5,374 research outputs found
ResiDual Transformer Alignment with Spectral Decomposition
No full textWhen examined through the lens of their residual streams, a puzzling property emerges in transformer networks: residual contributions (e.g., attention heads) sometimes specialize in specific tasks or input attributes. In this paper, we analyze this phenomenon in vision transformers, focusing on the spectral geometry of residuals, and explore its implications for modality alignment in vision-language models. First, we link it to the intrinsically low-dimensional structure of visual head representations, zooming into their principal components and showing that they encode specialized roles across a wide variety of input data distributions. Then, we analyze the effect of head specialization in multimodal models, focusing on how improved alignment between text and specialized heads impacts zero-shot classification performance. This specialization-performance link consistently holds across diverse pre-training data, network sizes, and objectives, demonstrating a powerful new mechanism for boosting zero-shot classification through targeted alignment. Ultimately, we translate these insights into actionable terms by introducing ResiDual, a technique for spectral alignment of the residual stream. Much like panning for gold, it lets the noise from irrelevant unit principal components (i.e., attributes) wash away to amplify task-relevant ones. Remarkably, this dual perspective on modality alignment yields fine-tuning level performance on different data distributions while modelling an extremely interpretable and parameter-efficient transformation, as we extensively show on 70 pre-trained network-dataset combinations (7 models, 10 datasets)
Francesco Scorza Barcellona o della passione agiografica
No full textL'autrice traccia, sul filo dei ricordi, il profilo scientifico ed umano di Francesco Scorza Barcellona ed introduce gli studi raccolti nel volume.The author traces, on the thread of memories, the scientific and human profile of Francesco Scorza Barcelona and introduces the studies collected in the book
Delle lodi di don Francesco Medici de' principi di Toscana : orazione /
No full textEngraved t.p. vignette of Medici arms. On recto of 2nd leaf is Callot's full-page engraved port. of Francesco de' Medici, son of Grand Duke Ferdinand I. It is in Lieure's 1st state. The port. also appeared the same year in Alessandro Adimari's Esequie dell' ill.mo & ecc.mo principe don Francesco Medici (Florence : Gio. Donato & Bernardino Giunti); see Lieure. Large woodcut Giunti device on p. [31].Lieure, J. Jacques Callot,Mode of access: Internet.At head of front pastedown is bookplate of Francesco Riccardi de Vernaccia (Florence, ca. 1780; see Gelli, p. 387). Below it is the label of Horatius (Orazio) Landau, with stamped shelfmark 53021. At foot is the bookplate of U. Manganelli, signed with the initials SER. A dedicatory inscription at foot of t.p. has been scratched out.Binding: modern marbled paper, backed in green vellum. Date, author & title writte on spine.Port. trimmed into the image at right margin and foot
Frankenstein
No full text@inproceedings{orsini2015graph,
title={Graph invariant kernels},
author={Orsini, Francesco and Frasconi, Paolo and De Raedt, Luc},
booktitle={IJCAI Proceedings-International Joint Conference on Artificial Intelligence. IJCAI},
year={2015}
Relative representations enable zero-shot latent space communication
Full text linkNeural networks embed the geometric structure of a data manifold lying in a
high-dimensional space into latent representations. Ideally, the distribution
of the data points in the latent space should depend only on the task, the
data, the loss, and other architecture-specific constraints. However, factors
such as the random weights initialization, training hyperparameters, or other
sources of randomness in the training phase may induce incoherent latent spaces
that hinder any form of reuse. Nevertheless, we empirically observe that, under
the same data and modeling choices, the angles between the encodings within
distinct latent spaces do not change. In this work, we propose the latent
similarity between each sample and a fixed set of anchors as an alternative
data representation, demonstrating that it can enforce the desired invariances
without any additional training. We show how neural architectures can leverage
these relative representations to guarantee, in practice, invariance to latent
isometries and rescalings, effectively enabling latent space communication:
from zero-shot model stitching to latent space comparison between diverse
settings. We extensively validate the generalization capability of our approach
on different datasets, spanning various modalities (images, text, graphs),
tasks (e.g., classification, reconstruction) and architectures (e.g., CNNs,
GCNs, transformers).Comment: ICLR 2023 notable top 5%, 26 pages, 11 figures, 18 table
Didattica e ricerca. L’esperienza del Laboratorio aragonese e sforzesco
Full text linkIn the workshop on Diplomatic correspondences in the Italian Renaissance, directed in January-May 2018 by Francesco Storti and Francesco Senatore in the Istituto italiano per gli studi filosofici of Naples, 12 undergraduate and graduate students learned to read and trascribe italian diplomatic dispatches of the 15th cent., producing a specimen of scientific edition. In the opinion of the author, such activities and similar ones are very useful for future school teachers and scholars and they could be introduced in the university and pHD courses of history. Students of humanistic sciences should have concrete contacts with the historical research on primary sources. On the other hand, editing of historical sources could become a marginal activity because it is not enough appreciated in the current research evaluation processes.
Keywords: Didattica, edizione documentaria, diplomazia rinascimentale; education, documentary edition, renaissance diplomac
LA MADONNA ESTERHAZY DI RAFFAELLO
No full textLA MADONNA ESTERHÁZY DI RAFFAELLO A PALAZZO MARINO, di Francesco Murano,
Fotografie di Sergio Oriani
Il curatore della mostra, Stefano Zuffi, ha voluto inquadrare criticamente l’opera di Raffaello
proveniente dal Museo di Budapest, ponendola tra la Vergine delle Rocce di Francesco Melzi,
proveniente dall’Istituto delle Suore Orsoline di Milano e la Madonna della Rosa di Giovanni
Antonio Boltraffio, proveniente dal Museo Poldi Pezzoli della stessa città. L’architetto Corrado
Anselmi ha progettato un affascinante allestimento, concependolo come un percorso che si
sviluppa attraverso tre soste utili all’ammirazione delle opere. L’autore Francesco Murano,
spiega con ricchezza di particolari i vincoli del progetto illuminotecnico e le scelte luministiche.Once again, for the Christmas season, the Municipally of Milan decided to display an important work in the frame of the Palazzo Marino’s Alessi Hall. For this year 2014, the choice has been to propose the Estheràzy Madonna by Raphael, from the Museum of Budapest.
The exhibition curator, Stefano Zuffi, wanted to critically frame to work of Raphael by placing it between the Vegine delle Rocce (Virgin of the Rocks) by Francesco Melzi, coming from the Milan’s Istituto delle Suore Orsoline, and the Madonna della Rosa by Giovanni Boltraffio, from the Milanese Museo Poldi Pezzoli.
The Architect Corrado Anselmi designed a fascinating display, convinced as a path that develops through three stops that are useful for the enjoyment of the works.
The author Francesco Murano, explains with many details the lighting design project
- …
