Interpretable Convolutional Neural Networks for Decoding and Analyzing Neural Time Series Data

Machine learning is widely adopted to decode multi-variate neural time series, including electroencephalographic (EEG) and single-cell recordings. Recent solutions based on deep learning (DL) outperformed traditional decoders by automatically extracting relevant discriminative features from raw or minimally pre-processed signals. Convolutional Neural Networks (CNNs) have been successfully applied to EEG and are the most common DL-based EEG decoders in the state-of-the-art (SOA). However, the current research is affected by some limitations. SOA CNNs for EEG decoding usually exploit deep and heavy structures with the risk of overfitting small datasets, and architectures are often defined empirically. Furthermore, CNNs are mainly validated by designing within-subject decoders. Crucially, the automatically learned features mainly remain unexplored; conversely, interpreting these features may be of great value to use decoders also as analysis tools, highlighting neural signatures underlying the different decoded brain or behavioral states in a data-driven way. Lastly, SOA DL-based algorithms used to decode single-cell recordings rely on more complex, slower to train and less interpretable networks than CNNs, and the use of CNNs with these signals has not been investigated. This PhD research addresses the previous limitations, with reference to P300 and motor decoding from EEG, and motor decoding from single-neuron activity. CNNs were designed light, compact, and interpretable. Moreover, multiple training strategies were adopted, including transfer learning, which could reduce training times promoting the application of CNNs in practice. Furthermore, CNN-based EEG analyses were proposed to study neural features in the spatial, temporal and frequency domains, and proved to better highlight and enhance relevant neural features related to P300 and motor states than canonical EEG analyses. Remarkably, these analyses could be used, in perspective, to design novel EEG biomarkers for neurological or neurodevelopmental disorders. Lastly, CNNs were developed to decode single-neuron activity, providing a better compromise between performance and model complexity.

Preharvest factors affecting quality and antioxidant levels on Abate Fétel pears: study of superficial scald with a multivariate statistical approach and neural networks

There are only a few insights concerning the influence that agronomic and management variability may have on superficial scald (SS) in pears. Abate Fétel pears were picked during three seasons (2018, 2019 and 2020) from thirty commercial orchards in the Emilia Romagna region, Italy. Using a multivariate statistical approach, high heterogeneity between farms for SS development after cold storage with regular atmosphere was demonstrated. Indeed, some factors seem to affect SS in all growing seasons: high yields, soil texture, improper irrigation and Nitrogen management, use of plant growth regulators, late harvest, precipitations, Calcium and cow manure, presence of nets, orchard age, training system and rootstock. Afterwards, we explored the spatio/temporal variability of fruit attributes in two pear orchards. Environmental and physiological spatial variables were recorded by a portable RTK GPS. High spatial variability of the SS index was observed. Through a geostatistical approach, some characteristics, including soil electrical conductivity and fruit size, have been shown to be negatively correlated with SS. Moreover, regression tree analyses were applied suggesting the presence of threshold values of antioxidant capacity, total phenolic content, and acidity against SS. High pulp firmness and IAD values before storage, denoting a more immature fruit, appeared to be correlated with low SS. Finally, a convolution neural networks (CNN) was tested to detect SS and the starch pattern index (SPI) in pears for portable device applications. Preliminary statistics showed that the model for SS had low accuracy but good precision, and the CNN for SPI denoted good performances compared to the Ctifl and Laimburg scales. The major conclusion is that Abate Fétel pears can potentially be stored in different cold rooms, according to their origin and quality features, ensuring the best fruit quality for the final consumers. These results might lead to a substantial improvement in the Italian pear industry.

Directly training spiking neural networks for cyber-physical systems: from supervised to reinforcement learning

Spiking Neural Networks (SNNs) are bio-inspired Artificial Neural Networks (ANNs) utilizing discrete spiking signals, akin to neuron communication in the brain, making them ideal for real-time and energy-efficient Cyber-Physical Systems (CPSs). This thesis explores their potential in Structural Health Monitoring (SHM), leveraging low-cost MEMS accelerometers for early damage detection in motorway bridges. The study focuses on Long Short-Term SNNs (LSNNs), although their complex learning processes pose challenges. Comparing LSNNs with other ANN models and training algorithms for SHM, findings indicate LSNNs' effectiveness in damage identification, comparable to ANNs trained using traditional methods. Additionally, an optimized embedded LSNN implementation demonstrates a 54% reduction in execution time, but with longer pre-processing due to spike-based encoding. Furthermore, SNNs are applied in UAV obstacle avoidance, trained directly using a Reinforcement Learning (RL) algorithm with event-based input from a Dynamic Vision Sensor (DVS). Performance evaluation against Convolutional Neural Networks (CNNs) highlights SNNs' superior energy efficiency, showing a 6x decrease in energy consumption. The study also investigates embedded SNN implementations' latency and throughput in real-world deployments, emphasizing their potential for energy-efficient monitoring systems. This research contributes to advancing SHM and UAV obstacle avoidance through SNNs' efficient information processing and decision-making capabilities within CPS domains.

