Arrow Research search

Author name cluster

Andrei Giurgiu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

2 papers
2 author rows

Possible papers

2

ICML Conference 2019 Conference Paper

Parameter-Efficient Transfer Learning for NLP

  • Neil Houlsby
  • Andrei Giurgiu
  • Stanislaw Jastrzebski
  • Bruna Morrone
  • Quentin de Laroussilhe
  • Andrea Gesmundo
  • Mona Attariyan
  • Sylvain Gelly

Fine-tuning large pretrained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter’s effectiveness, we transfer the recently proposed BERT Transformer model to $26$ diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within $0. 8%$ of the performance of full fine-tuning, adding only $3. 6%$ parameters per task. By contrast, fine-tuning trains $100%$ of the parameters per task.

Details

I&C Journal 2014 Journal Article

Computing in social networks

  • Andrei Giurgiu
  • Rachid Guerraoui
  • Kévin Huguenin
  • Anne-Marie Kermarrec

This paper defines the problem of Scalable Secure computing in a Social network: we call it the S 3 problem. In short, nodes, directly reflecting on associated users, need to compute a symmetric function f: V n → U of their inputs in a set of constant size, in a scalable and secure way. Scalability means that the spatial, computational and message complexity of the distributed computation does not grow too fast with the number of nodes n. Security encompasses (1) accuracy and (2) privacy: accuracy holds when the distance from the output to the ideal result is negligible with respect to the maximum distance between any two possible results; privacy is characterized by how the information disclosed by the computation helps faulty nodes infer inputs of non-faulty nodes, which we capture in our context by the very notion of probabilistic anonymity. We first prove that under mild regularity conditions the problem of computing an arbitrary function can be reduced to that of component-wise addition of vectors of integers. More specifically, if the function f is Lipschitz-continuous and the maximum distance between two possible results is Ω ( n ), any protocol that S 3 -computes component-wise addition of vectors of integers S 3 -computes f. We then present AG-S3, a protocol that S 3 -computes a class of aggregation functions, that is that can be expressed as a commutative monoid operation on U: f ( x 1, …, x n ) = x 1 ⊕ ⋯ ⊕ x n, assuming the number of faulty participants is at most n / log 2 n. We further prove that AG-S3 S 3 -computes component-wise addition of vectors of integers thus extending its application spectrum to regular functions. Key to our protocol is a dedicated overlay structure that enables secret sharing and distributed verifications which leverage the social aspect of the network: nodes care about their reputation and do not want to be tagged as misbehaving.

Details DOI
v2026.09.13