LOCEN Research Topics

LOCEN research topics focus on three classes of phenomena believed to be paramount for the understanding of truly intelligent behaviour:

1. Motivations, both as sources of behaviour selection and energisation and as sources of learning signals:
   • Extrinsic motivations: related to the homeostatic regulations (body integrity, food, water, sex). These motivations are at the basis of the survival and reproductive capabilities of practically all organisms. The study of extrinsic motivations of the group, that started with the project ICEA and remains central for the group, pivots on amygdala and striatum (especially the nucleus accumbens), the neuromodulators (expecially dopamine), but it also involves other areas such as the hypothalamus. We are also interested in the study of "psychiatric" disturbances,  seen as a pathological functioning of the extrinsic motivation/neuromodulatory systems, and so we have developped a model of the psychobiological rat-models on stress copying.
   • Intrinsic motivations: related to the acquisition of knowledge and skills (e.g., exploratory drive, curiosity, novelty, suprise, sense of grouth). These motivations are maximally expressed in children at play (or in scientists willing to know!). The study of intrinsic motivations is central for the project IM-CLeVeR which is now the main project of the group. This study pivots on superior colliculus and its capacity to trigger phasic dopamine signals guiding trial-and-error learning,  on the hippocampus capacity to detect novelty, and on the prefrontal cortex capacity to detect the success/failure in the achievement of goals.
2. Sensorimotor behaviour:
   • Navigation: we have studied this especially within the project ICEA, now over. Navigation remains important for the group as it is one of the most studied behaviours in rats: the bio-psycholgy/neo-behaviourist literature is very important as a source of target behaviours for the study of the group, expecially for the study of the relation betwen goal-directed behaviours and habits. 
   • Eye control and attention: This has been an important research topic within the project MindRACES, now over. The topic remains fundamental for the group as we believe that the study of many behaviors can be distorted if one does not take into due consideration attention. So we tend to explicitly model (overt) attention in various of our models.
   • Eye-arm-hand coordination: the study of the tria-and-error aquisition of this coordination is one of the major topics of investigation of the group and the project IM-CLeVeR. We are particularly interested in the hierarchical organisation of such coordination, so we study it with bio-inspired hierarchical reinforcement learning models, mostly based on the actor-critic model ideas, and with bio-constrained models that take into consideration the hierarchical organisation of the three main striato-cortical looops (pivoting on lateral, medial, and ventral striatum). We also have important models on the cortical hierarchical organisation of motor behaviour encompassing both the dorsal neural pathways, encoding affordances for the control of reach grasp and look (with the involvement of canonical and mirror neurons), and the ventral neural pathways, pivoting on prefrontal cortex and its capacity to select affordances on the basis of the organism's homeostatic needs and the external context.
3. Biologically plausible learning:
   • Trial-and-error learning: as mentioned, we study this topic with both bio-inspired models (hierarchical reinforcement-learning  actor-critic models) and bio-constrained models (detail models of basal ganglia-cortical loops). We are also very interested in models of developmental-psychology data studying how motor skills develop in children (e.g., reaching; grasping in the future).
   • Hebbian learning: our bio-constrained models are usually based on Hebbian learning and on dopamine-driven Hebbian learning. We are also working on differential Hebbian learning expecially for the models on amygdala.
   • Unsupervised learning: we use Kononen learning in some of our bio-inspired models (e.g., on compatibility effects).



LOCEN Research Method: Computational Embodied Neuroscience

LOCEN's research method has been called Computational Embodied Neuroscience (CEN). This method has been influenced by both the Simulation of Adaptive Behaviour approach (SAB) and by Computational Neuroscience (CN).

CEN draws from the SAB approach the importance given to natural evolution to understand behaviour in terms of its adaptive function, the view of behaviour as a phenomenon emergent from the dynamical interaction between the brain, the body, and the environment, the aim to understand the functioning of organisms as a whole, and the view of brain as a complex adaptive system.

CEN draws from Computational Neuroscience the goal to finely reproduce and explain the experimental data on behaviour and brain anatomy and physiology, and the rigor of models. In general, CEN can be said to be a ''top down'' approach that wants to arrive to understand brain starting from behaviour and its adaptive function for organisms.

