Hanna Mazzawi

Hanna Mazzawi, Xavi Gonzalvo, Michael Wunder et al.

In recent years, deep learning has made remarkable progress in a wide range of domains, with a particularly notable impact on natural language processing tasks. One of the challenges associated with training deep neural networks in the context of LLMs is the need for large amounts of computational resources and time. To mitigate this, network growing algorithms offer potential cost savings, but their underlying mechanisms are poorly understood. We present two notable contributions in this paper. First, we present Deep Fusion, an efficient approach to network training that leverages pre-trained initializations of smaller networks. Second, we propose a theoretical framework using backward error analysis to illustrate the dynamics of mid-training network growth. Our experiments show how Deep Fusion is a practical and effective approach that not only accelerates the training process but also reduces computational requirements, maintaining or surpassing traditional training methods' performance in various NLP tasks and T5 model sizes. Finally, we validate our theoretical framework, which guides the optimal use of Deep Fusion, showing that with carefully optimized training dynamics, it significantly reduces both training time and resource consumption.

Charles Weill, Javier Gonzalvo, Vitaly Kuznetsov et al.

AdaNet is a lightweight TensorFlow-based (Abadi et al., 2015) framework for automatically learning high-quality ensembles with minimal expert intervention. Our framework is inspired by the AdaNet algorithm (Cortes et al., 2017) which learns the structure of a neural network as an ensemble of subnetworks. We designed it to: (1) integrate with the existing TensorFlow ecosystem, (2) offer sensible default search spaces to perform well on novel datasets, (3) present a flexible API to utilize expert information when available, and (4) efficiently accelerate training with distributed CPU, GPU, and TPU hardware. The code is open-source and available at: https://github.com/tensorflow/adanet.

Benoit Dherin, Michael Munn, Hanna Mazzawi et al.

One of the most striking features of Large Language Models (LLM) is their ability to learn in context. Namely at inference time an LLM is able to learn new patterns without any additional weight update when these patterns are presented in the form of examples in the prompt, even if these patterns were not seen during training. The mechanisms through which this can happen are still largely unknown. In this work, we show that the stacking of a self-attention layer with an MLP, allows the transformer block to implicitly modify the weights of the MLP layer according to the context. We argue through theory and experimentation that this simple mechanism may be the reason why LLMs can learn in context and not only during training. Specifically, we show under mild simplifying assumptions how a transformer block implicitly transforms a context into a low-rank weight-update of the MLP layer.

Benoit Dherin, Benny Avelin, Anders Karlsson et al.

Despite exceptional achievements, training neural networks remains computationally expensive and is often plagued by instabilities that can degrade convergence. While learning rate schedules can help mitigate these issues, finding optimal schedules is time-consuming and resource-intensive. This work explores theoretical issues concerning training stability in the constant-learning-rate (i.e., without schedule) and small-batch-size regime. Surprisingly, we show that the order of gradient updates affects stability and convergence in gradient-based optimizers. We illustrate this new line of thinking using backward-SGD, which processes batch gradient updates like SGD but in reverse order. Our theoretical analysis shows that in contractive regions (e.g., around minima) backward-SGD converges to a point while the standard forward-SGD generally only converges to a distribution. This leads to improved stability and convergence which we demonstrate experimentally. While full backward-SGD is computationally intensive in practice, it highlights opportunities to exploit reverse training dynamics (or more generally alternate iteration orders) to improve training. To our knowledge, this represents a new and unexplored avenue in deep learning optimization.

Hanna Mazzawi, Benoit Dherin, Michael Munn et al.

A growing body of research has demonstrated that the behavior of large language models can be effectively controlled at inference time by directly modifying their internal states, either through vector additions to their activations or through updates to their weight matrices. These techniques, while powerful, are often guided by empirical heuristics, such as deriving steering vectors from the average activations of contrastive prompts. This work provides a theoretical foundation for these interventions, explaining how they emerge from the fundamental computations of the transformer architecture. Building on the recent finding that a prompt's influence can be mathematically mapped to implicit weight updates (Dherin et al., 2025), we generalize this theory to deep, multi-block transformers. We show how the information contained in any chunk of a user prompt is represented and composed internally through weight vectors and weight matrices. We then derive a principled method for condensing this information into token-independent thought vectors and thought matrices. These constructs provide a theoretical explanation for existing vector- and matrix-based model editing techniques and offer a direct, computationally-grounded method for transmuting textual input into reusable weight updates.

Benoit Dherin, Michael Munn, Hanna Mazzawi et al.

