196,115 research outputs found
Obiettivi del programma formativo per l’integrazione delle cure dei malati gravi e linee di indirizzo sanitarie
No full textSemantic processing with and without awareness. Insights from computational linguistics and semantic priming.
No full textDuring my PhD, I’ve explored how native speakers access semantic information from lexical stimuli, and weather consciousness plays a role in the process of meaning construction. In a first study, I exploited the metaphor linking time and space to assess the specific contribution of linguistically–coded information to the emergence of priming. In fact, time is metaphorically arranged on either the horizontal or the sagittal axis in space (Clark, 1973), but only the latter comes up in language (e.g., "a bright future in front of you"). In a semantic categorization task, temporal target words (e.g., earlier, later) were primed by spatial words that were processed either consciously (unmasked) or unconsciously (masked). With visible primes, priming was observed for both lateral and sagittal words; yet, only the latter ones led to a significant effect when the primes were masked. Thus, unconscious word processing may be limited to those aspects of meaning that emerge in language use. In a second series of experiments, I tried to better characterize these aspects by taking advantage of Distributional Semantic Models (DSMs; Marelli, 2017), which represent word meaning as vectors built upon word co–occurrences in large textual database. I compared state–of–the–art DSMs with Pointwise Mutual Information (PMI; Church & Hanks, 1990), a measure of local association between words that is merely based on their surface co–occurrence. In particular, I tested how the two indexes perform on a semantic priming dataset comprising visible and masked primes, and different stimulus onset asynchronies between the two stimuli. Subliminally, none of the predictor alone elicited significant priming, although participants who showed some residual prime visibility showed larger effect. Post-hoc analyses showed that for subliminal priming to emerge, the additive contribution of both PMI and DSM was required. Supraliminally, PMI outperforms DSM in the fit to the behavioral data. According to these results, what has been traditionally thought of as unconscious semantic priming may mostly rely on local associations based on shallow word cooccurrence. Of course, masked priming is only one possible way to model unconscious perception. In an attempt to provide converging evidence, I also tested overt and covert semantic facilitation by presenting prime words in the unattended vs. attended visual hemifield of brain–injured patients suffering from neglect. In seven sub–acute cases, data show more solid PMI–based than DSM–based priming in the unattended hemifield, confirming the results obtained from healthy participants. Finally, in a fourth work package, I explored the neural underpinnings of semantic processing as revealed by EEG (Kutas & Federmeier, 2011). As the behavioral results of the previous study were much clearer when the primes were visible, I focused on this condition only. Semantic congruency was dichotomized in order to compare the ERP evoked by related and unrelated pairs. Three different types of semantic similarity were taken into account: in a first category, primes and targets were often co–occurring but far in the DSM (e.g., cheese-mouse), while in a second category the two words were closed in the DSM, but not likely to co-occur (e.g., lamp-torch). As a control condition, we added a third category with pairs that were both high in PMI and close in DSMs (e.g., lemon-orange). Mirroring the behavioral results, we observed a significant PMI effect in the N400 time window; no such effect emerged for DSM. References Church, K. W., & Hanks, P. (1990). Word association norms, mutual information, and lexicography. Computational linguistics, 16(1), 22-29. Clark, H. H. (1973). Space, time, semantics, and the child. In Cognitive development and acquisition of language (pp. 27-63). Academic Press. Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting: finding meaning in the N400 component of the event-related brain potential (ERP). Annual review of psychology, 62, 621-647. Marelli, M. (2017). Word-Embeddings Italian Semantic Spaces: a semantic model for psycholinguistic research. Psihologija, 50(4), 503-520. Commentat
Reduced precision floating-point optimization for Deep Neural Network On-Device Learning on microcontrollers
No full textEnabling On-Device Learning (ODL) for Ultra-Low-Power Micro-Controller Units (MCUs) is a key step for post-deployment adaptation and fine-tuning of Deep Neural Network (DNN) models in future TinyML applications. This paper tackles this challenge by introducing a novel reduced precision optimization technique for ODL primitives on MCU-class devices, leveraging the State-of-Art advancements in RISC-V RV32 architectures with support for vectorized 16-bit floating-point (FP16) Single-Instruction Multiple-Data (SIMD) operations. Our approach for the Forward and Backward steps of the Back Propagation training algorithm is composed of specialized shape transform operators and Matrix Multiplication (MM) kernels, accelerated with parallelization and loop unrolling. When evaluated on a single training step of a 2D Convolution layer, the SIMD-optimized FP16 primitives result up to 1.72x faster than the FP32 baseline on a RISC-V-based 8+1-core MCU. An average computing efficiency of 3.11 Multiply and Accumulate operations per clock cycle (MAC/clk) and 0.81 MAC/clk is measured for the end-to-end training tasks of a ResNet8 and a DS-CNN for Image Classification and Keyword Spotting, respectively - requiring 17.1 ms and 6.4 ms on the target platform to compute a training step on a single sample. Overall, our approach results more than two orders of magnitude faster than existing ODL software frameworks for single-core MCUs and outperforms by 1.6x previous FP32 parallel implementations on a Continual Learning setup.& COPY; 2023 Elsevier B.V. All rights reserved
