Solve Memory to Solve Cognition Paul Baxter Lincoln Centre for Autonomous Systems School of Computer Science University of Lincoln, U.K. Email: pbaxter@lincoln.ac.uk Abstract—The foundations of cognition and cognitive be- consistent with more purely theoretical considerations, e.g. haviour are consistently proposed to be built upon the capability [11], that emphasise the dynamical process properties of to predict (at various levels of abstraction). For autonomous memory over passive information storage. cognitive agents, this implicitly assumes a foundational role for memory, as a mechanism by which prior experience can be By taking on this interpretation of memory, a more re- brought to bear in the service of present and future behaviour. fined process definition memory may be ventured: memory In this contribution, this idea is extended to propose that an is a distributed associative structure that is created through active process of memory provides the substrate for cognitive experience (the formation associations), and which forms the processing, particularly when considering it as fundamentally substrate for activation dynamics (through externally driven associative and from a developmental perspective. It is in this context that the claim is made that in order to solve the question activity, and internal reactivation) that gives rise to cognitive of cognition, the role and function of memory must be fully processing [12], [5]. The creation of structure through expe- resolved. rience is consistent with developmental accounts, and enforce the consideration of not only interaction with the environment, I. P REDICTION , C OGNITION , AND M EMORY but also the social context of the learning agent (if human- There are a range of competencies that are involved in cog- like cognition is to be considered). Previous explorations have nition: an ongoing challenge is to identify common functional suggested how this framework can be used (in principle) and organisational principles of operation. This will facilitate to account for human-level cognitive competencies within a both the understanding of natural cognition (particularly that memory-centred cognitive architecture [13], although there of humans), and the creation of synthetic artefacts that can remain many gaps in this account that require addressing be of use to individuals and society. One such principle is before it can be considered definitive. that of prediction [1], prospection [2], or indeed simulation [3], as being fundamental to cognition. A further requirement II. A PPLICATION AND I MPLICATIONS is the need to incorporate an account of development [4] as a Following this definition, take for example the role that means of an individual to gain cognitive competencies through such a memory-centred cognitive architecture could play in experience (of the physical and social world), rather than a facilitating social robot behaviour, as a prototypical example priori programming. of a cognitive competence that needs to be fulfilled. It is It is suggested that one common dependency of these princi- uncontroversial to suggest that humans incrementally acquire ples is a requirement for memory. At this point, the definition social skills (though perhaps based on some inherently present of memory provided is only in the broadest sense: i.e. memory mechanisms) over time and through development. The role is a process that acquires information through experience in of memory within this is therefore also not controversial, the service of current and future behaviour [5]. While broad, particularly when skills such as intent prediction (based on it nevertheless commits to a fundamental function/role for prior experience) are also considered [14]. Using an associative memory in behaviour [6]. It is on this basis that the remainder network that learns from the behaviour of the interaction of this contribution is focused: taking memory as fundamental, partner [15], following the use of simple associative learning how can it be characterised such that it serves cognition (and in [16], it has been found that a degree of behavioural the development thereof)? alignment between a child and a robot is observed within In one particular perspective grounded in neuropsycholog- real-time interactions - an effect readily seen in human-human ical data, emphasis is placed on the associative and network interactions. While only a basic illustration of human-like nature of memory. This is apparent in the “Network Memory” competence, this nevertheless demonstrates the importance of framework for example [7], which proposes a hierarchical memory for social HRI [17], and thus establishes associativity and heterarchical organisation of overlapping distributed as- as a candidate foundational mechanism for a social cognitive sociative networks that that formed through experience, and architecture. Similarly, with associativity being considered whose reactivation gives rise to the dynamics that instantiate sufficient for generating predictions as noted above, and pre- cognition [8]. Such a perspective is not unusual, e.g. [1], diction/anticipation being considered essential for sociality in despite the apparent contradiction to multi-system accounts terms of supporting coordination [18], then such an account of memory organisation, e.g. [9], [10], with it being also of memory remains consistent. Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 58 An alternative implementation using similar principles of contribution goes beyond this: that a full account of memory associativity and interactive learning has been applied to a may be sufficient to provide an account of cognition. range of embodied and developmental psychology models R EFERENCES related to language. The Epigenetic Robotics Architecture (ERA) [19] emphasises associative learning, and is instantiated [1] M. Bar, “The proactive brain: using analogies and associations to generate predictions,” Trends in Cognitive Sciences, vol. 11, no. 7, through linked self-organising maps (SOM), arranged through pp. 280–289, 2007. a “hub” SOM that learns from body posture. This structure, [2] D. Vernon, M. Beetz, and G. Sandini, “Prospection in Cognition: The learning from a blank initial state, can provide an account Case for Joint Episodic-Procedural Memory in Cognitive Robotics,” Frontiers in Robotics and AI, vol. 2, no. July, pp. 1–14, 2015. of