Joost van Amersfoort

Gemma Team, Morgane Riviere, Shreya Pathak et al. · deepmind

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.

Dustin Tran, Jeremiah Liu, Michael W. Dusenberry et al. · oxford

A recent trend in artificial intelligence is the use of pretrained models for language and vision tasks, which have achieved extraordinary performance but also puzzling failures. Probing these models' abilities in diverse ways is therefore critical to the field. In this paper, we explore the reliability of models, where we define a reliable model as one that not only achieves strong predictive performance but also performs well consistently over many decision-making tasks involving uncertainty (e.g., selective prediction, open set recognition), robust generalization (e.g., accuracy and proper scoring rules such as log-likelihood on in- and out-of-distribution datasets), and adaptation (e.g., active learning, few-shot uncertainty). We devise 10 types of tasks over 40 datasets in order to evaluate different aspects of reliability on both vision and language domains. To improve reliability, we developed ViT-Plex and T5-Plex, pretrained large model extensions for vision and language modalities, respectively. Plex greatly improves the state-of-the-art across reliability tasks, and simplifies the traditional protocol as it improves the out-of-the-box performance and does not require designing scores or tuning the model for each task. We demonstrate scaling effects over model sizes up to 1B parameters and pretraining dataset sizes up to 4B examples. We also demonstrate Plex's capabilities on challenging tasks including zero-shot open set recognition, active learning, and uncertainty in conversational language understanding.

Milad Alizadeh, Shyam A. Tailor, Luisa M Zintgraf et al.

Pruning neural networks at initialization would enable us to find sparse models that retain the accuracy of the original network while consuming fewer computational resources for training and inference. However, current methods are insufficient to enable this optimization and lead to a large degradation in model performance. In this paper, we identify a fundamental limitation in the formulation of current methods, namely that their saliency criteria look at a single step at the start of training without taking into account the trainability of the network. While pruning iteratively and gradually has been shown to improve pruning performance, explicit consideration of the training stage that will immediately follow pruning has so far been absent from the computation of the saliency criterion. To overcome the short-sightedness of existing methods, we propose Prospect Pruning (ProsPr), which uses meta-gradients through the first few steps of optimization to determine which weights to prune. ProsPr combines an estimate of the higher-order effects of pruning on the loss and the optimization trajectory to identify the trainable sub-network. Our method achieves state-of-the-art pruning performance on a variety of vision classification tasks, with less data and in a single shot compared to existing pruning-at-initialization methods.

Haiwen Huang, Joost van Amersfoort, Yarin Gal

Uncertainty estimation is a key component in any deployed machine learning system. One way to evaluate uncertainty estimation is using "out-of-distribution" (OoD) detection, that is, distinguishing between the training data distribution and an unseen different data distribution using uncertainty. In this work, we show that current feature density based uncertainty estimators cannot perform well consistently across different OoD detection settings. To solve this, we propose to decompose the learned representations and integrate the uncertainties estimated on them separately. Through experiments, we demonstrate that we can greatly improve the performance and the interpretability of the uncertainty estimation.

Andrew Jesson, Panagiotis Tigas, Joost van Amersfoort et al.

Estimating personalized treatment effects from high-dimensional observational data is essential in situations where experimental designs are infeasible, unethical, or expensive. Existing approaches rely on fitting deep models on outcomes observed for treated and control populations. However, when measuring individual outcomes is costly, as is the case of a tumor biopsy, a sample-efficient strategy for acquiring each result is required. Deep Bayesian active learning provides a framework for efficient data acquisition by selecting points with high uncertainty. However, existing methods bias training data acquisition towards regions of non-overlapping support between the treated and control populations. These are not sample-efficient because the treatment effect is not identifiable in such regions. We introduce causal, Bayesian acquisition functions grounded in information theory that bias data acquisition towards regions with overlapping support to maximize sample efficiency for learning personalized treatment effects. We demonstrate the performance of the proposed acquisition strategies on synthetic and semi-synthetic datasets IHDP and CMNIST and their extensions, which aim to simulate common dataset biases and pathologies.

Jishnu Mukhoti, Joost van Amersfoort, Philip H. S. Torr et al.