Modern deep learning algorithms use variations of gradient descent as their main learning methods. Gradient descent can be understood as the simplest Ordinary Differential Equation (ODE) solver; namely, the Euler method applied to the gradient flow differential equation. Since Euler, many ODE solvers have been devised that follow the gradient flow equation more precisely and more stably. Runge-Kutta (RK) methods provide a family of very powerful explicit and implicit high-order ODE solvers. However, these higher-order solvers have not found wide application in deep learning so far. In this work, we evaluate the performance of higher-order RK solvers when applied in deep learning, study their limitations, and propose ways to overcome these drawbacks. In particular, we explore how to improve their performance by naturally incorporating key ingredients of modern neural network optimizers such as preconditioning, adaptive learning rates, and momentum.

Hanna Mazzawi, Pranjal Awasthi, Xavi Gonzalvo et al.

Recent breakthroughs and successful deployment of large language and vision models in a constrained environment predominantly follow a two phase approach. First, large models are trained to achieve peak performance, followed by a model shrinking method to meet hardware constraints; Methods like distillation, compression or quantization help leverage the highly performant large models to induce smaller performant ones. Formally, this can be seen as the problem of identifying an optimal model of size $n$ from a larger model of size $k \cdot n$, where $k > 1$ is the overparameterization factor. This paper explores the hypothesis that a single training run can simultaneously train a larger model for performance and derive a smaller model for deployment. Our contribution is an effective architectural change, namely, {\it Majority Kernels} that is compatible with the main standard architectures such as multi-layer perceptrons (MLPs), Residual networks (ResNets), and Transformers. We demonstrate that applying our technique can modify the training dynamics resulting in performance gains across architectures and tasks while maintaining the inference performance consistent. Furthermore, our approach adds minimal overhead to the cost incurred (wall clock time) at training time. The proposed approach shows strong performance on a wide variety of datasets and models, even outperforming strong baselines such as distilled ensembles as well as combinatorial optimization methods based on submodular optimization.

Scott Yak, Javier Gonzalvo, Hanna Mazzawi

Can we use deep learning to predict when deep learning works? Our results suggest the affirmative. We created a dataset by training 13,500 neural networks with different architectures, on different variations of spiral datasets, and using different optimization parameters. We used this dataset to train task-independent and architecture-independent generalization gap predictors for those neural networks. We extend Jiang et al. (2018) to also use DNNs and RNNs and show that they outperform the linear model, obtaining $R^2=0.965$. We also show results for architecture-independent, task-independent, and out-of-distribution generalization gap prediction tasks. Both DNNs and RNNs consistently and significantly outperform linear models, with RNNs obtaining $R^2=0.584$.

Vladimir Macko, Charles Weill, Hanna Mazzawi et al.

Finding the best neural network architecture requires significant time, resources, and human expertise. These challenges are partially addressed by neural architecture search (NAS) which is able to find the best convolutional layer or cell that is then used as a building block for the network. However, once a good building block is found, manual design is still required to assemble the final architecture as a combination of multiple blocks under a predefined parameter budget constraint. A common solution is to stack these blocks into a single tower and adjust the width and depth to fill the parameter budget. However, these single tower architectures may not be optimal. Instead, in this paper we present the AdaNAS algorithm, that uses ensemble techniques to compose a neural network as an ensemble of smaller networks automatically. Additionally, we introduce a novel technique based on knowledge distillation to iteratively train the smaller networks using the previous ensemble as a teacher. Our experiments demonstrate that ensembles of networks improve accuracy upon a single neural network while keeping the same number of parameters. Our models achieve comparable results with the state-of-the-art on CIFAR-10 and sets a new state-of-the-art on CIFAR-100.

Hasan Abasi, Nader H. Bshouty, Hanna Mazzawi

We give a new deterministic algorithm that non-adaptively learns a hidden hypergraph from edge-detecting queries. All previous non-adaptive algorithms either run in exponential time or have non-optimal query complexity. We give the first polynomial time non-adaptive learning algorithm for learning hypergraph that asks almost optimal number of queries.

Hasan Abasi, Nader H. Bshouty, Hanna Mazzawi

In this paper, we study the problem of learning a monotone DNF with at most $s$ terms of size (number of variables in each term) at most $r$ ($s$ term $r$-MDNF) from membership queries. This problem is equivalent to the problem of learning a general hypergraph using hyperedge-detecting queries, a problem motivated by applications arising in chemical reactions and genome sequencing. We first present new lower bounds for this problem and then present deterministic and randomized adaptive algorithms with query complexities that are almost optimal. All the algorithms we present in this paper run in time linear in the query complexity and the number of variables $n$. In addition, all of the algorithms we present in this paper are asymptotically tight for fixed $r$ and/or $s$.