Matthew Willetts

Fabian Falck, Christopher Williams, Dominic Danks et al. · oxford

U-Net architectures are ubiquitous in state-of-the-art deep learning, however their regularisation properties and relationship to wavelets are understudied. In this paper, we formulate a multi-resolution framework which identifies U-Nets as finite-dimensional truncations of models on an infinite-dimensional function space. We provide theoretical results which prove that average pooling corresponds to projection within the space of square-integrable functions and show that U-Nets with average pooling implicitly learn a Haar wavelet basis representation of the data. We then leverage our framework to identify state-of-the-art hierarchical VAEs (HVAEs), which have a U-Net architecture, as a type of two-step forward Euler discretisation of multi-resolution diffusion processes which flow from a point mass, introducing sampling instabilities. We also demonstrate that HVAEs learn a representation of time which allows for improved parameter efficiency through weight-sharing. We use this observation to achieve state-of-the-art HVAE performance with half the number of parameters of existing models, exploiting the properties of our continuous-time formulation.

Matthew Willetts

We show that when a dynamic-weight AMM rebalances by creating arbitrage opportunities, the per-step log loss is the KL divergence between successive weight vectors. The Fisher-Rao metric is therefore the natural Riemannian metric on the weight simplex. The loss-minimising interpolation under the leading-order expansion of this KL cost is SLERP (Spherical Linear Interpolation) in the Hellinger coordinates $η_i = \sqrt{w_i}$: a geodesic on the positive orthant of the unit sphere, traversed at constant speed. The SLERP midpoint equals the (AM+GM)/normalise heuristic of prior work (Willetts & Harrington, 2024), so the heuristic lies on the geodesic. This identity holds for any number of tokens and any magnitude of weight change; using this link, all dyadic points on the geodesic can be reached by recursive AM-GM bisection without trigonometric functions. SLERP's relative sub-optimality on the full KL cost is proportional to the squared magnitude of the overall weight change and to $1/f^2$, where $f$ is the number of interpolation steps. Under driftless GBM prices, the fractional value loss from each weight update is price-independent, and the cross term between weight and price changes telescopes, so the constant-price geometry carries over. LVR exposure introduces a finite optimal step count $f^*$, which lies in the perturbative regime where SLERP remains near-optimal.

Fabian Falck, Haoting Zhang, Matthew Willetts et al.

Work in deep clustering focuses on finding a single partition of data. However, high-dimensional data, such as images, typically feature multiple interesting characteristics one could cluster over. For example, images of objects against a background could be clustered over the shape of the object and separately by the colour of the background. In this paper, we introduce Multi-Facet Clustering Variational Autoencoders (MFCVAE), a novel class of variational autoencoders with a hierarchy of latent variables, each with a Mixture-of-Gaussians prior, that learns multiple clusterings simultaneously, and is trained fully unsupervised and end-to-end. MFCVAE uses a progressively-trained ladder architecture which leads to highly stable performance. We provide novel theoretical results for optimising the ELBO analytically with respect to the categorical variational posterior distribution, correcting earlier influential theoretical work. On image benchmarks, we demonstrate that our approach separates out and clusters over different aspects of the data in a disentangled manner. We also show other advantages of our model: the compositionality of its latent space and that it provides controlled generation of samples.

Matthew Willetts, Brooks Paige

In this paper, we investigate the algorithmic stability of unsupervised representation learning with deep generative models, as a function of repeated re-training on the same input data. Algorithms for learning low dimensional linear representations -- for example principal components analysis (PCA), or linear independent components analysis (ICA) -- come with guarantees that they will always reveal the same latent representations (perhaps up to an arbitrary rotation or permutation). Unfortunately, for non-linear representation learning, such as in a variational auto-encoder (VAE) model trained by stochastic gradient descent, we have no such guarantees. Recent work on identifiability in non-linear ICA have introduced a family of deep generative models that have identifiable latent representations, achieved by conditioning on side information (e.g. informative labels). We empirically evaluate the stability of these models under repeated re-estimation of parameters, and compare them to both standard VAEs and deep generative models which learn to cluster in their latent space. Surprisingly, we discover side information is not necessary for algorithmic stability: using standard quantitative measures of identifiability, we find deep generative models with latent clusterings are empirically identifiable to the same degree as models which rely on auxiliary labels. We relate these results to the possibility of identifiable non-linear ICA.