Un modello di calcolo delle forze idrodinamiche indotte da onde irregolari su condotte sottomarine snelle posate sul fondo: primi risultati e verifiche sperimentali
No full text
Towards reliability-based design of rockfall hybrid barriers and attenuators: A focus on the resistances
No full textThe conventional design approach for geotechnical structures presented in Eurocode 7 (EC7) shows limitations when dealing with rockfalls. To overcome these limitations, we propose the application of Reliability Based Design (RBD), which describes the relationship between the actions and the system's resistance through the definition of a reliability index. In this work, particular attention has been given to innovative rockfall protection structures such as hybrid barriers and/or attenuators. Considering the applicability of RDB approach, the paper focuses on the response of these structures to the impact of the block and their absorption capacity at different stress levels. In this context, numerical modelling represents a powerful solution to reproduce the behaviour of these structures subjected to dynamic impacts at different Kinetic Energy levels
Persistent quantum confinement in a Germanium quantum dot solid
No full textQuantum dot solar cells have the advantage that they can harvest the full spectrum of the sun by layers of quantum dots made out of the same semiconductor but with just a different size. This means that such quantum dots must maintain their quantum confinement and therefore their band gap, upon being in an ensemble (solid) in which all quantum dots are connected. When the quantum dot does not have a protective shell or ligands at the surface, which often hinders charge transport, the preservation of quantum confinement in a highly connected solid remains a question. In this work a germanium quantum dot solid is investigated by probing the quantum dots with scanning tunnelling spectroscopy in which only one or few quantum dots are targeted for each interrogation. Besides the band gap, the discrete energy levels at the edges of the band gap, which accompany quantum confinement, are here used as a key-tool to investigate the quantum confinement. This work forms a next step in assessing how quantum confinement is highly persistent and understanding how it can be utilized in solar-cell architectures based on nanoparticle assemblies
Foliar and root applications of the rare sugar tagatose control powdery mildew in soilless grown cucumbers
No full textBiodegradation of a mixture of PAHs was assessed in forest soil microcosms performed either without or
with bioaugmentation using individual fungi and bacterial and a fungal consortia. Respiratory activity,
metabolic intermediates and extent of PAH degradation were determined. In all microcosms the low
molecular weight PAH’s naphthalene, phenanthrene and anthracene, showed a rapid initial rate of
removal. However, bioaugmentation did not significantly affect the biodegradation efficiency for these
compounds. Significantly slower degradation rates were demonstrated for the high molecular weight
PAH’s pyrene, benz[a]anthracene and benz[a]pyrene. Bioaugmentation did not improve the rate or extent
of PAH degradation, except in the case of Aspergillus sp. Respiratory activity was determined by CO2 evolution
and correlated roughly with the rate and timing of PAH removal. This indicated that the PAHs were
being used as an energy source. The native microbiota responded rapidly to the addition of the PAHs and
demonstrated the ability to degrade all of the PAHs added to the soil, indicating their ability to remediate
PAH-contaminated soils
A TinyML Platform for On-Device Continual Learning with Quantized Latent Replays
No full textIn the last few years, research and development on Deep Learning models & techniques for ultra-low-power devices- in a word, TinyML - has mainly focused on a train-then-deploy assumption, with static models that cannot be adapted to newly collected data without cloud-based data collection and fine-tuning. Latent Replay-based Continual Learning (CL) techniques (Pellegrini et al., 2020) enable online, serverless adaptation in principle, but so far they have still been too computation- and memory-hungry for ultra-low-power TinyML devices, which are typically based on microcontrollers. In this work, we introduce a HW/SW platform for end-to-end CL based on a 10-core FP32 -enabled parallel ultra-low-power (PULP) processor. We rethink the baseline Latent Replay CL algorithm, leveraging quantization of the frozen stage of the model and Latent Replays (LRs) to reduce their memory cost with minimal impact on accuracy. In particular, 8-bit compression of the LR memory proves to be almost lossless (-0.26% with 3000LR) compared to the full-precision baseline implementation, but requires 4times less memory, while 7-bit can also be used with an additional minimal accuracy degradation (up to 5%). We also introduce optimized primitives for forward and backward propagation on the PULP processor, together with data tiling strategies to fully exploit its memory hierarchy, while maximizing efficiency. Our results show that by combining these techniques, continual learning can be achieved in practice using less than 64MB of memory - an amount compatible with embedding in TinyML devices. On an advanced 22nm prototype of our platform, called VEGA, the proposed solution performs on average 65 times faster than a low-power STM32 L4 microcontroller, being 37times more energy efficient - enough for a lifetime of 535h when learning a new mini-batch of data once every minute
- …