PCM-based in-memory computing: architectures, circuits and applications

Analog In-memory Computing (AIMC) has been proposed in the context of Beyond Von Neumann architectures as a valid strategy to reduce internal data transfers energy consumption and latency, and to improve compute efficiency. The aim of AIMC is to perform computations within the memory unit, typically leveraging the physical features of memory devices. Among resistive Non-volatile Memories (NVMs), Phase-change Memory (PCM) has become a promising technology due to its intrinsic capability to store multilevel data. Hence, PCM technology is currently investigated to enhance the possibilities and the applications of AIMC. This thesis aims at exploring the potential of new PCM-based architectures as in-memory computational accelerators. In a first step, a preliminar experimental characterization of PCM devices has been carried out in an AIMC perspective. PCM cells non-idealities, such as time-drift, noise, and non-linearity have been studied to develop a dedicated multilevel programming algorithm. Measurement-based simulations have been then employed to evaluate the feasibility of PCM-based operations in the fields of Deep Neural Networks (DNNs) and Structural Health Monitoring (SHM). Moreover, a first testchip has been designed and tested to evaluate the hardware implementation of Multiply-and-Accumulate (MAC) operations employing PCM cells. This prototype experimentally demonstrates the possibility to reach a 95% MAC accuracy with a circuit-level compensation of cells time drift and non-linearity. Finally, empirical circuit behavior models have been included in simulations to assess the use of this technology in specific DNN applications, and to enhance the potentiality of this innovative computation approach.

Deploying and processing neural representations of signals

Neural representations (NR) have emerged in the last few years as a powerful tool to represent signals from several domains, such as images, 3D shapes, or audio. Indeed, deep neural networks have been shown capable of approximating continuous functions that describe a given signal with theoretical infinite resolution. This finding allows obtaining representations whose memory footprint is fixed and decoupled from the resolution at which the underlying signal can be sampled, something that is not possible with traditional discrete representations, e.g., grids of pixels for images or voxels for 3D shapes. During the last two years, many techniques have been proposed to improve the capability of NR to approximate high-frequency details and to make the optimization procedures required to obtain NR less demanding both in terms of time and data requirements, motivating many researchers to deploy NR as the main form of data representation for complex pipelines. Following this line of research, we first show that NR can approximate precisely Unsigned Distance Functions, providing an effective way to represent garments that feature open 3D surfaces and unknown topology. Then, we present a pipeline to obtain in a few minutes a compact Neural Twin® for a given object, by exploiting the recent advances in modeling neural radiance fields. Furthermore, we move a step in the direction of adopting NR as a standalone representation, by considering the possibility of performing downstream tasks by processing directly the NR weights. We first show that deep neural networks can be compressed into compact latent codes. Then, we show how this technique can be exploited to perform deep learning on implicit neural representations (INR) of 3D shapes, by only looking at the weights of the networks.

