The regularity or implicit bias in neural network optimization has been typically studied via the parameter norms or the landscape curvature, often overlooking the trajectory leading to these parameters. However, properties of the trajectory --- particularly its directionality --- capture critical aspects of how gradient descent navigates the landscape to converge to a solution. In this work, we introduce the notion of a Trajectory Map and derive natural complexity measures that highlight the directional characteristics of optimization trajectories. Our comprehensive analysis across vision and language modeling tasks reveals that (a) the trajectory's directionality at the macro-level saturates by the initial phase of training, wherein weight decay and momentum play a crucial but understated role; and (b) in subsequent training, trajectory directionality manifests in micro-level behaviors, such as oscillations, for which we also provide a theoretical analysis. This implies that neural optimization trajectories have, overall, a more linear form than zig-zaggy, as evident by high directional similarity, especially towards the end. To further hone this point, we show that when the trajectory direction gathers such an inertia, optimization proceeds largely unaltered even if the network is severely decapacitated (by freezing >99% of the parameters), --- thereby demonstrating the potential for significant computational and resource savings without compromising performance.

A simple design recipe for deep Transformers is to compose identical building blocks. But standard transformer blocks are far from simple, interweaving attention and MLP sub-blocks with skip connections \& normalisation layers in precise arrangements. This complexity leads to brittle architectures, where seemingly minor changes can significantly reduce training speed, or render models untrainable. In this work, we ask to what extent the standard transformer block can be simplified? Combining signal propagation theory and empirical observations, we motivate modifications that allow many block components to be removed with no loss of training speed, including skip connections, projection or value parameters, sequential sub-blocks and normalisation layers. In experiments on both autoregressive decoder-only and BERT encoder-only models, our simplified transformers match the per-iteration training speed and performance of standard transformers, while enjoying 16\% faster training throughput, and using 15\% fewer parameters.

Graph neural networks based on iterative one-hop message passing have been shown to struggle in harnessing the information from distant nodes effectively. Conversely, graph transformers allow each node to attend to all other nodes directly, but lack graph inductive bias and have to rely on ad-hoc positional encoding. In this paper, we propose a new architecture to reconcile these challenges. Our approach stems from the recent breakthroughs in long-range modeling provided by deep state-space models: for a given target node, our model aggregates other nodes by their shortest distances to the target and uses a linear RNN to encode the sequence of hop representations. The linear RNN is parameterized in a particular diagonal form for stable long-range signal propagation and is theoretically expressive enough to encode the neighborhood hierarchy. With no need for positional encoding, we empirically show that the performance of our model is comparable to or better than that of state-of-the-art graph transformers on various benchmarks, with a significantly reduced computational cost. Our code is open-source at https://github.com/skeletondyh/GRED.

Outlier Features (OFs) are neurons whose activation magnitudes significantly exceed the average over a neural network's (NN) width. They are well known to emerge during standard transformer training and have the undesirable effect of hindering quantisation in afflicted models. Despite their practical importance, little is known behind why OFs emerge during training, nor how one can minimise them. Our work focuses on the above questions, first identifying several quantitative metrics, such as the kurtosis over neuron activation norms, to measure OFs. With these metrics, we study how architectural and optimisation choices influence OFs, and provide practical insights to minimise OFs during training. As highlights, we introduce a novel unnormalised transformer block, the Outlier Protected block, and present a previously unknown benefit of non-diagonal preconditioning optimisers, finding both approaches to significantly reduce OFs and improve quantisation without compromising convergence speed, at scales of up to 7B parameters. Notably, our combination of OP block and non-diagonal preconditioner (SOAP) achieves 14.87 weight-and-activation int8 perplexity (from 14.71 in standard precision), compared to 63.4 int8 perplexity (from 16.00) with a default OF-prone combination of Pre-Norm model and Adam, when quantising OPT-125m models post-training.

In deep learning theory, the covariance matrix of the representations serves as a proxy to examine the network’s trainability. Motivated by the success of Transform- ers, we study the covariance matrix of a modified Softmax-based attention model with skip connections in the proportional limit of infinite-depth-and-width. We show that at initialization the limiting distribution can be described by a stochastic differential equation (SDE) indexed by the depth-to-width ratio. To achieve a well-defined stochastic limit, the Transformer’s attention mechanism is modified by centering the Softmax output at identity, and scaling the Softmax logits by a width-dependent temperature parameter. We examine the stability of the network through the corresponding SDE, showing how the scale of both the drift and diffu- sion can be elegantly controlled with the aid of residual connections. The existence of a stable SDE implies that the covariance structure is well-behaved, even for very large depth and width, thus preventing the notorious issues of rank degeneracy in deep attention models. Finally, we show, through simulations, that the SDE provides a surprisingly good description of the corresponding finite-size model. We coin the name shaped Transformer for these architectural modifications.

Avoiding feature collapse, when a Neural Network (NN) encoder maps all inputs to a constant vector, is a shared implicit desideratum of various methodological advances in self-supervised learning (SSL). To that end, whitened features have been proposed as an explicit objective to ensure uncollapsed features \cite{zbontar2021barlow,ermolov2021whitening,hua2021feature,bardes2022vicreg}. We identify power law behaviour in eigenvalue decay, parameterised by exponent $\beta{\geq}0$, as a spectrum that bridges between the collapsed & whitened feature extremes. We provide theoretical & empirical evidence highlighting the factors in SSL, like projection layers & regularisation strength, that influence eigenvalue decay rate, & demonstrate that the degree of feature whitening affects generalisation, particularly in label scarce regimes. We use our insights to motivate a novel method, PMP (PostMan-Pat), which efficiently post-processes a pretrained encoder to enforce eigenvalue decay rate with power law exponent $\beta$, & find that PostMan-Pat delivers improved label efficiency and transferability across a range of SSL methods and encoder architectures.

Deep ResNet architectures have achieved state of the art performance on many tasks. While they solve the problem of gradient vanishing, they might suffer from gradient exploding as the depth becomes large (Yang et al. 2017). Moreover, recent results have shown that ResNet might lose expressivity as the depth goes to infinity (Yang et al. 2017, Hayou et al. 2019). To resolve these issues, we introduce a new class of ResNet architectures, calledStable ResNet, that have the property of stabilizing the gradient while ensuring expressivity in the infinite depth limit.

We explore the link between deep ensembles and Gaussian processes (GPs) through the lens of the Neural Tangent Kernel (NTK): a recent development in understanding the training dynamics of wide neural networks (NNs). Previous work has shown that even in the infinite width limit, when NNs become GPs, there is no GP posterior interpretation to a deep ensemble trained with squared error loss. We introduce a simple modification to standard deep ensembles training, through addition of a computationally-tractable, randomised and untrainable function to each ensemble member, that enables a posterior interpretation in the infinite width limit. When ensembled together, our trained NNs give an approximation to a posterior predictive distribution, and we prove that our Bayesian deep ensembles make more conservative predictions than standard deep ensembles in the infinite width limit. Finally, using finite width NNs we demonstrate that our Bayesian deep ensembles faithfully emulate the analytic posterior predictive when available, and outperform standard deep ensembles in various out-of-distribution settings, for both regression and classification tasks.