Research | Natural Computation

Mission

The Natural Computation Lab conducts research activities at the intersection of nature-inspired computation and real-world problem-solving from two different, and complementary, points of view, characterized by a strong mutual interaction:

• the application of computational simulations able to replicate natural phenomena to better understand them;
• the development of Artificial Intelligence (AI) algorithms, which draw their metaphorical inspiration from systems and phenomena occurring in the natural world.

The Lab’s mission is to foster a deep understanding of natural systems as models of computation. It covers foundational research, such as the development of new algorithms and the design of innovative and interdisciplinary applications of natural computation methods to address the real world's complex problems.

In particular, the research activity focuses on a wide array of topics, including:

• Nature-Inspired Algorithms: foundational work in Evolutionary Computation, Developmental and Grammatical Computing, Swarm Intelligence, Immunocomputing, Neurocomputing, metaheuristics, and related methodologies.
• Explainable and Interpretable AI (XAI): advancing fair, interpretable, and transparent AI solutions specifically tailored to scientific domains and real-world applications, such as analysis of heuristic optimization algorithms, development of novel explainable AI methods, E-Health, Personalized and Precision Medicine, etc. This interdisciplinary effort involves collaboration with domain-specific experts to develop explainable systems that are both effective and user-friendly.
• Neuro-computational models of motor learning: understanding brain neural processes involved in the generation of human complex motor tasks to model the interaction of the different levels of the nervous system, and to evaluate to which extent lesions of different part of the nervous system affect the motor performance.

• Handwritten document processing: developing methods and algorithms for automatic reading through an approach based on neurocomputational models of motor learning and machine learning techniques, and a methodology to evaluate the performance of such systems in a user-centric perspective.

• Human-Robot Interaction: defining an index of comfort: evaluating the comfort by measuring the similarity to the human movement, which we term human likeness, so addressing a notable gap in the literature in the absence of a quantitative index for such a similarity. The creation of an objective measure accounting for the specific characteristics of the human movement will represent a significant advancement in the field and it will facilitate the development of collaborative robots both in industry and in rehabilitation medicine. A large experimental campaign of collaborative human-robot interaction sessions will then validate the comfort index.

Description

Explainable and Interpretable AI (XAI)

This research topic aims to investigate Interpretable Artificial Intelligence methodologies based on Natural Computation suitable for addressing complex problems in both medicine and healthcare, with special emphasis on solving diagnostic, prognostic, and signal processing problems. In particular, the research activity focuses on the definition and application of Natural Computing methodologies to two specific areas:

• early diagnosis of progressive neurodegenerative diseases (Parkinson’s disease, Alzheimer’s disease, etc.) would provide faster and more effective intervention strategies. In this context, Natural Computing methodologies are capable of automatically identifying diseased subjects through the analysis of handwritten forms drawn by potential patients. The use of such methodologies is particularly interesting in that it would make it possible to infer explicit and interpretable patterns of classification and, at the same time, to automatically identify a suitable subset of relevant features for a correct diagnosis.
• development of a Natural Computing approach to improve the treatment of Type 1 diabetic patients on insulin therapy. As a disease associated with a pancreas that generates an insufficient amount of insulin, one way to improve the quality of life of these patients is to make an artificial pancreas capable of artificially regulating external insulin dosage. In this context, Natural Computing methodologies that aim to extrapolate interpretable regression models able to estimate blood glucose through interstitial glucose values extracted from wearable sensors is a vital preliminary step in building the basic components of an artificial pancreas. In addition, given the high complexity of the blood glucose and insulin reciprocal metabolic interactions, the research activity aims to infer explicit mathematical models of such complex interactions.

Neuro-computational models of motor learning

The research activity focuses on the analysis of the neural processes involved in the generation of complex motor tasks and aims to understand how different levels of the nervous system interact, contributing to the gradual improvement of motor performance during learning. The objective is pursued through the realization of models that instantiate in a neural network the main anatomical, physiological, and biological characteristics of the cortical areas (motor, somatosensory, etc.) and subcortical systems (basal ganglia and cerebellum) involved in the realization of motor tasks. Understanding these mechanisms provides an important contribution to various fields of application, from the design of robotic limbs to the development of new rehabilitation treatments for neurodegenerative diseases involving movement, such as Parkinson's disease, Huntington chorea, and Tourette's syndrome, as well as autism spectrum disorder related pathologies.

Handwritten document processing

Understanding the role of the nervous system and the interaction with the musculoskeletal system during the learning of complex movements is used to develop learning models for reading handwritten documents that integrate motor and cognitive aspects for the identification of the elementary movements used by the writing with the cognitive representation of the task. This approach allows the development of a universal approach, independent of the specific language for all applications that require the transformation of the textual content of a digital image into its representation through character strings.

The research activity carried out has also highlighted the need to have a methodology for evaluating the cost/benefit ratio in the use of these systems which, unlike the state of the art in which the performances are evaluated on a test set without taking into account of the application context, takes into account the methods of human-machine interaction and its effects on the overall performance of the system in the specific context.