Mathematical models of cognitive processes

The research activity carried out during the PhD course was focused on the development of mathematical models of some cognitive processes and their validation by means of data present in literature, with a double aim: i) to achieve a better interpretation and explanation of the great amount of data obtained on these processes from different methodologies (electrophysiological recordings on animals, neuropsychological, psychophysical and neuroimaging studies in humans), ii) to exploit model predictions and results to guide future research and experiments. In particular, the research activity has been focused on two different projects: 1) the first one concerns the development of neural oscillators networks, in order to investigate the mechanisms of synchronization of the neural oscillatory activity during cognitive processes, such as object recognition, memory, language, attention; 2) the second one concerns the mathematical modelling of multisensory integration processes (e.g. visual-acoustic), which occur in several cortical and subcortical regions (in particular in a subcortical structure named Superior Colliculus (SC)), and which are fundamental for orienting motor and attentive responses to external world stimuli. This activity has been realized in collaboration with the Center for Studies and Researches in Cognitive Neuroscience of the University of Bologna (in Cesena) and the Department of Neurobiology and Anatomy of the Wake Forest University School of Medicine (NC, USA). PART 1. Objects representation in a number of cognitive functions, like perception and recognition, foresees distribute processes in different cortical areas. One of the main neurophysiological question concerns how the correlation between these disparate areas is realized, in order to succeed in grouping together the characteristics of the same object (binding problem) and in maintaining segregated the properties belonging to different objects simultaneously present (segmentation problem). Different theories have been proposed to address these questions (Barlow, 1972). One of the most influential theory is the so called “assembly coding”, postulated by Singer (2003), according to which 1) an object is well described by a few fundamental properties, processing in different and distributed cortical areas; 2) the recognition of the object would be realized by means of the simultaneously activation of the cortical areas representing its different features; 3) groups of properties belonging to different objects would be kept separated in the time domain. In Chapter 1.1 and in Chapter 1.2 we present two neural network models for object recognition, based on the “assembly coding” hypothesis. These models are networks of Wilson-Cowan oscillators which exploit: i) two high-level “Gestalt Rules” (the similarity and previous knowledge rules), to realize the functional link between elements of different cortical areas representing properties of the same object (binding problem); 2) the synchronization of the neural oscillatory activity in the γ-band (30-100Hz), to segregate in time the representations of different objects simultaneously present (segmentation problem). These models are able to recognize and reconstruct multiple simultaneous external objects, even in difficult case (some wrong or lacking features, shared features, superimposed noise). In Chapter 1.3 the previous models are extended to realize a semantic memory, in which sensory-motor representations of objects are linked with words. To this aim, the network, previously developed, devoted to the representation of objects as a collection of sensory-motor features, is reciprocally linked with a second network devoted to the representation of words (lexical network) Synapses linking the two networks are trained via a time-dependent Hebbian rule, during a training period in which individual objects are presented together with the corresponding words. Simulation results demonstrate that, during the retrieval phase, the network can deal with the simultaneous presence of objects (from sensory-motor inputs) and words (from linguistic inputs), can correctly associate objects with words and segment objects even in the presence of incomplete information. Moreover, the network can realize some semantic links among words representing objects with some shared features. These results support the idea that semantic memory can be described as an integrated process, whose content is retrieved by the co-activation of different multimodal regions. In perspective, extended versions of this model may be used to test conceptual theories, and to provide a quantitative assessment of existing data (for instance concerning patients with neural deficits). PART 2. The ability of the brain to integrate information from different sensory channels is fundamental to perception of the external world (Stein et al, 1993). It is well documented that a number of extraprimary areas have neurons capable of such a task; one of the best known of these is the superior colliculus (SC). This midbrain structure receives auditory, visual and somatosensory inputs from different subcortical and cortical areas, and is involved in the control of orientation to external events (Wallace et al, 1993). SC neurons respond to each of these sensory inputs separately, but is also capable of integrating them (Stein et al, 1993) so that the response to the combined multisensory stimuli is greater than that to the individual component stimuli (enhancement). This enhancement is proportionately greater if the modality-specific paired stimuli are weaker (the principle of inverse effectiveness). Several studies have shown that the capability of SC neurons to engage in multisensory integration requires inputs from cortex; primarily the anterior ectosylvian sulcus (AES), but also the rostral lateral suprasylvian sulcus (rLS). If these cortical inputs are deactivated the response of SC neurons to cross-modal stimulation is no different from that evoked by the most effective of its individual component stimuli (Jiang et al 2001). This phenomenon can be better understood through mathematical models. The use of mathematical models and neural networks can place the mass of data that has been accumulated about this phenomenon and its underlying circuitry into a coherent theoretical structure. In Chapter 2.1 a simple neural network model of this structure is presented; this model is able to reproduce a large number of SC behaviours like multisensory enhancement, multisensory and unisensory depression, inverse effectiveness. In Chapter 2.2 this model was improved by incorporating more neurophysiological knowledge about the neural circuitry underlying SC multisensory integration, in order to suggest possible physiological mechanisms through which it is effected. This endeavour was realized in collaboration with Professor B.E. Stein and Doctor B. Rowland during the 6 months-period spent at the Department of Neurobiology and Anatomy of the Wake Forest University School of Medicine (NC, USA), within the Marco Polo Project. The model includes four distinct unisensory areas that are devoted to a topological representation of external stimuli. Two of them represent subregions of the AES (i.e., FAES, an auditory area, and AEV, a visual area) and send descending inputs to the ipsilateral SC; the other two represent subcortical areas (one auditory and one visual) projecting ascending inputs to the same SC. Different competitive mechanisms, realized by means of population of interneurons, are used in the model to reproduce the different behaviour of SC neurons in conditions of cortical activation and deactivation. The model, with a single set of parameters, is able to mimic the behaviour of SC multisensory neurons in response to very different stimulus conditions (multisensory enhancement, inverse effectiveness, within- and cross-modal suppression of spatially disparate stimuli), with cortex functional and cortex deactivated, and with a particular type of membrane receptors (NMDA receptors) active or inhibited. All these results agree with the data reported in Jiang et al. (2001) and in Binns and Salt (1996). The model suggests that non-linearities in neural responses and synaptic (excitatory and inhibitory) connections can explain the fundamental aspects of multisensory integration, and provides a biologically plausible hypothesis about the underlying circuitry.

