intelligence is still adistant dream to many researchers. We embarked on this work because we believe that a modular bio-inspiredcognitive architecture could be a key step in designing a bio-realistic artificially intelligent system. This is aradically different approach to taking existing software (e.g. OpenCV) and adapting it to be more brain-like -it is fundamentally brain-inspired. It be one of the key steps towards achieving artificial intelligence.A second motivation was to build a model to simulate some of the behaviours of key modules in the humanbrain, including the thalamus, cortex, arousal system, basal ganglia, and amygdala, to help medical practitionersunderstand the effects of disease or accidental damage to the cortex. Wouldn’t it be great if we could simulatethe effect of removing the thalamus, impeding the release of noradrenaline, or triggering a seizure withoutrelying on living test subjects? A more extended model of the kind we present could help here.A third motivation was to design a cognitive architecture with future research in mind. The distributedmodular nature of the architecture we present is designed to be extended, whereby a research group canindependently enhance a single module in the system, such as the arousal system, without affecting the workingof the entire system. This would allow multiple research teams to incrementally enhance the entire system byfocusing on only a part of the system.A fourth motivation was to design a scaleable infrastructure that could be deployed on a distributed networkof servers, allowing the system to exist outside of just a single server. The implementation leverages a distributednetwork of servers, giving autonomy to computational units inspired by key units in the human brain. Whilethis allows the system to scale, it also allows the system as a whole to build in redundancies that will improveefficiency and robustness of the system as a whole. 1 a r X i v : . [ q - b i o . N C ] M a y inally, we hope that building on the research presented in this paper will eventually shed light on the natureof consciousness [6] [7]. Our bio-inspired cognitive architecture is modelled closely after a number of key structures in the human brain,namely the thalamus, cortex, arousal system, basal ganglia, and amygdala (see Fig 1). These structures playimportant roles in sensory processing, prediction, emotional tagging, attention direction and consciousness [8][9]. We model some aspects of these interactions.Figure 1: Some of the key autonomic and voluntary pathways between the thalamus, amygdala, arousal system,basal ganglia and the prefrontal cortex.
Emotional tagging of events and memories is key component to consciousness, intelligence and directing attentionin the cortex [10] [11] [12] [9]. The arousal system releases neuromodulators such as dopamine and noradrenalineinto the cortex via diffuse projections [13] [14] [11]. These long-range diffuse projections allow neuromodulatorsto impact an entire pattern of synaptic connections proportional to their synaptic activation level and thestrength of the neuromodulator released. The release of noradrenaline has a direct impact of increasing gain onperception, which in turn also directs cortical attention to the object or event causing the emotional response.For example, if someone has a fear of dogs, then seeing a dog may make them feel afraid, which in turn commandstheir attention to the dog. In this paper we model the way (a) the arousal system releases noradrenaline, and(b) how the noradrenaline level decays over time.
The thalamus acts as a relay station between visual, auditory and somatosensory sensory inputs and theprefrontal cortex [1] [15] [16] [17]. The connections between the thalamus and the prefrontal cortex are referredto as corticothalamic connections [18] and consist of both bottom-up (feed forward) and top-down (feedback)connections [4]. The thalamus is responsible for comparing the incoming signal to the top-down predictionsignal, and sending a moderated signal back to the cortex if the difference between the predicted signal andincoming data exceeds a threshold. This threshold can be affected by the level of neuromodulators releasedfrom the arousal system, which is a key feature of our model. In this paper we focus on modelling the way thethalamus (a) relays incoming information to the cortex, (b) receives a contextual prediction from the cortex,and (c) compares the incoming information to the prediction before deciding whether to send the next imageto the cortex. 2 .3 Prefrontal Cortex
These prefrontal cortex is responsible for a range of key functions such as cognition [19], attention direction [20][21] [22] [23], sensory selection [24] [25] [26], awareness [27], emotional control [28], prediction [5], and saliencedetection [29]. In this paper we focus on modelling the way the cortex (a) receives sensory inputs, (b) classifiesthe sensory input, (c) creates a prediction of the current context, (d) produces a desire to act, and (e) triggersthe release of neuromodulators. The model is a simplified integrated representation of these features.
In summary: we present a proof-of-concept of a brain-inspired cognitive architecture that incorporates sensoryprocessing, classification, contextual prediction, and emotional tagging. The architecture is robust and modular;meaning that it can be deployed centrally or across a network for servers. The software produced by this researchhas been published as an open source platform, and can be easily extended by future research teams.
