The scalability of deep learning models is fundamentally limited by computing resources, memory, and communication. Although methods like low-rank adaptation (LoRA) have reduced the cost of model finetuning, its application in model pre-training remains largely unexplored. This paper explores extending LoRA to model pre-training, identifying the inherent constraints and limitations of standard LoRA in this context. We introduce LoRA-the-Explorer (LTE), a novel bi-level optimization algorithm designed to enable parallel training of multiple low-rank heads across computing nodes, thereby reducing the need for frequent synchronization. Our approach includes extensive experimentation on vision transformers using various vision datasets, demonstrating that LTE is competitive with standard pre-training.
The push to train ever larger neural networks has motivated the study of initialization and training at large network width. A key challenge is to scale training so that a network's internal representations evolve nontrivially at all widths, a process known as feature learning. Here, we show that feature learning is achieved by scaling the spectral norm of weight matrices and their updates like $\sqrt{\texttt{fan-out}/\texttt{fan-in}}$, in contrast to widely used but heuristic scalings based on Frobenius norm and entry size. Our spectral scaling analysis also leads to an elementary derivation of \emph{maximal update parametrization}. All in all, we aim to provide the reader with a solid conceptual understanding of feature learning in neural networks.
When machine learning models are trained continually on a sequence of tasks, they are liable to forget what they learned on previous tasks -- a phenomenon known as catastrophic forgetting. Proposed solutions to catastrophic forgetting tend to involve storing information about past tasks, meaning that memory usage is a chief consideration in determining their practicality. This paper proposes a memory-efficient solution to catastrophic forgetting, improving upon an established algorithm known as orthogonal gradient descent (OGD). OGD utilizes prior model gradients to find weight updates that preserve performance on prior datapoints. However, since the memory cost of storing prior model gradients grows with the runtime of the algorithm, OGD is ill-suited to continual learning over arbitrarily long time horizons. To address this problem, this paper proposes SketchOGD. SketchOGD employs an online sketching algorithm to compress model gradients as they are encountered into a matrix of a fixed, user-determined size. In contrast to existing memory-efficient variants of OGD, SketchOGD runs online without the need for advance knowledge of the total number of tasks, is simple to implement, and is more amenable to analysis. We provide theoretical guarantees on the approximation error of the relevant sketches under a novel metric suited to the downstream task of OGD. Experimentally, we find that SketchOGD tends to outperform current state-of-the-art variants of OGD given a fixed memory budget.
The architecture of a deep neural network is defined explicitly in terms of the number of layers, the width of each layer and the general network topology. Existing optimisation frameworks neglect this information in favour of implicit architectural information (e.g. second-order methods) or architecture-agnostic distance functions (e.g. mirror descent). Meanwhile, the most popular optimiser in practice, Adam, is based on heuristics. This paper builds a new framework for deriving optimisation algorithms that explicitly leverage neural architecture. The theory extends mirror descent to non-convex composite objective functions: the idea is to transform a Bregman divergence to account for the non-linear structure of neural architecture. Working through the details for deep fully-connected networks yields automatic gradient descent: a first-order optimiser without any hyperparameters. Automatic gradient descent trains both fully-connected and convolutional networks out-of-the-box and at ImageNet scale. A PyTorch implementation is available at https://github.com/jxbz/agd and also in Appendix B. Overall, the paper supplies a rigorous theoretical foundation for a next-generation of architecture-dependent optimisers that work automatically and without hyperparameters.
The goal of this thesis is to develop the optimisation and generalisation theoretic foundations of learning in artificial neural networks. On optimisation, a new theoretical framework is proposed for deriving architecture-dependent first-order optimisation algorithms. The approach works by combining a "functional majorisation" of the loss function with "architectural perturbation bounds" that encode an explicit dependence on neural architecture. The framework yields optimisation methods that transfer hyperparameters across learning problems. On generalisation, a new correspondence is proposed between ensembles of networks and individual networks. It is argued that, as network width and normalised margin are taken large, the space of networks that interpolate a particular training set concentrates on an aggregated Bayesian method known as a "Bayes point machine". This correspondence provides a route for transferring PAC-Bayesian generalisation theorems over to individual networks. More broadly, the correspondence presents a fresh perspective on the role of regularisation in networks with vastly more parameters than data.
