We aim at developing a new type of neuro-inspired artificial neural network (NN), bio-inspired, for continual learning, based on the principle of Information Maximization proposed by Barlow. Information Maximization will serve to replicate the brain’s capabilities for large memory capacity, rapid acquisition, robust memory retrieval.
Recently, we successfully developed a NN that satisfies Information Maximization by exploiting random and unprecise neurons in order to encode information. The idea behind is to exploit randomness and quantization to generate orthogonal representations that do not overlap (sparse coding).
This PhD thesis aims at pursuing the recent works done to design new neural models, to explore new features and capabilities for continual learning. Some parallels and comparison will be made with current Machine Learning techniques and brain-inspired architectures . For instance, how the hippocampus acquire and store robustly information.

We aim at developing a new type of bio/neuro-inspired artificial neural network (ANN) for continual learning, based on the principle of Information Maximization proposed by Barlow (1969) to describe efficient coding in neurons. Information Maximization is a strong principle to understand how information is conveyed in the brain.

The hypothesis of efficient encoding states that neurons should encode information as efficiently as possible in order to maximize neural resource utilization. From an Information Theory viewpoint, compressing information is equivalent to a maximum entropy representation, suppressing redundant parts in signals, which can be reconstructed or approximated from the representation. Thus, efficient neural codes are more compact and can serve to encode more information for the same capacity of neurons, or reduce their number for the same size of data.

Achieving Information Maximization in ANNs will serve to better understand information processing in the brain. It will also serve to replicate the brain’s capabilities for large memory capacity, rapid acquisition, robust memory retrieval, as well as promote code compactness and low energy consumption, as fewer resources are required. These properties are difficult to obtain in current ANNs, as very large models like Transformers become the norm.

Recently, Pitti, Weidmann, and Quoy (PNAS 2022) successfully developed a NN that satisfies Information Maximization by exploiting random neurons in order to encode information. The idea behind is to exploit randomness — which implies high entropy — and quantization — which suppresses redundancy — to generate orthogonal representations that do not overlap (e.g., Olshausen 2004). This new associative memory has allowed the rapid learning of sparse and distributed memories with few neurons only, and has been shown to not forget catastrophically when acquiring new information (French 1999). Furthermore, experiments showed that this type of encoding achieves memory capacities close to Shannon’s information theoretic limit; hence following Barlow hypothesis of efficient encoding by Information Maximization.

This PhD thesis aims at pursuing the above-mentioned recent works to design new neural models based on Information Maximization and to explore new features and capabilities for continual learning (Annabi et al., NN 2022). Parallels and comparisons with current Machine Learning techniques and brain-inspired architectures will be made.

For instance, links can be developed on how the hippocampus robustly acquires and stores information without catastrophic forgetting (McClelland et al 1995, French 1999), and how complementary systems (e.g., the neocortex) can exchange and store information for large-scale memory capacity scaffolding and development (Pitti et al., TCDS 2022).

Other links can be made with recent Machine Learning neural networks like the modern Hopfield Network (mHN) (Krotov and Hopfield 2020) and the Sparse Distributed Memory system (SDM) (Bricken 2021), which are associative memory systems that have attracted a lot of attention recently. One line of research is to compare our NN with current state of the art methods and show how it can potentially learn and reconstruct information faster in the challenging one-shot learning scenario.

Furthermore, we expect to obtain models with higher memory capacity in comparison with mHN and SDM, since we will embed mechanisms that allow to convey maximum information. We believe that our neural architecture will expand its performance on larger databases during continual learning with a minimal computational cost.

Methodology to reach the scientific objectives of the project:

The thesis will be organized around three main research topics, focusing on improving the neural architecture for rapid acquisition and reconstruction using insights from Coding/Information Theory, comparison with other neural architectures in Machine Learning on known databases for continual learning, design of a complementary neural architecture for ‘fast’ encoding and ‘slow’ generalization.

In the first stage, we will develop an energy-efficient, high capacity and long-term memory network unaffected by catastrophic interference from new input. In the second stage, we will focus on coding efficiency and memory information capacity with respect to current Machine Learning algorithms. In the third stage, we will concentrate on continual learning with two complementary systems, modeling hippocampus-cortical interaction.

Location : ETIS laboratory, CNRS, ENSEA, CY Cergy-Paris University
Cergy is a mid-sized town north-west of Paris, at 40 minutes by train.
ETIS is the main computer science CNRS laboratory in the West of Paris.
The Neurocybernetics team (15 researchers and 15 PhD students/PostDoc) is an international team working in the field of Bio-inspiration for intelligent systems design. It is specialized in Cognitive Robotics and Brain-inspired models with many international projects and close to the international communities.