We extend Deep Deterministic Uncertainty (DDU), a method for uncertainty estimation using feature space densities, to semantic segmentation. DDU enables quantifying and disentangling epistemic and aleatoric uncertainty in a single forward pass through the model. We study the similarity of feature representations of pixels at different locations for the same class and conclude that it is feasible to apply DDU location independently, which leads to a significant reduction in memory consumption compared to pixel dependent DDU. Using the DeepLab-v3+ architecture on Pascal VOC 2012, we show that DDU improves upon MC Dropout and Deep Ensembles while being significantly faster to compute.

Lewis Smith, Joost van Amersfoort, Haiwen Huang et al.

ResNets constrained to be bi-Lipschitz, that is, approximately distance preserving, have been a crucial component of recently proposed techniques for deterministic uncertainty quantification in neural models. We show that theoretical justifications for recent regularisation schemes trying to enforce such a constraint suffer from a crucial flaw -- the theoretical link between the regularisation scheme used and bi-Lipschitzness is only valid under conditions which do not hold in practice, rendering existing theory of limited use, despite the strong empirical performance of these models. We provide a theoretical explanation for the effectiveness of these regularisation schemes using a frequency analysis perspective, showing that under mild conditions these schemes will enforce a lower Lipschitz bound on the low-frequency projection of images. We then provide empirical evidence supporting our theoretical claims, and perform further experiments which demonstrate that our broader conclusions appear to hold when some of the mathematical assumptions of our proof are relaxed, corresponding to the setup used in prior work. In addition, we present a simple constructive algorithm to search for counter examples to the distance preservation condition, and discuss possible implications of our theory for future model design.

Jishnu Mukhoti, Andreas Kirsch, Joost van Amersfoort et al.

Reliable uncertainty from deterministic single-forward pass models is sought after because conventional methods of uncertainty quantification are computationally expensive. We take two complex single-forward-pass uncertainty approaches, DUQ and SNGP, and examine whether they mainly rely on a well-regularized feature space. Crucially, without using their more complex methods for estimating uncertainty, a single softmax neural net with such a feature-space, achieved via residual connections and spectral normalization, *outperforms* DUQ and SNGP's epistemic uncertainty predictions using simple Gaussian Discriminant Analysis *post-training* as a separate feature-space density estimator -- without fine-tuning on OoD data, feature ensembling, or input pre-procressing. This conceptually simple *Deep Deterministic Uncertainty (DDU)* baseline can also be used to disentangle aleatoric and epistemic uncertainty and performs as well as Deep Ensembles, the state-of-the art for uncertainty prediction, on several OoD benchmarks (CIFAR-10/100 vs SVHN/Tiny-ImageNet, ImageNet vs ImageNet-O) as well as in active learning settings across different model architectures, yet is *computationally cheaper*.

Joost van Amersfoort, Lewis Smith, Andrew Jesson et al.

Inducing point Gaussian process approximations are often considered a gold standard in uncertainty estimation since they retain many of the properties of the exact GP and scale to large datasets. A major drawback is that they have difficulty scaling to high dimensional inputs. Deep Kernel Learning (DKL) promises a solution: a deep feature extractor transforms the inputs over which an inducing point Gaussian process is defined. However, DKL has been shown to provide unreliable uncertainty estimates in practice. We study why, and show that with no constraints, the DKL objective pushes "far-away" data points to be mapped to the same features as those of training-set points. With this insight we propose to constrain DKL's feature extractor to approximately preserve distances through a bi-Lipschitz constraint, resulting in a feature space favorable to DKL. We obtain a model, DUE, which demonstrates uncertainty quality outperforming previous DKL and other single forward pass uncertainty methods, while maintaining the speed and accuracy of standard neural networks.

Joost van Amersfoort, Milad Alizadeh, Sebastian Farquhar et al.

We introduce a method to speed up training by 2x and inference by 3x in deep neural networks using structured pruning applied before training. Unlike previous works on pruning before training which prune individual weights, our work develops a methodology to remove entire channels and hidden units with the explicit aim of speeding up training and inference. We introduce a compute-aware scoring mechanism which enables pruning in units of sensitivity per FLOP removed, allowing even greater speed ups. Our method is fast, easy to implement, and needs just one forward/backward pass on a single batch of data to complete pruning before training begins.

Joost van Amersfoort, Lewis Smith, Yee Whye Teh et al.