Alexander Camuto, Matthew Willetts

In this work we study Variational Autoencoders (VAEs) from the perspective of harmonic analysis. By viewing a VAE's latent space as a Gaussian Space, a variety of measure space, we derive a series of results that show that the encoder variance of a VAE controls the frequency content of the functions parameterised by the VAE encoder and decoder neural networks. In particular we demonstrate that larger encoder variances reduce the high frequency content of these functions. Our analysis allows us to show that increasing this variance effectively induces a soft Lipschitz constraint on the decoder network of a VAE, which is a core contributor to the adversarial robustness of VAEs. We further demonstrate that adding Gaussian noise to the input of a VAE allows us to more finely control the frequency content and the Lipschitz constant of the VAE encoder networks. To support our theoretical analysis we run experiments with VAEs with small fully-connected neural networks and with larger convolutional networks, demonstrating empirically that our theory holds for a variety of neural network architectures.

Ben Barrett, Alexander Camuto, Matthew Willetts et al.

We introduce an approach for training Variational Autoencoders (VAEs) that are certifiably robust to adversarial attack. Specifically, we first derive actionable bounds on the minimal size of an input perturbation required to change a VAE's reconstruction by more than an allowed amount, with these bounds depending on certain key parameters such as the Lipschitz constants of the encoder and decoder. We then show how these parameters can be controlled, thereby providing a mechanism to ensure \textit{a priori} that a VAE will attain a desired level of robustness. Moreover, we extend this to a complete practical approach for training such VAEs to ensure our criteria are met. Critically, our method allows one to specify a desired level of robustness \emph{upfront} and then train a VAE that is guaranteed to achieve this robustness. We further demonstrate that these Lipschitz--constrained VAEs are more robust to attack than standard VAEs in practice.

Alexander Camuto, Matthew Willetts, Umut Şimşekli et al.

We study the regularisation induced in neural networks by Gaussian noise injections (GNIs). Though such injections have been extensively studied when applied to data, there have been few studies on understanding the regularising effect they induce when applied to network activations. Here we derive the explicit regulariser of GNIs, obtained by marginalising out the injected noise, and show that it penalises functions with high-frequency components in the Fourier domain; particularly in layers closer to a neural network's output. We show analytically and empirically that such regularisation produces calibrated classifiers with large classification margins.

Alexander Camuto, Matthew Willetts, Stephen Roberts et al.

We make inroads into understanding the robustness of Variational Autoencoders (VAEs) to adversarial attacks and other input perturbations. While previous work has developed algorithmic approaches to attacking and defending VAEs, there remains a lack of formalization for what it means for a VAE to be robust. To address this, we develop a novel criterion for robustness in probabilistic models: $r$-robustness. We then use this to construct the first theoretical results for the robustness of VAEs, deriving margins in the input space for which we can provide guarantees about the resulting reconstruction. Informally, we are able to define a region within which any perturbation will produce a reconstruction that is similar to the original reconstruction. To support our analysis, we show that VAEs trained using disentangling methods not only score well under our robustness metrics, but that the reasons for this can be interpreted through our theoretical results.

Matthew Willetts, Xenia Miscouridou, Stephen Roberts et al.

Successfully training Variational Autoencoders (VAEs) with a hierarchy of discrete latent variables remains an area of active research. Vector-Quantised VAEs are a powerful approach to discrete VAEs, but naive hierarchical extensions can be unstable when training. Leveraging insights from classical methods of inference we introduce \textit{Relaxed-Responsibility Vector-Quantisation}, a novel way to parameterise discrete latent variables, a refinement of relaxed Vector-Quantisation that gives better performance and more stable training. This enables a novel approach to hierarchical discrete variational autoencoders with numerous layers of latent variables (here up to 32) that we train end-to-end. Within hierarchical probabilistic deep generative models with discrete latent variables trained end-to-end, we achieve state-of-the-art bits-per-dim results for various standard datasets. % Unlike discrete VAEs with a single layer of latent variables, we can produce samples by ancestral sampling: it is not essential to train a second autoregressive generative model over the learnt latent representations to then sample from and then decode. % Moreover, that latter approach in these deep hierarchical models would require thousands of forward passes to generate a single sample. Further, we observe different layers of our model become associated with different aspects of the data.

Alexander Camuto, Matthew Willetts, Brooks Paige et al.

Separating high-dimensional data like images into independent latent factors, i.e independent component analysis (ICA), remains an open research problem. As we show, existing probabilistic deep generative models (DGMs), which are tailor-made for image data, underperform on non-linear ICA tasks. To address this, we propose a DGM which combines bijective feature maps with a linear ICA model to learn interpretable latent structures for high-dimensional data. Given the complexities of jointly training such a hybrid model, we introduce novel theory that constrains linear ICA to lie close to the manifold of orthogonal rectangular matrices, the Stiefel manifold. By doing so we create models that converge quickly, are easy to train, and achieve better unsupervised latent factor discovery than flow-based models, linear ICA, and Variational Autoencoders on images.

