DenseFormer: Enhancing Information Flow in Transformers via Depth Weighted Averaging

The transformer architecture from Vaswani et al. (2017) is now ubiquitous across application domains, from natural language processing to speech processing and image understanding. We propose DenseFormer, a simple modification to the standard architecture that improves the perplexity of the model without increasing its size -- adding a few thousand parameters for large-scale models in the 100B parameters range. Our approach relies on an additional averaging step after each transformer block, which computes a weighted average of current and past representations -- we refer to this operation as Depth-Weighted-Average (DWA). The learned DWA weights exhibit coherent patterns of information flow, revealing the strong and structured reuse of activations from distant layers. Experiments demonstrate that DenseFormer is more data efficient, reaching the same perplexity of much deeper transformer models, and that for the same perplexity, these new models outperform transformer baselines in terms of memory efficiency and inference time.

DeepEMD: A Transformer-based Fast Estimation of the Earth Mover's Distance

The Earth Mover's Distance (EMD) is the measure of choice between point clouds. However the computational cost to compute it makes it prohibitive as a training loss, and the standard approach is to use a surrogate such as the Chamfer distance. We propose an attention-based model to compute an accurate approximation of the EMD that can be used as a training loss for generative models. To get the necessary accurate estimation of the gradients we train our model to explicitly compute the matching between point clouds instead of EMD itself. We cast this new objective as the estimation of an attention matrix that approximates the ground truth matching matrix. Experiments show that this model provides an accurate estimate of the EMD and its gradient with a wall clock speed-up of more than two orders of magnitude with respect to the exact Hungarian matching algorithm and one order of magnitude with respect to the standard approximate Sinkhorn algorithm, allowing in particular to train a point cloud VAE with the EMD itself. Extensive evaluation show the remarkable behaviour of this model when operating out-of-distribution, a key requirement for a distance surrogate. Finally, the model generalizes very well to point clouds during inference several times larger than during training.

SequeL: A Continual Learning Library in PyTorch and JAX

Continual Learning is an important and challenging problem in machine learning, where models must adapt to a continuous stream of new data without forgetting previously acquired knowledge. While existing frameworks are built on PyTorch, the rising popularity of JAX might lead to divergent codebases, ultimately hindering reproducibility and progress. To address this problem, we introduce SequeL, a flexible and extensible library for Continual Learning that supports both PyTorch and JAX frameworks. SequeL provides a unified interface for a wide range of Continual Learning algorithms, including regularization-based approaches, replay-based approaches, and hybrid approaches. The library is designed towards modularity and simplicity, making the API suitable for both researchers and practitioners. We release SequeL\footnote{\url{https://github.com/nik-dim/sequel}} as an open-source library, enabling researchers and developers to easily experiment and extend the library for their own purposes.

Flatten the Curve: Efficiently Training Low-Curvature Neural Networks

The highly non-linear nature of deep neural networks causes them to be susceptible to adversarial examples and have unstable gradients which hinders interpretability. However, existing methods to solve these issues, such as adversarial training, are expensive and often sacrifice predictive accuracy. In this work, we consider curvature, which is a mathematical quantity which encodes the degree of non-linearity. Using this, we demonstrate low-curvature neural networks (LCNNs) that obtain drastically lower curvature than standard models while exhibiting similar predictive performance, which leads to improved robustness and stable gradients, with only a marginally increased training time. To achieve this, we minimize a data-independent upper bound on the curvature of a neural network, which decomposes overall curvature in terms of curvatures and slopes of its constituent layers. To efficiently minimize this bound, we introduce two novel architectural components: first, a non-linearity called centered-softplus that is a stable variant of the softplus non-linearity, and second, a Lipschitz-constrained batch normalization layer. Our experiments show that LCNNs have lower curvature, more stable gradients and increased off-the-shelf adversarial robustness when compared to their standard high-curvature counterparts, all without affecting predictive performance. Our approach is easy to use and can be readily incorporated into existing neural network models.

HyperMixer: An MLP-based Green AI Alternative to Transformers

Transformer-based architectures are the model of choice for natural language understanding, but they come at a significant cost, as they have quadratic complexity in the input length and can be difficult to tune. In the pursuit of Green AI, we investigate simple MLP-based architectures. We find that existing architectures such as MLPMixer, which achieves token mixing through a static MLP applied to each feature independently, are too detached from the inductive biases required for natural language understanding. In this paper, we propose a simple variant, HyperMixer, which forms the token mixing MLP dynamically using hypernetworks. Empirically, we demonstrate that our model performs better than alternative MLP-based models, and on par with Transformers. In contrast to Transformers, HyperMixer achieves these results at substantially lower costs in terms of processing time, training data, and hyperparameter tuning.