We propose a method for training a deterministic deep model that can find and reject out of distribution data points at test time with a single forward pass. Our approach, deterministic uncertainty quantification (DUQ), builds upon ideas of RBF networks. We scale training in these with a novel loss function and centroid updating scheme and match the accuracy of softmax models. By enforcing detectability of changes in the input using a gradient penalty, we are able to reliably detect out of distribution data. Our uncertainty quantification scales well to large datasets, and using a single model, we improve upon or match Deep Ensembles in out of distribution detection on notable difficult dataset pairs such as FashionMNIST vs. MNIST, and CIFAR-10 vs. SVHN.

Andreas Kirsch, Joost van Amersfoort, Yarin Gal

We develop BatchBALD, a tractable approximation to the mutual information between a batch of points and model parameters, which we use as an acquisition function to select multiple informative points jointly for the task of deep Bayesian active learning. BatchBALD is a greedy linear-time $1 - \frac{1}{e}$-approximate algorithm amenable to dynamic programming and efficient caching. We compare BatchBALD to the commonly used approach for batch data acquisition and find that the current approach acquires similar and redundant points, sometimes performing worse than randomly acquiring data. We finish by showing that, using BatchBALD to consider dependencies within an acquisition batch, we achieve new state of the art performance on standard benchmarks, providing substantial data efficiency improvements in batch acquisition.

Jo Schlemper, Jose Caballero, Andy Aitken et al.

In large scale systems, approximate nearest neighbour search is a crucial algorithm to enable efficient data retrievals. Recently, deep learning-based hashing algorithms have been proposed as a promising paradigm to enable data dependent schemes. Often their efficacy is only demonstrated on data sets with fixed, limited numbers of classes. In practical scenarios, those labels are not always available or one requires a method that can handle a higher input variability, as well as a higher granularity. To fulfil those requirements, we look at more flexible similarity measures. In this work, we present a novel, flexible, end-to-end trainable network for large-scale data hashing. Our method works by transforming the data distribution to behave as a uniform distribution on a product of spheres. The transformed data is subsequently hashed to a binary form in a way that maximises entropy of the output, (i.e. to fully utilise the available bit-rate capacity) while maintaining the correctness (i.e. close items hash to the same key in the map). We show that the method outperforms baseline approaches such as locality-sensitive hashing and product quantisation in the limited capacity regime.

Joost van Amersfoort, Wenzhe Shi, Alejandro Acosta et al.

Frame interpolation attempts to synthesise frames given one or more consecutive video frames. In recent years, deep learning approaches, and notably convolutional neural networks, have succeeded at tackling low- and high-level computer vision problems including frame interpolation. These techniques often tackle two problems, namely algorithm efficiency and reconstruction quality. In this paper, we present a multi-scale generative adversarial network for frame interpolation (\mbox{FIGAN}). To maximise the efficiency of our network, we propose a novel multi-scale residual estimation module where the predicted flow and synthesised frame are constructed in a coarse-to-fine fashion. To improve the quality of synthesised intermediate video frames, our network is jointly supervised at different levels with a perceptual loss function that consists of an adversarial and two content losses. We evaluate the proposed approach using a collection of 60fps videos from YouTube-8m. Our results improve the state-of-the-art accuracy and provide subjective visual quality comparable to the best performing interpolation method at x47 faster runtime.

Joost van Amersfoort, Anitha Kannan, Marc'Aurelio Ranzato et al.

In this work we propose a simple unsupervised approach for next frame prediction in video. Instead of directly predicting the pixels in a frame given past frames, we predict the transformations needed for generating the next frame in a sequence, given the transformations of the past frames. This leads to sharper results, while using a smaller prediction model. In order to enable a fair comparison between different video frame prediction models, we also propose a new evaluation protocol. We use generated frames as input to a classifier trained with ground truth sequences. This criterion guarantees that models scoring high are those producing sequences which preserve discriminative features, as opposed to merely penalizing any deviation, plausible or not, from the ground truth. Our proposed approach compares favourably against more sophisticated ones on the UCF-101 data set, while also being more efficient in terms of the number of parameters and computational cost.

Otto Fabius, Joost R. van Amersfoort

In this paper we propose a model that combines the strengths of RNNs and SGVB: the Variational Recurrent Auto-Encoder (VRAE). Such a model can be used for efficient, large scale unsupervised learning on time series data, mapping the time series data to a latent vector representation. The model is generative, such that data can be generated from samples of the latent space. An important contribution of this work is that the model can make use of unlabeled data in order to facilitate supervised training of RNNs by initialising the weights and network state.