Federated Learning through Nature-Inspired Algorithms

In the past few years, Federated Learning (FL) has emerged as an effective approach for training Neural Networks (NNs) over a computing network while preserving data privacy. Most existing FL approaches require defining a priori 1) a predefined structure for all the NNs running on the clients and 2) an explicit aggregation procedure. These can be limiting factors in cases where pre-defining such algorithmic details is difficult. Recently, NCLab research activity proposed NEvoFed, an FL method that leverages NeuroEvolution running on the clients, in which the NN structures are heterogeneous and the aggregation is implicitly accomplished on the client side. Currently, the research activity is devoted to designing a novel approach to FL that does not require learning models, i.e., neural network parameters, to be distributed over the networks, thus taking a step towards security improvement. The only information exchanged in client/server communication is the performance of each model on local data, allowing the emergence of optimal NN architectures without needing any kind of models’ aggregation. Another appealing feature of the framework to design is that it should be used with any Machine Learning algorithm provided that, during the learning phase, the model updates do not depend on the input data.

Financial Networks Optimization

Banking is, by its nature, a risky activity: during their daily operation, banks collect short-term liabilities (typically customer deposits) and invest in long-term assets (shares, buildings, loans to firms, etc.). However, such deposits and, therefore, the internal liquidity of the banks fluctuate over time, for example, due to payments and withdrawals made by these deposit holders. This exposes banks to the risk of a liquidity shortage, which should then be covered through the sale of their long-term assets, typically at discounted prices. To avoid incurring this situation, banks insure each other against liquidity risk through the cross-holding of liquidity positions. Such interbank deposits create actual financial networks which, on one side, allow the easy reallocate of money from banks in surplus to banks in deficit, on the other, create the possibility, in the event of a substantial loss of liquidity or bankruptcy of one or more banks of the collapse of all the others linked to them due to the financial contagion. Economists represent financial networks as weighted, directed graphs, where nodes represent banks or other intermediaries and edges represent financial transactions or interbank exposures. The research activity is aimed at identifying the best financial network topologies that minimize the risk of systemic crises in the presence of liquidity shocks with different characteristics. Borrowing the graph representation of the networks from economics, the problem is reformulated as a high-dimensional optimization problem of the weights and the structure of a graph, and it is addressed by using distributed Evolutionary Computation and Swarm Intelligence algorithms.

Human-Robot Interaction

Since the emergence of Human-Robot Interaction (HRI) as a field}, several studies have been conducted over the use of new robotic systems in the manufacturing industry with respect to the ``social'' dimension, particularly, in relation to the limits of cognitive and perceptual workload for robot operators}, that can compromise acceptance.

A strategy to tackle user robot acceptance goes through the design of robots resembling humans in both their physical aspect and their behavior. To this extent, recent research has led to the development of brain-inspired intelligent robots that are capable of emulating humans and animals in both their external structures and internal mechanisms. This is achieved through the integration of visual cognition, decision-making, motion control, and musculoskeletal systems.

This research project is sustained by the concept of motor contagion, which posits that an individual's motor actions are involuntarily influenced by observing and experiencing others’ movements. The degree of motor contagion is influenced by the human likeness of a robot: the greater the resemblance between robot and human movements and appearance, the stronger the motor contagion effect. Movement kinematics, such as the velocity profile and the trajectory of the limb, might also vary the degree of motor resonance evoked in the observer. Specifically, high-speed movements have been observed to increase anxiety and risk perception among human co-workers, who, consequently, exhibit unpleasant and inefficient behavior. Therefore, time-parametrization is crucial in designing motion planning processes for human-robot interaction.

In such a context, the research activities aim at defining a global, quantitative index of comfort in human-robot-interaction that can be computed without the need to experience the movement itself. Such an index does not currently exist: the available indices are either qualitative, a posteriori or quantitative, primarily based on local motion characteristics. Furthermore, since in most application scenarios, like motor rehabilitation or collaborative robotics, it is preferable that planning is independent of experience, it is crucial to have an a priori comfort index in the pursuit of planning human-like trajectories.

Teaching

The NCLab provides support for the courses in Artificial Intelligence (MSc in Electrical Engineering for Digital Energy) and Natural Computing (MSc in Information Engineering for Digital Medicine and MSc in Information Engineering), for the thesis activities of students of the BSc and MSc in Computer Engineering, and PhD students in Information Engineering.

As part of the project activities of the Artificial Intelligence and Natural Computing courses, the laboratory provides hardware platforms and development environments for the design and simulation of intelligent agents, and the development and performance evaluation of Natural Computation methodologies.

Equipment

Wacom Intuos A4

• Description: Graphics tablet
• Main use: Online acquisition of handwriting and drawing samples

Wacom Bamboo folio

• Description: Graphics tablet
• Main use: Online acquisition of handwriting and drawing samples

ASL-5000

• Description: Eye-tracking with magnetic head tracking
• Main use: validation of computational models of saccadic movements during scene observation under controlled conditions

ASL-500 ME

• Description: mobile eye-tracking
• Main use: validation of computational models of saccadic movements during scene observation under controlled conditions

Server Linux

• Description: high-performance multicore and GPU server
• Main use: implementation of Evolutionary Computation, Swarm Intelligence, Immunocomputing, Neuroevolution, and Deep Neural Network systems

Server Windows

• Description: high-performance multicore server RAID architecture
• Main use: implementation of neuro-computational models