Our work builds on existing research by connecting together models of individual modules in the brain, namelymodels of the thalamus, cortex, arousal system.In modelling the thalamus, cortex and corticothalamic connections, we reviewed the work done on predictiveprocessing by Andy Clark [30], as well as the models of Retinal Predictive Coding [31], Linear Predictive Coding[32], Cortical Predictive Coding [33], Restricted Boltzmann machine (RBM) [34], Free Energy Predictive Coding[35], BC-DIM Predictive Coding [36], Predictive Sparse Decomposition [37], Stacked Denoising Auto-encoders[38], Deep Predictive Coding Networks [39], PredNet [40], Multilevel Predictor Estimator [41], Deep PredictiveCoding [42], LPCNet [43], and the CTNN model [44]. We have chosen to incorporate the CTNN model becauseunlike the other models, it is input agnostic, multi-modal and computationally efficient.In modelling the arousal system, we reviewed the work done on modelling non-local effects of neuromod-ulators by Edelman and the primary emotional systems identified by Panksepp [11]. These models includethe Husbands’ model of gas diffusion [45], Juvina’s model of valuation and arousal[46], Edelman’s brain-baseddevices [2] [47], and the Salience-affected Neural Network (SANN) [48]. We have chosen to use the SANNarchitecture to model the effects of salience because unlike the other models salience affects a specific neuralactivation pattern as a whole, tagging it with a salience signal, and with one-time learning.Attempts have previously been made to map the connectome, such as the Blue Brain Project [49], butwe are focused on the functional structures of each module in the brain, rather than the connections betweenindividual neurons. Our research is therefore at a level abstracted from mapping the connectome.
The distributed computational model we use is represented in Fig 2, developed for computational purposes fromthe set of interactions summarised in Fig 1. The thalamus, arousal system and memory structures were eachmodelled as their own independent Python Flask web-servers. Each web-server could be called by any otherweb-server, to request information (e.g. current state of the cortex) or to request an action (e.g. classifying anew image). For more technical information on the API protocols see [50].
The arousal system was modelled as the first stand-alone server. It maintains as an internal state the levelsof neuromodulators such as noradrenaline in the cortex. Any other system (e.g. the cortex) may trigger therelease of a neuromodulator, and may request the current level of neuromodulator.Once a neoromodulator has been released, it is programmed to decay at a fixed rate as represented in(1).Here x is the number of seconds that has passed. y = e − . ∗ x (1) The thalamus system was modelled as the second stand-alone web-server. It calculates an internal thresholddepending on the level of noradrenaline released by the arousal system. If the difference between the incomingimage and the contextual prediction generated by the cortex system exceeds the threshold, the thalamus sendsthe input image to the cortex system for processing. A key point here is that the thalamus threshold is directly3igure 2: Overview of the distributed system architecture.impacted by the level of noradrenaline release by the arousal system. The higher the level of noradrenaline, thelower the internal threshold in the thalamus system. This models the work of Solms and Friston on the relationbetween arousal and conscousness.
The cortex system was modelled as the third stand-alone web-server, and incorporates the SANN [48] andCTNN models previously published by Remmelzwaal et al [44]. The cortex system accepts an input image, andprocesses this input image using a number of internal models, including an auto-encoder (based on the CTNN)to reduce image dimensionality and to generate a reconstruction of the current context, and deep neural network(based on the SANN) that classifies and returns the salience tag for each image (see Fig 2).Figure 3: Illustration of information flow in the cortex system.4 .4 Graphical User Interface (GUI)
A graphical user interface (GUI) shown in Fig 4 was created to allow the researcher to vary the input image,and to give the researcher visibility of key variables in the system (e.g. thalamus threshold, cortical prediction,neuromodulator level). Figure 4: The Graphical User Interface (GUI)
In this paper we generated a dataset of black and white silhouettes of animals in three classes, namely Bird,Cat and Dog (shown in Fig 1). This dataset was inspired by [51], and is a novel dataset presented for the firsttime here.
In this model only the memory structures in the Cortical System are trainable. Via the GUI or API interfacethe researcher can initiate training of both the auto-encoder (used to predict context) and the SANN (to classifyand determine salience). For the purposes of demonstration, a single image of a DOG was tagged with one-timesalience training.
The experiments conducted demonstrate various key results.