Miguel Morin, Matthew Willetts

We show that the stochasticity in training ResNets for image classification on GPUs in TensorFlow is dominated by the non-determinism from GPUs, rather than by the initialisation of the weights and biases of the network or by the sequence of minibatches given. The standard deviation of test set accuracy is 0.02 with fixed seeds, compared to 0.027 with different seeds---nearly 74\% of the standard deviation of a ResNet model is non-deterministic. For test set loss the ratio of standard deviations is more than 80\%. These results call for more robust evaluation strategies of deep learning models, as a significant amount of the variation in results across runs can arise simply from GPU randomness.

Matthew Willetts, Alexander Camuto, Stephen Roberts et al.

We develop a new method for regularising neural networks. We learn a probability distribution over the activations of all layers of the model and then insert imputed values into the network during training. We obtain a posterior for an arbitrary subset of activations conditioned on the remainder. This is a generalisation of data augmentation to the hidden layers of a network, and a form of data-aware dropout. We demonstrate that our training method leads to higher test accuracy and lower test-set cross-entropy for neural networks trained on CIFAR-10 and SVHN compared to standard regularisation baselines: our approach leads to networks with better calibrated uncertainty over the class posteriors all the while delivering greater test-set accuracy.

Matthew Willetts, Stephen Roberts, Chris Holmes

In clustering we normally output one cluster variable for each datapoint. However it is not necessarily the case that there is only one way to partition a given dataset into cluster components. For example, one could cluster objects by their colour, or by their type. Different attributes form a hierarchy, and we could wish to cluster in any of them. By disentangling the learnt latent representations of some dataset into different layers for different attributes we can then cluster in those latent spaces. We call this "disentangled clustering". Extending Variational Ladder Autoencoders (Zhao et al., 2017), we propose a clustering algorithm, VLAC, that outperforms a Gaussian Mixture DGM in cluster accuracy over digit identity on the test set of SVHN. We also demonstrate learning clusters jointly over numerous layers of the hierarchy of latent variables for the data, and show component-wise generation from this hierarchical model.

Matthew Willetts, Alexander Camuto, Tom Rainforth et al.

Variational autoencoders (VAEs) have recently been shown to be vulnerable to adversarial attacks, wherein they are fooled into reconstructing a chosen target image. However, how to defend against such attacks remains an open problem. We make significant advances in addressing this issue by introducing methods for producing adversarially robust VAEs. Namely, we first demonstrate that methods proposed to obtain disentangled latent representations produce VAEs that are more robust to these attacks. However, this robustness comes at the cost of reducing the quality of the reconstructions. We ameliorate this by applying disentangling methods to hierarchical VAEs. The resulting models produce high-fidelity autoencoders that are also adversarially robust. We confirm their capabilities on several different datasets and with current state-of-the-art VAE adversarial attacks, and also show that they increase the robustness of downstream tasks to attack.

Matthew Willetts, Stephen J Roberts, Christopher C Holmes

In semi-supervised learning for classification, it is assumed that every ground truth class of data is present in the small labelled dataset. Many real-world sparsely-labelled datasets are plausibly not of this type. It could easily be the case that some classes of data are found only in the unlabelled dataset -- perhaps the labelling process was biased -- so we do not have any labelled examples to train on for some classes. We call this learning regime $\textit{semi-unsupervised learning}$, an extreme case of semi-supervised learning, where some classes have no labelled exemplars in the training set. First, we outline the pitfalls associated with trying to apply deep generative model (DGM)-based semi-supervised learning algorithms to datasets of this type. We then show how a combination of clustering and semi-supervised learning, using DGMs, can be brought to bear on this problem. We study several different datasets, showing how one can still learn effectively when half of the ground truth classes are entirely unlabelled and the other half are sparsely labelled.

Matthew Willetts, Aiden Doherty, Stephen Roberts et al.

We introduce 'semi-unsupervised learning', a problem regime related to transfer learning and zero-shot learning where, in the training data, some classes are sparsely labelled and others entirely unlabelled. Models able to learn from training data of this type are potentially of great use as many real-world datasets are like this. Here we demonstrate a new deep generative model for classification in this regime. Our model, a Gaussian mixture deep generative model, demonstrates superior semi-unsupervised classification performance on MNIST to model M2 from Kingma and Welling (2014). We apply the model to human accelerometer data, performing activity classification and structure discovery on windows of time series data.