What Does It Mean to Be a Transformer? Insights from a Theoretical Hessian Analysis

The Transformer architecture has inarguably revolutionized deep learning, overtaking classical architectures like multi-layer perceptrons (MLPs) and convolutional neural networks (CNNs). At its core, the attention block differs in form and functionality from most other architectural components in deep learning -- to the extent that Transformers are often accompanied by adaptive optimizers, layer normalization, learning rate warmup, and more, in comparison to MLPs/CNNs. The root causes behind these outward manifestations, and the precise mechanisms that govern them, remain poorly understood. In this work, we bridge this gap by providing a fundamental understanding of what distinguishes the Transformer from the other architectures -- grounded in a theoretical comparison of the (loss) Hessian. Concretely, for a single self-attention layer, (a) we first entirely derive the Transformer's Hessian and express it in matrix derivatives; (b) we then characterize it in terms of data, weight, and attention moment dependencies; and (c) while doing so further highlight the important structural differences to the Hessian of classical networks. Our results suggest that various common architectural and optimization choices in Transformers can be traced back to their highly non-linear dependencies on the data and weight matrices, which vary heterogeneously across parameters. Ultimately, our findings provide a deeper understanding of the Transformer's unique optimization landscape and the challenges it poses.

Revisiting Scalable Hessian Diagonal Approximations for Applications in Reinforcement Learning

Second-order information is valuable for many applications but challenging to compute. Several works focus on computing or approximating Hessian diagonals, but even this simplification introduces significant additional costs compared to computing a gradient. In the absence of efficient exact computation schemes for Hessian diagonals, we revisit an early approximation scheme proposed by Becker and LeCun (1989, BL89), which has a cost similar to gradients and appears to have been overlooked by the community. We introduce HesScale, an improvement over BL89, which adds negligible extra computation. On small networks, we find that this improvement is of higher quality than all alternatives, even those with theoretical guarantees, such as unbiasedness, while being much cheaper to compute. We use this insight in reinforcement learning problems where small networks are used and demonstrate HesScale in second-order optimization and scaling the step-size parameter. In our experiments, HesScale optimizes faster than existing methods and improves stability through step-size scaling. These findings are promising for scaling second-order methods in larger models in the future.

Lowering PyTorch's Memory Consumption for Selective Differentiation

Memory is a limiting resource for many deep learning tasks. Beside the neural network weights, one main memory consumer is the computation graph built up by automatic differentiation (AD) for backpropagation. We observe that PyTorch's current AD implementation neglects information about parameter differentiability when storing the computation graph. This information is useful though to reduce memory whenever gradients are requested for a parameter subset, as is the case in many modern fine-tuning tasks. Specifically, inputs to layers that act linearly in their parameters (dense, convolution, or normalization layers) can be discarded whenever the parameters are marked as non-differentiable. We provide a drop-in, differentiability-agnostic implementation of such layers and demonstrate its ability to reduce memory without affecting run time.

Can We Remove the Square-Root in Adaptive Gradient Methods? A Second-Order Perspective

Adaptive gradient optimizers like Adam(W) are the default training algorithms for many deep learning architectures, such as transformers. Their diagonal preconditioner is based on the gradient outer product which is incorporated into the parameter update via a square root. While these methods are often motivated as approximate second-order methods, the square root represents a fundamental difference. In this work, we investigate how the behavior of adaptive methods changes when we remove the root, i.e. strengthen their second-order motivation. Surprisingly, we find that such square-root-free adaptive methods close the generalization gap to SGD on convolutional architectures, while maintaining their root-based counterpart's performance on transformers. The second-order perspective also has practical benefits for the development of adaptive methods with non-diagonal preconditioner. In contrast to root-based counterparts like Shampoo, they do not require numerically unstable matrix square roots and therefore work well in low precision, which we demonstrate empirically. This raises important questions regarding the currently overlooked role of adaptivity for the success of adaptive methods since the success is often attributed to sign descent induced by the root.

Structured Inverse-Free Natural Gradient: Memory-Efficient & Numerically-Stable KFAC for Large Neural Nets

Second-order methods for deep learning -- such as KFAC -- can be useful for neural net training. However, they are often memory-inefficient and numerically unstable for low-precision training since their preconditioning Kronecker factors are dense, and require high-precision matrix inversion or decomposition. Consequently, such methods are not widely used for training large neural networks such as transformer-based models. We address these two issues by (i) formulating an inverse-free update of KFAC and (ii) imposing structures in each of the Kronecker factors, resulting in a method we term structured inverse-free natural gradient descent (SINGD). On large modern neural networks, we show that, in contrast to KFAC, SINGD is memory efficient and numerically robust, and often outperforms AdamW even in half precision. Hence, our work closes a gap between first-order and second-order methods in modern low precision training for large neural nets.