Firstly, the experiments reveal two distinct operations of behaviour, which closely model attention in the brain.The first mode of operation occurs when the system is presented with an input image that has a low-salienceassociation (see Fig 5A). When initially presented with such an image, the thalamus notices a significant changein input, and engages the cortex to classify the image. The cortex is engaged for one computational iteration,and returns a contextual prediction to the thalamus. For subsequent viewings with the same image the thalamusdoes not engage the cortex. During this mode, the cortex directs attention momentarily to identify an object,but then the cortex is free to assign attention to another classification or prediction task.The second mode of operation occurs when the system experiences an input image with a high associatedsalience (see Fig 5B). Similar to the first mode of operation, the thalamus notices a significant change ininput, and engages the cortex to classify the image. Again, the cortex is engaged for one computationaliteration, and returns a contextual prediction to the thalamus. However this time, the cortex identifies a highsalience association with the high level concept, and two things happen: firstly, the cortex triggers a release ofnoradrenaline from the arousal system, and secondly, the release of noradrenaline lowers the internal threshold5igure 5: Two modes of operation. Low-salience (A) and High-salience (B) inputsof the thalamus System. This has a direct impact on how the system processes subsequent images. Witha lower internal threshold, the thalamus System sends images to the cortex more frequently, demanding thecomputational power and attention of the cortex system. In this mode, the cortex directs all of its attention tothe incoming images, ignoring other lower-priority tasks.
Secondly, the simulations conducted in this paper confirm neuroscience understanding: that the thalamusand arousal system indeed play a crucial roles in processing incoming data, and directing cortical attention.Without the Thalamus constantly comparing incoming signals to a contextual prediction, our brains wouldbe overwhelmed with a non-stop overload of incoming information to process. Without salience tagging ofmemories and event enabled by the Arousal system, the cortex wouldn’t receive the guidance it requires todirect its attention to the most salient issue
In addition to modelling the thalamus and arousal system, we have also demonstrated that a bio-inspiredarchitecture can optimize performance of complex processing tasks (e.g. classification and prediction). We haveshow that our architecture frees up computational processing during low-salience events, especially when thereis high levels of similarity between one image an another. We experience this in the physical world almost everyday when we drive - driving at a steady speed down a road is a highly monotonous task, whereby your brainis free to engage in conversations or to sing along to a song on the radio. However, if a person steps onto theroad, your attention is immediately directed to managing the salience situation, abandoning almost instantlythe conversation you were having, or song you were singing. We have demonstrated this effect by adding 2sub-systems inspired by the functions of the the thalamus and the arousal system in the brain.
In this paper we have demonstrated with the aid of a fully-connected simulation of the thalamus, arousal systemand cortex, how these three key structures rely on each other to allow for:1. Attention direction [20] [21] [22] [23]; in our model the cortex is engaged with only those images thatdiffer significantly from the predicted context, a level which is modulated by the affective response fromprevious images. This has the impact of freeing up the cortex for other processing during low-salienceevents (e.g. driving a car), and demands attention from the cortex during high-salience events (e.g. seeinga dog). What is important to observe is that in our model the cortex is not actively directing its ownattention; it is a emergent feature guided by the arousal system and the thalamus.6. Salience detection [29]; our system allows for events to be tagged with affect (salience), and for thatsalience to be returned along with the label if that combination of sensory inputs is experiences again inthe future.3. A simple form of prediction [5]; the cortex generates a prediction of the current context by leveraging anauto-encoder. The model sends this prediction of the current context back to the thalamus system to becompared with incoming sensory information.
We have designed a cognitive architecture specifically with future research in mind. The distributed modularnature of the architecture we present here allows research groups to independently enhance and extend a singlemodule in the system (e.g. the arousal system) without affecting the functioning of the entire system. The innerworking of each module is unknown to other modules, and each module can only interact with other modulesusing predefined API calls. For example, the thalamus can send the cortex a new sensory image, and can requesta prediction, but it is not concerned with how the cortex processes the information it is sent, or generates theprediction. We believe that this approach to designing the cognitive architecture is key in allowing this researchto be extended in the future.
10 Future Work
In this paper we have presented an open-source brain-inspired distributed cognitive architecture. The GraphicalUser Interface (GUI) allows researchers to examine and interact with the model. We hope that future cognitivearchitectures can benefit from the principles presented in this paper.This model can be extended to include recurrent connections, or to include prediction over a period of time(e.g. to simulate autonomous behaviour when driving a car). While the current model accepts only images, itcan be extended to multi-modal inputs such as auditory of somatosensory inputs, as is demonstrated in [44]. Itmay also be valuable to explore whether the system proposed in this paper experiences subjective states, usingan evaluation framework such as the Independent Core Observer Model (ICOM) [53].We believe that this system could be extended to allow for:1. Emotional control [28]; how does an emotional response translate to an action? that can guide behaviourof the system. This would extend the ”desire to act” feature of our current system.2. Sensory selection [24] [25] [26]; how can the cortex select to attention from a specific sensory input? Thiswould require extending the thalamus model to allow sending of multiple modes of sensory data, and forthe cortex to be able to process multiple modes of sensory data independently of each other.The ultimate aim of this paper has been to introduce a cognitive architecture that may help lead to bettermodelling and understanding of awareness [27], consciousness [6], and subjective experiences [52]. We believethe design principles introduced here can be important in that search.
