Alexander Mordvintsev
Authored Publications
Sort By
Transformers learn in-context by gradient descent
João Sacramento
International Conference on Machine Learning (2023), pp. 35151-35174
Preview abstract
Transformers have become the state-of-the-art neural network architecture across numerous
domains of machine learning. This is partly due to their celebrated ability to transfer and
to learn in-context based on a few examples. Nevertheless, the mechanism of why and
how Transformers become in-context learners is not well understood and remains mostly an
intuition. Here, we argue that training Transformers on auto-regressive tasks can be closely
related to well-known gradient-based meta-learning formulations. We do so by providing
a simple construction that shows the equivalence of data transformations induced by 1) a
single linear self-attention layer and by 2) gradient-descent on a regression loss. Motivated by
that construction, we show empirically that when training self-attention only Transformers
on simple regression tasks either the models learned by GD and Transformers show great
similarity or, remarkably, the solutions found by gradient descent converge in weight space to
our construction. This allows us, at least on our simple regression tasks, to mechanistically
understand the inner workings of Transformers that enables in-context learning within.
Finally, we discuss intriguing parallels to a mechanism identified as crucial for in-context
learning termed induction-head (Olsson et al., 2022) and show how it could be generalized
by in-context learning by gradient descent within Transformers.
View details
Preview abstract
Training an end-to-end differentiable, self-organising cellular automata model of morphogenesis, able to both grow and regenerate specific patterns.
View details
Preview abstract
We present a Message Passing based Learning Protocol (MPLP) for artificial neural networks. With this protocol, every synapse (weights and biases), and activation is considered an independent agent, responsible for ingesting incoming messages, updating their own states, and outputting n-dimensional messages for their neighbours. We show how this protocol can be used instead of a traditional gradient-based approach for traditional feed-forward neural networks, and present a framework capable of generalizing neural networks to explore more flexible architectures. We meta-learn the MPLP through end-to-end gradient-based meta-optimisation. Finally, we discuss where the strengths of MPLP lay, and where we foresee possible limitations.
View details
The Building Blocks of Interpretability
Christopher Olah
Arvind Satyanarayan
Ian Johnson
Shan Carter
Ludwig Schubert
Katherine Ye
Distill (2018)
Preview abstract
Interpretability techniques are normally studied in isolation. We explore the powerful interfaces that arise when you combine them -- and the rich structure of this combinatorial space.
View details
Associative Domain Adaptation
Philip Haeusser
Thomas Frerix
Daniel Cremers
International Conference on Computer Vision (ICCV), IEEE (2017) (to appear)
Preview abstract
We propose associative domain adaptation, a novel technique for end-to-end domain adaptation with neural networks, the task of inferring class labels for an unlabeled target domain based on the statistical properties of a labeled source domain.
Our training scheme follows the paradigm that in order to effectively derive class labels for the target domain, a network should produce statistically domain invariant embeddings, while minimizing the classification error on the labeled source domain.
We accomplish this by reinforcing associations between source and target data directly in embedding space.
Our method can easily be added to any existing classification network with no structural and almost no computational overhead.
We demonstrate the effectiveness of our approach on various benchmarks and achieve state-of-the-art results across the board with a generic convolutional neural network architecture not specifically tuned to the respective tasks. Finally, we show that the proposed association loss produces embeddings that are more effective for domain adaptation compared to methods employing maximum mean discrepancy as a similarity measure in embedding space.
View details
Learning by Association - A versatile semi-supervised training method for neural networks
Philip Haeusser
Daniel Cremers
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2017)
Preview abstract
In many real-world scenarios, labeled data for a specific training task is costly to obtain. Semi-supervised methods make use of abundantly available unlabeled data and a smaller number of labeled examples.
We propose a new framework for semi-supervised training of deep neural networks that is inspired by learning in humans. "Associations" are made from embeddings of labeled samples to those of unlabeled ones and back. The optimization schedule encourages correct association cycles that end up at the same class where the association was started from and penalizes wrong associations that end at a different class.
The implementation is easy to use and can be added to any existing end-to-end training setup.
We demonstrate the capabilities of our approach on several data sets and show that it can improve performance on classification tasks up to state of the art, making use of additionally available unlabeled data. We also show how to apply this to the task of domain adaptation, surpassing current state-of-the-art results.
View details
Preview abstract
Neural network feature visualization is a powerful technique. It can answer questions about what a network — or parts of a network — are looking for by generating idealized examples of what the network is trying to find.
Over the last few years, the field has made great strides in feature visualization. Actually getting it to work, however, involves a number of details. In this article, we examine the major issues and explore common approaches to solving them.
We find that remarkably simple methods can produce state-of-the-art visualizations — and that, surprisingly, these visualizations are often limited by optimization problems that can be solved with standard techniques.
View details