how aspects of language can extend cognitive processing [3] G. Hesslow, “Conscious thought as simulation of behaviour and percep- [20], and of how word learning in infants is mediated by body tion,” Trends in cognitive sciences, vol. 6, no. 6, pp. 242–247, 2002. posture [21]. In each of these examples, the computational in- [4] J. Weng, J. McClelland, A. Pentland, O. Sporns, I. Stockman, M. Sur, and E. Thelen, “Autonomous Mental Development by Robots and stantiation of ERA is the same, but the functionality observed Animals,” Science, vol. 291, pp. 599–600, 2001. differs based on the interaction context of the experiment. [5] R. Wood, P. Baxter, and T. Belpaeme, “A Review of long-term memory Given the fundamentally associative nature of the learning in natural and synthetic systems,” Adaptive Behavior, vol. 20, no. 2, pp. 81–103, 2012. process, this is consistent with the memory-centred account of [6] A. M. Glenberg, “What Memory is For,” The Behavioral and Brain social human-robot interaction competence described above. Sciences, vol. 20, no. 1, pp. 1–19; discussion 19–55, 1997. In many principled but low-level approaches – including the [7] J. M. Fuster, “Network Memory,” Trends in Neurosciences, vol. 20, no. 10, pp. 451–9, 1997. those systems based on the developmental systems paradigm, [8] J. M. Fuster and S. L. Bressler, “Past Makes Future: Role of pFC in as subscribed to here – there is often a gap between the theo- Prediction,” Journal of Cognitive Neuroscience, vol. 27, no. 4, pp. 639– retical consistency and the complexity of the applied resulting 654, 2015. [9] L. R. Squire, “Memory systems of the brain: a brief history and current system, with simplified (or rather constrained) problems typi- perspective.,” Neurobiology of learning and memory, vol. 82, no. 3, cally targeted. While the memory-centred approach advocated pp. 171–7, 2004. here suffers similarly, the range of applications outlined in [10] G. Repovs and A. Baddeley, “The multi-component model of working memory: explorations in experimental cognitive psychology,” Neuro- the previous paragraphs cover a number of aspects of “higher science, vol. 139, no. 1, pp. 5–21, 2006. level” (indeed, human-level) cognition that go beyond the [11] A. Riegler, “Constructive memory,” Kybernetes, vol. 34, no. 1, pp. 89– typical domains for low-level associative systems. The efforts 104, 2005. [12] P. Baxter and W. Browne, “Memory as the substrate of cognition: a described here remain relatively sparse and currently lack a developmental cognitive robotics perspective,” in Proceedings of the computational integration into a single coherent system that Tenth International Conference on Epigenetic Robotics (B. Johansson, existing psychologically-derived cognitive architectures (such E. Sahin, and C. Balkenius, eds.), (Örenäs Slott, Sweden), pp. 19–26, 2010. as SOAR, ACT-R, etc) attempt. Nevertheless, there appears to [13] P. Baxter, R. Wood, A. Morse, and T. Belpaeme, “Memory-Centred be a convergence of principles of operation that the present Architectures: Perspectives on Human-level Cognitive Competencies,” in work seeks to extend: cognition founded on formation and Proceedings of the AAAI Fall 2011 symposium on Advances in Cognitive Systems (P. Langley, ed.), (Arlington, Virginia, U.S.A.), pp. 26–33, manipulation of memory, and memory as associative and AAAI Press, 2011. developmental. At the least, what is proposed here is a re- [14] Y. Demiris, “Prediction of intent in robotics and multi-agent systems,” framing of the problem: not to look at cognition from the Cognitive Processing, vol. 8, no. 3, pp. 151–8, 2007. [15] P. E. Baxter, J. de Greeff, and T. Belpaeme, “Cognitive architecture for perspective of the ‘computation’ or the behavioural outcome humanrobot interaction: Towards behavioural alignment,” Biologically as is typical, but rather to re-evaluate the problem from the Inspired Cognitive Architectures, vol. 6, pp. 30–39, 2013. perspective of memory. [16] K. Dautenhahn and A. Billard, “Studying robot social cognition within a developmental psychology framework,” in Third European Workshop on Advanced Mobile Robots (Eurobot’99), (Zurich, Switzerland), pp. 187– III. T HE S UFFICIENCY OF AN ACCOUNT OF M EMORY 194, 1999. The outcome of this discussion is a commitment to a [17] P. Baxter, “Memory-Centred Cognitive Architectures for Robots Inter- acting Socially with Humans,” in 2nd Workshop on Cognitive Archi- fundamentally associative structure of memory, with this main- tectures for Social Human-Robot Interaction at HRI’16, (Christchurch, taining consistency with the developmental perspective, and New Zealand), 2016. as illustrated through the social human-robot interaction and [18] E. Di Paolo and H. De Jaegher, “The interactive brain hypothesis,” Frontiers in Human Neuroscience, vol. 6, pp. 1–16, 2012. language examples. The outline described in this abstract [19] A. F. Morse, J. De Greeff, T. Belpaeme, and A. Cangelosi, “Epige- points to a framework within which the relationship between netic Robotics Architecture (ERA),” IEEE Transactions on Autonomous memory and cognition can be understood, although there Mental Development, vol. 2, no. 4, pp. 325–339, 2010. [20] A. F. Morse, P. Baxter, T. Belpaeme, L. B. Smith, and A. Cangelosi, remain a number of open questions that need to be resolved, “The Power of Words,” in Joint IEEE International Conference on such as reconciliation with empirical evidence supporting the Development and Learning and on Epigenetic Robotics, (Frankfurt am multi-systems organisation of memory, e.g. [10], and the in- Main, Germany), pp. 1–6, IEEE Press, 2011. [21] A. F. Morse, V. L. Benitez, T. Belpaeme, A. Cangelosi, and L. B. Smith, terplay of memory with non-memory mechanisms underlying “Posture affects how robots and infants map words to objects,” PLoS cognition (such as affective processes, e.g. [22]). Nevertheless, ONE, vol. 10, no. 3, 2015. the proposal is that even these aspects could be approached [22] A. R. Damasio, “The somatic marker hypothesis and the possible functions of the prefrontal cortex,” Philosophical Transactions Of the from the perspective of memory. In all, this leads to the view Royal Society B, vol. 351, pp. 1413–1420, 1996. that in order to ‘solve’ cognition, the problem of memory must be fully resolved. Indeed, the suggestion of the present Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 59