On the Disconnect Between Theory and Practice of Overparametrized Neural Networks

The infinite-width limit of neural networks (NNs) has garnered significant attention as a theoretical framework for analyzing the behavior of large-scale, overparametrized networks. By approaching infinite width, NNs effectively converge to a linear model with features characterized by the neural tangent kernel (NTK). This establishes a connection between NNs and kernel methods, the latter of which are well understood. Based on this link, theoretical benefits and algorithmic improvements have been hypothesized and empirically demonstrated in synthetic architectures. These advantages include faster optimization, reliable uncertainty quantification and improved continual learning. However, current results quantifying the rate of convergence to the kernel regime suggest that exploiting these benefits requires architectures that are orders of magnitude wider than they are deep. This assumption raises concerns that practically relevant architectures do not exhibit behavior as predicted via the NTK. In this work, we empirically investigate whether the limiting regime either describes the behavior of large-width architectures used in practice or is informative for algorithmic improvements. Our empirical results demonstrate that this is not the case in optimization, uncertainty quantification or continual learning. This observed disconnect between theory and practice calls into question the practical relevance of the infinite-width limit.

Convolutions Through the Lens of Tensor Networks

Despite their simple intuition, convolutions are more tedious to analyze than dense layers, which complicates the generalization of theoretical and algorithmic ideas. We provide a new perspective onto convolutions through tensor networks (TNs) which allow reasoning about the underlying tensor multiplications by drawing diagrams, and manipulating them to perform function transformations, sub-tensor access, and fusion. We demonstrate this expressive power by deriving the diagrams of various autodiff operations and popular approximations of second-order information with full hyper-parameter support, batching, channel groups, and generalization to arbitrary convolution dimensions. Further, we provide convolution-specific transformations based on the connectivity pattern which allow to re-wire and simplify diagrams before evaluation. Finally, we probe computational performance, relying on established machinery for efficient TN contraction. Our TN implementation speeds up a recently-proposed KFAC variant up to 4.5x and enables new hardware-efficient tensor dropout for approximate backpropagation.

The Geometry of Neural Nets' Parameter Spaces Under Reparametrization

Model reparametrization -- transforming the parameter space via a bijective differentiable map -- is a popular way to improve the training of neural networks. But reparametrizations have also been problematic since they induce inconsistencies in, e.g., Hessian-based flatness measures, optimization trajectories, and modes of probability density functions. This complicates downstream analyses, e.g. one cannot make a definitive statement about the connection between flatness and generalization. In this work, we study the invariance quantities of neural nets under reparametrization from the perspective of Riemannian geometry. We show that this notion of invariance is an inherent property of any neural net, as long as one acknowledges the assumptions about the metric that is always present, albeit often implicitly, and uses the correct transformation rules under reparametrization. We present discussions on measuring the flatness of minima, in optimization, and in probability-density maximization, along with applications in studying the biases of optimizers and in Bayesian inference.

ViViT: Curvature access through the generalized Gauss-Newton's low-rank structure

Curvature in form of the Hessian or its generalized Gauss-Newton (GGN) approximation is valuable for algorithms that rely on a local model for the loss to train, compress, or explain deep networks. Existing methods based on implicit multiplication via automatic differentiation or Kronecker-factored block diagonal approximations do not consider noise in the mini-batch. We present ViViT, a curvature model that leverages the GGN's low-rank structure without further approximations. It allows for efficient computation of eigenvalues, eigenvectors, as well as per-sample first- and second-order directional derivatives. The representation is computed in parallel with gradients in one backward pass and offers a fine-grained cost-accuracy trade-off, which allows it to scale. As examples for ViViT's usefulness, we investigate the directional gradients and curvatures during training, and how noise information can be used to improve the stability of second-order methods.

Cockpit: A Practical Debugging Tool for Training Deep Neural Networks

When engineers train deep learning models, they are very much "flying blind". Commonly used approaches for real-time training diagnostics, such as monitoring the train/test loss, are limited. Assessing a network's training process solely through these performance indicators is akin to debugging software without access to internal states through a debugger. To address this, we present Cockpit, a collection of instruments that enable a closer look into the inner workings of a learning machine, and a more informative and meaningful status report for practitioners. It facilitates the identification of learning phases and failure modes, like ill-chosen hyperparameters. These instruments leverage novel higher-order information about the gradient distribution and curvature, which has only recently become efficiently accessible. We believe that such a debugging tool, which we open-source for PyTorch, represents an important step to improve troubleshooting the training process, reveal new insights, and help develop novel methods and heuristics.