133 research outputs found
Fortissat Science Alliance podcast: Alexiei Dingli
No full textAlexiei Dingli was a Professor of Artificial Intelligence in the Faculty of ICT at the University of Malta. He took part in the Fortissat Science Alliance podcast recordings in December 2022.What is the Fortissat Science Alliance?The Fortissat Science Alliance was a Wellcome Trust & Children In Need "Curiosity" project. This scheme provided informal STEM learning opportunities for young people who attended the community centre Getting Better Together Shotts (GBT Shotts) between 2019 and 2023. Due to the COVID-19 pandemic, deliveries had to pivot online so the podcast was founded. These recordings were made via Zoom with warm-up STEM activities sent to every young person in advance, along with a profile page for each researcher, so that they were relaxed and able to ask excellent questions.Link to episode on Spotify.Depending on the broadcast date, podcast deliveries were co-sponsored by Glasgow Science Festival, EXPLORATHON 2021, or EXPLORATHON 2022/23.For the duration of the project, it was supported jointly by Children in Need and the Wellcome Trust. In 2021, EXPLORATHON episodes were supported by the European Commission [grant agreement ID 101036101]. In 2022-23, EXPLORATHON episodes were supported by the Engineering & Physical Sciences Research Council [grant number EP/X020894/1].Author contributions to contentAlexiei Dingli was the guest featured on this episode. Rebecca Hay was the youth worker coordinating the young people who conducted the interviews as well as co-editing and broadcasting the recordings. Iain Hamilton co-edited the episodes. Kirsty Ross was the STEM consultant for the project and uploaded completed episodes to Figshare.</p
1.3-μm uncooled 10 Gb/s directly modulated MQW AlGaInAs/InP laser diodes
No full textIn this paper, we report a novel 1.3-μm uncooled AlGaInAs/InP multiple quantum well (MQW) ridge waveguide laser diodes. By optimizing the design of MQW structure and facet coatings, together with the application of reversed-mesa ridge waveguide (RM-RWG) structure, polyimide planarization, and lift-off processes technology, an uncooled 1.3-μm, 10-Gb/s directly modulated MQW ridge waveguide laser diode was successfully fabricated. The threshold current and the slope efficiency were 7 mA and 0.48 mW/mA, respectively. The directly modulated bandwidths of 11 and 9.2 GHz were achieved at room temperature and 80 Celsius degrees, respectively
Gelis exareolatus
No full textGelis exareolatus (Fӧrster, 1850) Materials examined: MALTA, Qammiegħ, 15. i.2006, 1 ♀, emerged from Coleophora acrisella Milliére (Lepidoptera: Coleophoridae), MZ, MSC; Ħad-Dingli, 27. x.2009, 1 ♂ & 1♀, emerged from Coleophora festivella Toll, MZ, MSC; Ħad-Dingli, 20. x.2011, 1 ♀, emerged from Coleophora fretella Zeller, which was collected on 13.xii.2010, MZ, MSC; Għajn Tuffieħa, 4. i.2015, 1 ♂, emerged from Coleophora fretella, which was collected on 7.xii.2014, MZ, MSC [parasitoids determined by Martin Schwarz; hosts determined by Michael Zerafa]. Notes: New record for the Maltese Islands. Gelis exareolatus has been recorded as a primary idiobiont ectoparasitoid of Coleophora spp. (especially on Juncus), Exoteleia dodecella (Linnaeus) as well as Kermania pistaciella Amsel (Schwarz & Shaw, 1999; Yu et al., 2012; Mehrnejad, 2002). It is also known as a pseudohyperparasitoid, parasitizing the cocoons of braconids attacking K. pistaciella (van Achterberg & Mehrnejad, 2002).Published as part of Mifsud, David, Farrugia, Lucia & Shaw, Mark R., 2019, Braconid and ichneumonid (Hymenoptera) parasitoid wasps of Lepidoptera from the Maltese Islands, pp. 47-60 in Zootaxa 4567 (1) on page 52, DOI: 10.11646/zootaxa.4567.1.3, http://zenodo.org/record/259303
A fault isolation method based on parity equations with application to a lathe-spindle system
No full textEfficient Scaling of Large Models: Principles in Optimization and Data Aspects
No full textDeep learning has advanced remarkably in recent decades. Yet, its theoretical foundations, particularly in the realm of large models, still lag behind. This thesis focuses on research that combines strong theoretical foundations with practical applications in efficiently scaling up large models.