11 Supporting Material
The source code as well as records of the tests conducted in this paper are publicly available online [50]. Foradditional information, please contact the corresponding author.
12 Acknowledgements
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profitsectors.
References [1] Henry, A.J. and Usrey, M.W. Corticothalamic feedback and sensory processing.
Current opinion inneurobiology , 13(4):440–445, 2003.[2] Gerald M Edelman. Learning in and from brain-based devices.
Science , 318(5853):1103–1105, 2007.73] Wilson, JC and Hodgson, RM. Log-polar mapping applied to pattern representation and recognition.
Computer Vision and Image Processing , 245–277, 1992.[4] Ellis, GFR Top-down effects in the brain.
Physics of life reviews , 2018[5] George Ellis and Mark Solms.
Beyond Evolutionary Psychology . Cambridge University Press, 2017.[6] Solms, M. and Friston, K. How and why consciousness arises: Some considerations from physics andphysiology. Journal of Consciousness Studies, 25(5-6), pp.202-238. 2018[7] Solms, M. The hard problem of consciousness and the free energy principle. Frontiers in psychology, 9,p.2714. 2019[8] George FR Ellis and Judith A Toronchuk. Neural development: Affective and immune system influences.
Consciousness and Emotion: Agency, conscious choice, and selective perception , 1:81, 2005.[9] Mark Solms and Karl Friston. How and why consciousness arises: some considerations from physics andphysiology.
J. Conscious. Stud . 25: 202-238, 2019.[10] Damasio, A.
Descartes’ error: Emotion, reason, and the human brain . Harcourt Brace, 1994[11] Jaak Panksepp.
Affective neuroscience: The foundations of human and animal emotions . Oxford univer-sity press, 2004.[12] Panksepp, J. and Biven, L.
The archaeology of mind: neuroevolutionary origins of human emotions
WWNorton and Company, 2012.[13] Gerald M Edelman.
Neural Darwinism: The theory of neuronal group selection.
Basic books, 1987.[14] Gerald M Edelman.
Wider than the sky: The phenomenal gift of consciousness . Yale University Press,2004.[15] Lee M Miller, Monty A Escab´ı, Heather L Read, and Christoph E Schreiner. Spectrotemporal receptivefields in the lemniscal auditory thalamus and cortex.
Journal of neurophysiology , 87(1):516–527, 2002.[16] Adam M Sillito and Helen E Jones. Corticothalamic interactions in the transfer of visual information.
Philosophical Transactions of the Royal Society of London B: Biological Sciences , 357(1428):1739–1752,2002.[17] Farran Briggs and W Martin Usrey. Emerging views of corticothalamic function.
Current opinion inneurobiology , 18(4):403–407, 2008.[18] Sherman. MS. and Guillery, RW.
Exploring the thalamus and its role in cortical function. Cambridge,MA , 2009.[19] Edward G Jones. Thalamic circuitry and thalamocortical synchrony.
Philosophical Transactions of theRoyal Society of London B: Biological Sciences , 357(1428):1659–1673, 2002.[20] Stewart Shipp. The brain circuitry of attention.
Trends in cognitive sciences , 8(5):223–230, 2004.[21] Jos´e de Bourbon-Teles, Paul Bentley, Saori Koshino, Kushal Shah, Agneish Dutta, Paresh Malhotra,Tobias Egner, Masud Husain, and David Soto. Thalamic control of human attention driven by memoryand learning.
Current biology , 24(9):993–999, 2014.[22] Ralf D Wimmer, L Ian Schmitt, Thomas J Davidson, Miho Nakajima, Karl Deisseroth, and Michael MHalassa. Thalamic control of sensory selection in divided attention.
Nature , 526(7575):705, 2015.[23] Karl Friston. Does predictive coding have a future?
Nature neuroscience , 21(8):1019, 2018.[24] Yuri B Saalmann and Sabine Kastner. Neural mechanisms of spatial attention in the visual thalamus.
The Oxford Handbook of Attention , page 399, 2014.[25] Jos´e de Bourbon-Teles, Paul Bentley, Saori Koshino, Kushal Shah, Agneish Dutta, Paresh Malhotra,Tobias Egner, Masud Husain, and David Soto. Thalamic control of human attention driven by memoryand learning.
Current biology , 24(9):993–999, 2014.[26] Basilis Zikopoulos and Helen Barbas. Circuits for multisensory integration and attentional modulationthrough the prefrontal cortex and the thalamic reticular nucleus in primates.