Contacts : alexandre.pitti@ensea.fr claudio.weidmann@cyu.fr
References :
Annabi, L. Pitti, A. Quoy, M. (2022) Continual Sequence Modeling With Predictive Coding Front. Neurorobot. 16:845955
Barlow, H.B. Possible principles underlying the transformation of sensory messages. Rosenblith, W. (Ed.), Sensory Communication. MIT Press, Cambridge, MA., 1961.
Bricken, T., & Pehlevan, C. (2021). Attention approximates sparse distributed memory. arXiv preprint arXiv:2111.05498.
French, Robert M. Catastrophic forgetting in connectionist networks Trends in Cognitive Sciences 3, 4 (1999): 128 35.
Krotov, D., & Hopfield, J. J. (2016). Dense associative memory for pattern recognition. Advances in Neural Information Processing Systems, 29, 1172–1180.
Krotov, D., & Hopfield, J. (2020). Large associative memory problem in neurobiology and machine learning. arXiv preprint arXiv:2008.06996.
McClelland, James & Mcnaughton, Bruce & O’Reilly, Randall. (1995). Why There are Complementary Learning Systems in the Hippocampus and Neocortex: Insights from the Successes and Failures of Connectionist Models of Learning and Memory. Psychological review. 102. 419-57. 10.1037/0033-295X.102.3.419.
McClelland James L., McNaughton Bruce L. and Lampinen Andrew K. (2020) Integration of new information in memory: new insights from a complementary learning systems perspective Phil. Trans. R. Soc. B37520190637. 20190637
Olshausen BA, Field DJ. Sparse coding of sensory inputs. Curr Opin Neurobiol. 2004 Aug;14(4):481-7.
Pitti, A. Weidmann, C. Quoy, M. (2022) Digital computing through randomness and order in neural networks, PNAS, 119 (33) e21153351
Pitti, A. Quoy, M. Lavandier, C. Boucenna, S. and Weidmann, C. (2022) In Search of a Neural Model for Serial Order: A Brain Theory for Memory Development and Higher Level Cognition, IEEE Transactions on Cognitive and Developmental Systems, 14, 2, 279-291.

Given the centrality of emotions in human cognition and interaction, taking a biologically-inspired approach to model emotions in interactive social robots is key to develop robots that at the same time can interact naturally and in socially appropriate ways with humans, and be used as scientific models to study human affective cognition. I take a "strong" approach to emotion modeling from a bio-inspired, embodied and interactional perspective, to endow robots with affective cognition capabilities that underpin and shape their emotion perception and expression skills, and take them closer to being, in a proper sense, social, interactive, and agents. In this PhD project I investigate questions around the role of oxytocin and embodied interactions on affective cognition and social interaction dynamics. The work to be carried out will develop and test a robot model for decision making in social situations embodying these research questions.

BACKGROUND AND APPROACH
Neuroscience research over the last couple of decades has provided evidence that affective phenomena (emotions, motivations, moods) pervade intelligence at many levels (Damasio, 1994; Pessoa, 2013), being inseparable from the cognition-action loop, and artificial intelligence and robotics have echoed this view (Picard, 1997; Cañamero, 2021).
In humans and other animals, emotions are part of the bioregulatory mechanisms that contribute to the maintenance of the stability of an organism's internal environment (to its viability and homeostasis), needed to survive in changing environments, including human social environments. Emotions have a number of beneficial survival-related functions, as put forward by Darwin, as well as playing key roles in decision making and social behavior, as they provide values and motives to make decisions adapted to the physical and social environment (Damasio, 1994).
These aspects are relevant for autonomous and social robots that inhabit changing environments presenting similar kinds of challenges, particularly the natural environments of humans, where they have to make timely decisions that are adaptive for themselves and adapted to the current circumstances (e.g., appropriate to overcome a specific danger or to the preferences and behavior of a specific person). However, research in affective computing, social robotics, human-robot interaction, or social signal processing, pays particular attention to the external manifestations of emotions (emotion signals, emotional expression and displays) and their influence on humans, as well as giving technology the ability to recognize those signals in humans to be able to respond to them. To most researchers, those elements are sufficient for affective human-robot interaction. Nevertheless, current interactive robots are still far from being social agents.
Towards overcoming this problem, I take the stance that modeling affective cognition beyond affect expression and recognition, while at the same time grounding them, is important both to develop social robots better suited to interactions with humans in the long term, as proper adaptive interactive social agents, and to better understand human emotions. Concerning the latter point, affective robot models properly grounded in neuroscience and other affective sciences, can provide very useful tools and highly controllable and testable models to contribute to the study and understanding of human emotions (Cañamero, 2021), by operationalizing specific neuroscientific hypotheses and research questions and generating, testing in the real world, and quantitatively analyzing observable behavior stemming from those hypotheses and questions.