In the first part of the thesis, we focus on the training dynamics of neural nets by covering the theory of overparametrized neural nets. We will briefly introduce the theory of Neural Tangent Kernel (NTK), and proceed with Hyperparameter Transfer, an important application of the Tensor Program framework. We cover some of the earliest papers that establish NTK as a research field, along with the limitations of NTK. Hyperparameter Transfer is a novel and efficient paradigm for hyperparameter tuning by providing the optimal strategy for scaling up models. We introduce the characterization of the training dynamics for deep neural nets and offer an efficient hyperparameter selection scheme where optimal hyperparameters selected by tuning on shallow nets also work for deep nets.
In the second part of the thesis, we focus on the data aspect of large model scaling. We will first introduce Skill-Mix, a novel and unique evaluation that sidesteps issues of traditional large language model (LLM) evaluations like data contamination and cramming for leaderboard. Skill-Mix randomly selects k language skills, then prompts the LLM to produce a concise text that demonstrates the chosen skills. The exponentially growing number of skill combinations provably prevent data contamination and can further reveal the novelty of successful answers by powerful LLMs. We then introduce ConceptMix, an extension of Skill-Mix to evaluate the capabilities of text-to-image models to combine k random selected visual concepts. Finally, we uncover the capabilities of LLMs to learn and generalize skill compositions given good responses from Skill-Mix. The results show that a few thousand of such data is enough to significantly improve the model performance in unseen skill combinations, beating models with much larger sizes. It suggests incorporating skill-rich synthetic text into training is an efficient way to scale up the data
Designing trust in highly automated virtual assistants: A taxonomy of levels of autonomy
No full textThis paper presents a guiding framework and a multi-level taxonomy of automation levels specially adapted to Virtual Assistants in the context of Human-Human-Interaction. This trust-based framework incorporates interaction phases, trust-affecting design principles and design techniques. It also introduces a taxonomy of levels of autonomy explaining each level from a trust perspective. To test the proposed Levels a survey was conducted addressing different contexts. Participants preferred to have total control of the system. Level 1 is the preferred option on average. Levels 2 and 3, account for 40.50% of the participants preference to be in control of the autonomous system. If we combine levels 1, 2, and 3; This presents an average of 68.75% of participants demanding the initiative. The neutral level (level 4) is preferred by 15.75% of the participants on average. On Levels where the initiative resides on the system (levels 5, 6, and 7), only 14.75% of participants would decentralise their decision. Based on the research findings, the author recommends designers to combine a holistic perspective on trust with contextual awareness, to be able to integrate the impact of contexts on interactions. Trust formation is a dynamic process that starts before a user’s first contact with the system and continues long thereafter. Furthermore, as autonomous systems continuously evolve, factors-affecting trust change during user interactions with the system and over time; thus, Human-Human-Interaction concepts need to be able to adapt. Future work will be dedicated to further understanding other areas such as reparation and accountability
Dynamic Fault Detection in Automotive Engines using Principal Component Analysis based Hybrid Radial Basis Function and Input Training Neural Network Models
No full textFor automobile engines, a novel fault detection (FD) system is created using PCA in this paper. To identify errors, two distinct neural networks are used. Radial basis function (RBF) is the first neural network, while input training neural network (ITNN) is the second. The mean value engine model (MVEM) with Matlab/Simulink is used to build the approach and evaluate its performance. The MVEM has been used to simulate three faults. The outputs of the MVEM are used as input data for the RBF. ITNN received the RBF output as an input and the output were the estimation of speed, pressure and temperature. According to the simulation results, faults with an amplitude variation of 10 – 20% were successfully identified under dynamic conditions across the whole working range. The corresponding detection thresholds are 0.36, 0.68, and 0.284 for speed, pressure and temperature respectively and any error exceeding the allowable threshold will be easily detected. Therefore, the simulation demonstrates that all three flaws may be easily identified and yields satisfactory results
- …