Reviews in the neurosciences ,18(6):417–438, 2007. 827] Geraint Rees. Visual attention: the thalamus at the centre?
Current biology , 19(5):R213–R214, 2009.[28] Lihua Sun, Jari Per¨akyl¨a, Markus Polvivaara, Juha ¨Ohman, Jukka Peltola, Kai Lehtim¨aki, Heini Huhtala,and Kaisa M Hartikainen. Human anterior thalamic nuclei are involved in emotion–attention interaction.
Neuropsychologia , 78:88–94, 2015.[29] Howard Bowman, Marco Filetti, Brad Wyble, and Christian Olivers. Attention is more than predictionprecision.
Behavioral and Brain Sciences , 36(3):206–208, 2013.[30] Andy Clark.
Embodied prediction . Open MIND. Frankfurt am Main: MIND Group, 2015.[31] Mandyam Veerambudi Srinivasan, Simon Barry Laughlin, and Andreas Dubs. Predictive coding: a freshview of inhibition in the retina.
Proceedings of the Royal Society of London. Series B. Biological Sciences ,216(1205):427–459, 1982.[32] Douglas O’Shaughnessy. Linear predictive coding.
IEEE potentials , 7(1):29–32, 1988.[33] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretationof some extra-classical receptive-field effects.
Nature neuroscience , 2(1):79, 1999.[34] Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets.
Neural computation , 18(7):1527–1554, 2006.[35] Karl Friston and Stefan Kiebel. Predictive coding under the free-energy principle.
Philosophical Trans-actions of the Royal Society B: Biological Sciences , 364(1521):1211–1221, 2009.[36] Michael W Spratling. Predictive coding as a model of cognition.
Cognitive processing , 17(3):279–305,2016.[37] Koray Kavukcuoglu, Marc’Aurelio Ranzato, and Yann LeCun. Fast inference in sparse coding algorithmswith applications to object recognition. arXiv preprint arXiv:1010.3467 , 2010.[38] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stackeddenoising auto-encoders: Learning useful representations in a deep network with a local denoising criterion.
Journal of machine learning research , 11(Dec):3371–3408, 2010.[39] Rakesh Chalasani and Jose C Principe. Deep predictive coding networks. arXiv preprint arXiv:1301.3541 ,2013.[40] William Lotter, Gabriel Kreiman, and David Cox. Deep predictive coding networks for video predictionand unsupervised learning. arXiv preprint arXiv:1605.08104 , 2016.[41] Hyun Kim, Jong-Hyeok Lee, and Seung-Hoon Na. Predictor-estimator using multilevel task learning withstack propagation for neural quality estimation. In
Proceedings of the Second Conference on MachineTranslation , pages 562–568, 2017.[42] Haiguang Wen, Kuan Han, Junxing Shi, Yizhen Zhang, Eugenio Culurciello, and Zhongming Liu. Deeppredictive coding network for object recognition. arXiv preprint arXiv:1802.04762 , 2018.[43] Jean-Marc Valin and Jan Skoglund. Lpcnet: Improving neural speech synthesis through linear predic-tion. In
ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing(ICASSP) , pages 5891–5895. IEEE, 2019.[44] Remmelzwaal, L.A., Mishra, A. and Ellis, G.F. CTNN: Corticothalamic-inspired neural network. arXivpreprint arXiv:1910.12492. 2019.[45] Phil Husbands. Evolving robot behaviours with diffusing gas networks. In
European Workshop on Evo-lutionary Robotics , pages 71–86. Springer, 1998.[46] Ion Juvina, Othalia Larue, and Alexander Hough. Modeling valuation and core affect in a cognitivearchitecture: The impact of valence and arousal on memory and decision-making.
Cognitive SystemsResearch , 48:4–24, 2018.[47] Jeffrey L Krichmar and Gerald M Edelman. Brain-based devices for the study of nervous systems andthe development of intelligent machines.
Artificial Life , 11(1-2):63–77, 2005.[48] Remmelzwaal, L.A., Ellis, G.F. and Tapson, J. One-time learning and reverse salience signal with asalience affected neural network (SANN). arXiv preprint arXiv:1908.03532. 2019.949] Markram, H. The blue brain project. Nature Reviews Neuroscience, 7(2), pp.153-160. 2006[50] Remmelzwaal, L.A. A distributed brain-inspired cognitive architecture. https://bitbucket.org/leenremm/affect_in_the_cortex [Online; accessed 8-May-2020].[51] Alessio, C. Animals-10: Animal pictures of 10 different categories