Tensor-Train decomposition for image classification problems

In these last years a great effort has been put in the development of new techniques for automatic object classification, also due to the consequences in many applications such as medical imaging or driverless cars. To this end, several mathematical models have been developed from logistic regression to neural networks. A crucial aspect of these so called classification algorithms is the use of algebraic tools to represent and approximate the input data. In this thesis, we examine two different models for image classification based on a particular tensor decomposition named Tensor-Train (TT) decomposition. The use of tensor approaches preserves the multidimensional structure of the data and the neighboring relations among pixels. Furthermore the Tensor-Train, differently from other tensor decompositions, does not suffer from the curse of dimensionality making it an extremely powerful strategy when dealing with high-dimensional data. It also allows data compression when combined with truncation strategies that reduce memory requirements without spoiling classification performance. The first model we propose is based on a direct decomposition of the database by means of the TT decomposition to find basis vectors used to classify a new object. The second model is a tensor dictionary learning model, based on the TT decomposition where the terms of the decomposition are estimated using a proximal alternating linearized minimization algorithm with a spectral stepsize.

Flexible Computing Systems For AI Acceleration At The Extreme Edge Of The IoT

Embedding intelligence in extreme edge devices allows distilling raw data acquired from sensors into actionable information, directly on IoT end-nodes. This computing paradigm, in which end-nodes no longer depend entirely on the Cloud, offers undeniable benefits, driving a large research area (TinyML) to deploy leading Machine Learning (ML) algorithms on micro-controller class of devices. To fit the limited memory storage capability of these tiny platforms, full-precision Deep Neural Networks (DNNs) are compressed by representing their data down to byte and sub-byte formats, in the integer domain. However, the current generation of micro-controller systems can barely cope with the computing requirements of QNNs. This thesis tackles the challenge from many perspectives, presenting solutions both at software and hardware levels, exploiting parallelism, heterogeneity and software programmability to guarantee high flexibility and high energy-performance proportionality. The first contribution, PULP-NN, is an optimized software computing library for QNN inference on parallel ultra-low-power (PULP) clusters of RISC-V processors, showing one order of magnitude improvements in performance and energy efficiency, compared to current State-of-the-Art (SoA) STM32 micro-controller systems (MCUs) based on ARM Cortex-M cores. The second contribution is XpulpNN, a set of RISC-V domain specific instruction set architecture (ISA) extensions to deal with sub-byte integer arithmetic computation. The solution, including the ISA extensions and the micro-architecture to support them, achieves energy efficiency comparable with dedicated DNN accelerators and surpasses the efficiency of SoA ARM Cortex-M based MCUs, such as the low-end STM32M4 and the high-end STM32H7 devices, by up to three orders of magnitude. To overcome the Von Neumann bottleneck while guaranteeing the highest flexibility, the final contribution integrates an Analog In-Memory Computing accelerator into the PULP cluster, creating a fully programmable heterogeneous fabric that demonstrates end-to-end inference capabilities of SoA MobileNetV2 models, showing two orders of magnitude performance improvements over current SoA analog/digital solutions.

Uncovering the encoding and transmission of behaviourally relevant information in neural activity

Most cognitive functions require the encoding and routing of information across distributed networks of brain regions. Information propagation is typically attributed to physical connections existing between brain regions, and contributes to the formation of spatially correlated activity patterns, known as functional connectivity. While structural connectivity provides the anatomical foundation for neural interactions, the exact manner in which it shapes functional connectivity is complex and not yet fully understood. Additionally, traditional measures of directed functional connectivity only capture the overall correlation between neural activity, and provide no insight on the content of transmitted information, limiting their ability in understanding neural computations underlying the distributed processing of behaviorally-relevant variables. In this work, we first study the relationship between structural and functional connectivity in simulated recurrent spiking neural networks with spike timing dependent plasticity. We use established measures of time-lagged correlation and overall information propagation to infer the temporal evolution of synaptic weights, showing that measures of dynamic functional connectivity can be used to reliably reconstruct the evolution of structural properties of the network. Then, we extend current methods of directed causal communication between brain areas, by deriving an information-theoretic measure of Feature-specific Information Transfer (FIT) quantifying the amount, content and direction of information flow. We test FIT on simulated data, showing its key properties and advantages over traditional measures of overall propagated information. We show applications of FIT to several neural datasets obtained with different recording methods (magneto and electro-encephalography, spiking activity, local field potentials) during various cognitive functions, ranging from sensory perception to decision making and motor learning. Overall, these analyses demonstrate the ability of FIT to advance the investigation of communication between brain regions, uncovering the previously unaddressed content of directed information flow.