Weight norm $\|w\|$ and margin $\gamma$ participate in learning theory via the normalized margin $\gamma/\|w\|$. Since standard neural net optimizers do not control normalized margin, it is hard to test whether this quantity causally relates to generalization. This paper designs a series of experimental studies that explicitly control normalized margin and thereby tackle two central questions. First: does normalized margin always have a causal effect on generalization? The paper finds that no -- networks can be produced where normalized margin has seemingly no relationship with generalization, counter to the theory of Bartlett et al. (2017). Second: does normalized margin ever have a causal effect on generalization? The paper finds that yes -- in a standard training setup, test performance closely tracks normalized margin. The paper suggests a Gaussian process model as a promising explanation for this behavior.
Do neural networks generalise because of bias in the functions returned by gradient descent, or bias already present in the network architecture? Por qu\'e no los dos? This paper finds that while typical networks that fit the training data already generalise fairly well, gradient descent can further improve generalisation by selecting networks with a large margin. This conclusion is based on a careful study of the behaviour of infinite width networks trained by Bayesian inference and finite width networks trained by gradient descent. To measure the implicit bias of architecture, new technical tools are developed to both analytically bound and consistently estimate the average test error of the neural network--Gaussian process (NNGP) posterior. This error is found to be already better than chance, corroborating the findings of Valle-P\'erez et al. (2019) and underscoring the importance of architecture. Going beyond this result, this paper finds that test performance can be substantially improved by selecting a function with much larger margin than is typical under the NNGP posterior. This highlights a curious fact: minimum a posteriori functions can generalise best, and gradient descent can select for those functions. In summary, new technical tools suggest a nuanced portrait of generalisation involving both the implicit biases of architecture and gradient descent. Code for this paper is available at: https://github.com/jxbz/implicit-bias/.
We study the problem of sparse nonlinear model recovery of high dimensional compositional functions. Our study is motivated by emerging opportunities in neuroscience to recover fine-grained models of biological neural circuits using collected measurement data. Guided by available domain knowledge in neuroscience, we explore conditions under which one can recover the underlying biological circuit that generated the training data. Our results suggest insights of both theoretical and practical interests. Most notably, we find that a sign constraint on the weights is a necessary condition for system recovery, which we establish both theoretically with an identifiability guarantee and empirically on simulated biological circuits. We conclude with a case study on retinal ganglion cell circuits using data collected from mouse retina, showcasing the practical potential of this approach.
How much information does a learning algorithm extract from the training data and store in a neural network's weights? Too much, and the network would overfit to the training data. Too little, and the network would not fit to anything at all. Na\"ively, the amount of information the network stores should scale in proportion to the number of trainable weights. This raises the question: how can neural networks with vastly more weights than training data still generalise? A simple resolution to this conundrum is that the number of weights is usually a bad proxy for the actual amount of information stored. For instance, typical weight vectors may be highly compressible. Then another question occurs: is it possible to compute the actual amount of information stored? This paper derives both a consistent estimator and a closed-form upper bound on the information content of infinitely wide neural networks. The derivation is based on an identification between neural information content and the negative log probability of a Gaussian orthant. This identification yields bounds that analytically control the generalisation behaviour of the entire solution space of infinitely wide networks. The bounds have a simple dependence on both the network architecture and the training data. Corroborating the findings of Valle-P\'erez et al. (2019), who conducted a similar analysis using approximate Gaussian integration techniques, the bounds are found to be both non-vacuous and correlated with the empirical generalisation behaviour at finite width.
Descent methods for deep networks are notoriously capricious: they require careful tuning of step size, momentum and weight decay, and which method will work best on a new benchmark is a priori unclear. To address this problem, this paper conducts a combined study of neural architecture and optimisation, leading to a new optimiser called Nero: the neuronal rotator. Nero trains reliably without momentum or weight decay, works in situations where Adam and SGD fail, and requires little to no learning rate tuning. Also, Nero's memory footprint is ~ square root that of Adam or LAMB. Nero combines two ideas: (1) projected gradient descent over the space of balanced networks; (2) neuron-specific updates, where the step size sets the angle through which each neuron's hyperplane turns. The paper concludes by discussing how this geometric connection between architecture and optimisation may impact theories of generalisation in deep learning.