Taming GANs with Lookahead

Generative Adversarial Networks are notoriously challenging to train. The underlying minimax optimization is highly susceptible to the variance of the stochastic gradient and the rotational component of the associated game vector field. We empirically demonstrate the effectiveness of the Lookahead meta-optimization method for optimizing games, originally proposed for standard minimization. The backtracking step of Lookahead naturally handles the rotational game dynamics, which in turn enables the gradient ascent descent method to converge on challenging toy games often analyzed in the literature. Moreover, it implicitly handles high variance without using large mini-batches, known to be essential for reaching state of the art performance. Experimental results on MNIST, SVHN, and CIFAR-10, demonstrate a clear advantage of combining Lookahead with Adam or extragradient, in terms of performance, memory footprint, and improved stability. Using 30-fold fewer parameters and 16-fold smaller minibatches we outperform the reported performance of the class-dependent BigGAN on CIFAR-10 by obtaining FID of $13.65$ \emph{without} using the class labels, bringing state-of-the-art GAN training within reach of common computational resources.

Gradient Alignment in Deep Neural Networks

One cornerstone of interpretable deep learning is the high degree of visual alignment that input-gradients, i.e.,the gradients of the output w.r.t. inputs, exhibit with the input data. This alignment is assumed to arise as a result of the model's generalization, justifying its use for interpretability. However, recent work has shown that it is possible to 'fool' models into having arbitrary gradients while achieving good generalization, thus falsifying the assumption above. This leaves an open question: if not generalization, what causes input-gradients to align with input data? In this work, we first show that it is simple to 'fool' input-gradients using the shift-invariance property of softmax, and that gradient structure is unrelated to model generalization. Second, we re-interpret the logits of standard classifiers as unnormalized log-densities of the data distribution, and find that we can improve this gradient alignment via a generative modelling objective called score-matching.To show this, we derive a novel approximation to the score-matching objective that eliminates the need for expensive Hessian computations, which may be of independent interest.Our experiments help us identify one factor that causes input-gradient alignment in models, that being the approximate generative modelling behaviour of the normalized logit distributions.

Fair Latency-Aware Metric for real-time video segmentation networks

As supervised semantic segmentation is reaching satisfying results, many recent papers focused on making segmentation network architectures faster, smaller and more efficient. In particular, studies often aim to reach the stage to which they can claim to be "real-time". Achieving this goal is especially relevant in the context of real-time video operations for autonomous vehicles and robots, or medical imaging during surgery. The common metric used for assessing these methods is so far the same as the ones used for image segmentation without time constraint: mean Intersection over Union (mIoU). In this paper, we argue that this metric is not relevant enough for real-time video as it does not take into account the processing time (latency) of the network. We propose a similar but more relevant metric called FLAME for video-segmentation networks, that compares the output segmentation of the network with the ground truth segmentation of the current video frame at the time when the network finishes the processing. We perform experiments to compare a few networks using this metric and propose a simple addition to network training to enhance results according to that metric.

Full-Jacobian Representation of Neural Networks

Non-linear functions such as neural networks can be locally approximated by affine planes. Recent works make use of input-Jacobians, which describe the normal to these planes. In this paper, we introduce full-Jacobians, which includes this normal along with an additional intercept term called the bias-Jacobians, that together completely describe local planes. For ReLU neural networks, bias-Jacobians correspond to sums of gradients of outputs w.r.t. intermediate layer activations. We first use these full-Jacobians for distillation by aligning gradients of their intermediate representations. Next, we regularize bias-Jacobians alone to improve generalization. Finally, we show that full-Jacobian maps can be viewed as saliency maps. Experimental results show improved distillation on small data-sets, improved generalization for neural network training, and sharper saliency maps.

Practical Deep Stereo (PDS): Toward applications-friendly deep stereo matching

End-to-end deep-learning networks recently demonstrated extremely good perfor- mance for stereo matching. However, existing networks are difficult to use for practical applications since (1) they are memory-hungry and unable to process even modest-size images, (2) they have to be trained for a given disparity range. The Practical Deep Stereo (PDS) network that we propose addresses both issues: First, its architecture relies on novel bottleneck modules that drastically reduce the memory footprint in inference, and additional design choices allow to handle greater image size during training. This results in a model that leverages large image context to resolve matching ambiguities. Second, a novel sub-pixel cross- entropy loss combined with a MAP estimator make this network less sensitive to ambiguous matches, and applicable to any disparity range without re-training. We compare PDS to state-of-the-art methods published over the recent months, and demonstrate its superior performance on FlyingThings3D and KITTI sets.