RESEARCH QUESTIONS AND PLANNED WORK
From a mechanistic and functional perspective, the project will focus on oxytocin, hypothesized to have multiple roles in social interaction, to investigate the following questions:
1. Which are suitable mechanisms underlying specific social emotions (and more generally affective phenomena) and how do they permit coordinated responses at the physiological, neuro-ethological, cognitive, and social levels? How can we model equivalent mechanisms in social robots to achieve a human-robot interaction dynamics similar to human-human interaction dynamics? In particular, the project will investigate the mechanisms and roles of the hormone oxytocin in the formation of long-term social bonds and in the development of “low-level” empathy (Preston & de Waal, 2002) between humans and robots.
From the perspective of the adaptation, the project will investigate the adaptive value of emotions in the social domain:
2. How do specific capacities (e.g., of perception such as increased attention to social cues and mechanisms (e.g., oxytocin as a mechanism underlying social bonding) characteristic of social emotional states affect (in terms of costs and benefits) the social adaptation and wellbeing of individuals? In particular, the project will investigate, both in groups of robots and in human-robot interaction, the adaptive value of social bonding promoted by (natural and simulated) oxytocin in terms of benefits such as empathizing with others and complying with group norms, and costs such as the emergence of competition with members of out-groups (de Dreu & Kret, 2016) under different social contexts.
The work to be carried out will build on previous work investigating the roles of oxytocin in the formation of affective social bonds in embodied motivationally-autonomous agents simulated in an artificial life environment. Drawing on the Social Salience hypothesis of oxytocin (Shamay-Tsoory & Abu-Akel, 2016) as an underlying mechanism, this work implemented and compared contrasting views around this hypothesis, and investigated and tested the roles of oxytocin in adaptation, survival, social allostasis, and the emergence of social groups based on affective bonds, the formation of alliances, and the dynamics of their interactions (Khan & Cañamero, 2018; Khan et al., 2018; Khan et al., 2020). The proposed PhD will investigate similar and related questions using real (instead of computer simulated) robots – groups of small mobile robots interacting among themselves and with humans – focusing specially on the characterization of oxytocin as an allostatic hormone that modulates both social and non-social behavior and decision making by maintaining stability through changing environments (Quintana & Guastella, 2020), on the interplay between basic elements of empathy such as synchrony, mimicry and simple emotional contagion (Prochazkova & Kret, 2017) and the formation of affective social bonds, and on testing in diverse and changing social environments (Olff et al, 2013).

REFERENCES
Cañamero, L. (2021). Embodied Robot Models for Interdisciplinary Emotion Research, in IEEE Transactions on Affective Computing 12(2) pp. 340-351, April-June 2021, doi: 10.1109/TAFFC.2019.2908162.
Damasio, A. (1994). Descartes’ Error: Emotion, reason, and the Human Brain. New York, NY: Avon Books.
Khan & Cañamero, L. (2018). Modelling Adaptation through Social Allostasis: Modulating the Effects of Social Touch with Oxytocin in Embodied Agents, Multimodal Technologies and Interaction, vol. 2, no. 4.
Khan, I. Lewis, M., and Cañamero, L. (2018). Adaptation and the Social Salience Hypothesis of Oxytocin: Early Experiments in a Simulated Agent Environment, in Proc. 2nd Symposium on Social Interactions in Complex Intelligent Systems (SICIS), Liverpool, UK, 2018, pp. 2–9.
Khan, Lewis, M., and Cañamero, L. (2020). Modelling the Social Buffering Hypothesis in an Artificial Life Environment”, in Proceedings of the Artificial Life Conference 2020 (ALIFE 2020), Montreal, Canada, 2020, pp. 393–401. The MIT Press.
Olff M1, Frijling JL, Kubzansky LD, Bradley B, Ellenbogen MA, Cardoso C, Bartz JA, Yee JR, van Zuiden M. (2013). The role of oxytocin in social bonding, stress regulation and mental health: an update on the moderating effects of context and interindividual differences. Psychoneuroendocrinology, 38(9):1883-94.
Pessoa, L. (2013). The Cognitive-Emotional Brain: From Interactions to Integration. Cambridge, MA: The MIT Press.
Picard, R.W. (1997). Affective Computing. Cambridge, MA: The MIT Press.
Preston, S. D. & de Waal, F. B. M. (2002). Empathy: Its ultimate and proximate bases. Behavior and Brain Sciences, 25, 1-72
Prochazkova, E. & Kret, M.E. (2017). Connecting minds and sharing emotions through mimicry: A neurocognitive model of emotional contagion. Neuroscience & Biobehavioral Reviews, Volume 80, September 2017, Pages 99-114.
Quintana, D.S. & Guastella, A.J. (2020). An Allostatic Theory of Oxytocin. Trends in Cognitive Science 24(7): 515-528.
Shamay-Tsoory S.G. & Abu-Akel A. (2016). The Social Salience Hypothesis of Oxytocin. Biological Psychiatry, 79(3): 194-202.