In detail, CEN is based on 8 principles that can be summarised as follows:

1. Computational models. Use computational models to understand brain, body, and behaviour as they form a whole complex dynamic system (see below). Models allows the formalisation of theories and this yields rigor, completeness, clarity. Models produce quantitative empirically-testable predictions. Computational models, contrary to mathematical models that have to be solved analytically, allow studying complex systems as they can be solved numerically. Computational models allow the integration of interdisciplianary knowledge. Computational models produce knowledge in three ways: (a) Predictions: models generate detailed testable predictions; (b) Generation of new hypotheses: very often when you translate a theory into a model you discover that it has ''knowledge gaps'': at this point you have to formulate new hypotheses that might lead you to a new discoveries; (c) Theoretical advancement and integration: neuroscience and psychology are producing a huge amount of data but they have problems to produce unified pictures: computational models have a huge power in integrating such information and in generating unifying theories.
2.  Theoretical cumulativity. The hallmark of science is cumulativity. This means that a good scientific method is one which assures the cumulativity of knowledge in time, by allowing you to rank theories and models and discard those that explain less of the phenomena of interest. This in particular means that we should avoid building a different model for each of the phenomena we investigate, but to build models based on principles that account for an increasing number of phenomena of a certain class.
3. Complex systems framework. Brain, body, and environement form a complex adaptive system and computational models and robots are the only means you can use to understand their emergent properties. 
4. Evolution framework. Seek explanations of brain and behaviour within the theoretical framework of evolution. The theory of evolution is the best theoretical framework to understand any biological phenomena, and in this respect brain and behaviour are not an exception. Organisms' behaviour is as it is as it evolved to increase their survival and reproductive chances. Brain is as it is as it evolved to produce such behaviour. So when we try to understand brain and behaviour the question on their adaptive function is paramount.
5. Behavioural constraints. Constrain models by requiring that they reproduce specific quantitative data on behaviour, furnished by psychology, ethology, and other disciplines on behaviour. In fact anecdotal and qualitative comparisons of models with data are not enough compelling to select models.
6. Brain constraints. Constrain the assumptions on the architecture, functioning, and learning mechanisms of models on the basis of data on brain furnished by neuroscience. In fact you can always build different models to reproduce an observed behaviour: using neuroscientic constraints greatly aids the selection of models.
7. Embodiment constraints. Test your models within embodied systems, endowed with realistic bodies and interacting with realistic environments via noisy and quantitative sensors and actuators. This constrain allow developing and selecting models which are capable of scaling to reproduce the complexity of behaviour observed in real organisms. It also forces models to be inherently quantitative (i.e. non-symbolic) and robust in the face of noise of sensors and actuators.
8. Learning constraints. Reproduce learning processes which lead to the target behavious. This is important to understand not only how behaviour function, but also the mechanisms which lead to its acquisition. Indeed, a substantial part of brain structure serves organisms' need to acquire behaviour with experience and learning processes.

As for now it is difficult to completely satisfy all aforementioned constraints within the same models, LOCEN tends to build two types of models that differ for the emphasis given to the detail reproduction of the brain on one side, or to the complexity of the organisms' sensorimotor interactions with the environment (behaviour) on the other side:

(a) bio-inspired embodied models: these reproduce some important aspects of the complex sensorimotor interaction with the environment and rely upon models that are computationally powerful but rather abstract with respect to the brain anatomy and physiology (e.g., they are based on hierarchical actor-critic reinforcement-learning models);

(b) bio-constrained models, which reproduce the system-level anatomy and some aspects of the physiology of brain (for example formed by a system encompassing the amygdala for emotional regulation, the basal ganglia for action selection, the prefrontal cortex for goal-directed behaviour, and motor cortex for motor control), but are rather abstract and simplified in terms of embodiment and sensorimotor interactions with the environment.

You can find a more detailed description  CEN reserch method here:

• Mannella Francesco, Mirolli Marco, Baldassarre Gianluca (2010). The interplay of pavlovian and instrumental processes in devaluation experiments: a computational embodied neuroscience model tested with a simulated rat. In Tosh Colin and Ruxton Graeme (eds), "Modelling Perception With Artificial Neural Networks", pp. 93-113. Cambridge, Cambridge University press.
• Caligiore Daniele, Borghi Anna, Parisi Domenico, Baldassarre Gianluca (2010). TRoPICALS: A computational embodied neuroscience model of compatibility effects. Psycological Review, vol. 117, Issue 4, pp. 1188-1228.