We ask whether multilingual language models trained on unbalanced,
English-dominated corpora use English as an internal pivot language -- a
question of key importance for understanding how language models function and
the origins of linguistic bias. Focusing on the Llama-2 family of transformer
models, our study uses carefully constructed non-English prompts with a unique
correct single-token continuation. From layer to layer, transformers gradually
map an input embedding of the final prompt token to an output embedding from
which next-token probabilities are computed. Tracking intermediate embeddings
through their high-dimensional space reveals three distinct phases, whereby
intermediate embeddings (1) start far away from output token embeddings; (2)
already allow for decoding a semantically correct next token in the middle
layers, but give higher probability to its version in English than in the input
language; (3) finally move into an input-language-specific region of the
embedding space. We cast these results into a conceptual model where the three
phases operate in "input space", "concept space", and "output space",
respectively. Crucially, our evidence suggests that the abstract "concept
space" lies closer to English than to other languages, which may have important
consequences regarding the biases held by multilingual language models.

我们研究多语言语言模型在不平衡的以英语为主导的语料库上是否使用英语作为内部枢纽语言，以关键的重要性问题来理解语言模型的功能和语言偏见的起源。通过跟踪高维空间中的中间嵌入，我们发现了三个不同的阶段，这些阶段分别在 “输入空间”、“概念空间” 和 “输出空间” 中操作。关键的证据表明，抽象的 “概念空间” 更接近于英语而不是其他语言，这可能对多语言语言模型所持的偏见有重要影响。

多语言转换器的潜在语言：拉马在英文中起作用吗？

Do Llamas Work in English? On the Latent Language of Multilingual  Transformers

Deep neural networks that yield human interpretable decisions by
architectural design have lately become an increasingly popular alternative to
post hoc interpretation of traditional black-box models. Among these networks,
the arguably most widespread approach is so-called prototype learning, where
similarities to learned latent prototypes serve as the basis of classifying an
unseen data point. In this work, we point to an important shortcoming of such
approaches. Namely, there is a semantic gap between similarity in latent space
and similarity in input space, which can corrupt interpretability. We design
two experiments that exemplify this issue on the so-called ProtoPNet.
Specifically, we find that this network's interpretability mechanism can be led
astray by intentionally crafted or even JPEG compression artefacts, which can
produce incomprehensible decisions. We argue that practitioners ought to have
this shortcoming in mind when deploying prototype-based models in practice.

本文指出了原型学习中存在的一个重要缺陷：潜在空间中的相似性与输入空间中的相似性之间存在语义差距，这可能会损害可解释性。通过设计两个实验，我们在 ProtoPNet 上阐明了这个问题，并表明从业人员在实践中部署基于原型模型时应该考虑到这个缺陷。

潜在空间原型可解释性在深度神经网络中的局限性

This Looks Like That... Does it? Shortcomings of Latent Space Prototype  Interpretability in Deep Networks

We propose regularizing the empirical loss for semi-supervised learning by
acting on both the input (data) space, and the weight (parameter) space. We
show that the two are not equivalent, and in fact are complementary, one
affecting the minimality of the resulting representation, the other
insensitivity to nuisance variability. We propose a method to perform such
smoothing, which combines known input-space smoothing with a novel weight-space
smoothing, based on a min-max (adversarial) optimization. The resulting
Adversarial Block Coordinate Descent (ABCD) algorithm performs gradient ascent
with a small learning rate for a random subset of the weights, and standard
gradient descent on the remaining weights in the same mini-batch. It achieves
comparable performance to the state-of-the-art without resorting to heavy data
augmentation, using a relatively simple architecture.

本文提出了一种半监督学习的正则化方法，同时作用于输入空间和权重空间，通过组合已知输入空间平滑和基于 min-max 最优化的权重空间平滑，提出了 Adversarial Block Coordinate Descent 算法，使用相对简单的架构，无需重型数据增强，实现了与最先进技术可比的性能。