GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection

Training Large Language Models (LLMs) presents significant memory challenges, predominantly due to the growing size of weights and optimizer states. Common memory-reduction approaches, such as low-rank adaptation (LoRA), add a trainable low-rank matrix to the frozen pre-trained weight in each layer, reducing trainable parameters and optimizer states. However, such approaches typically underperform training with full-rank weights in both pre-training and fine-tuning stages since they limit the parameter search to a low-rank subspace and alter the training dynamics, and further, may require full-rank warm start. In this work, we propose Gradient Low-Rank Projection (GaLore), a training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods such as LoRA. Our approach reduces memory usage by up to 65.5% in optimizer states while maintaining both efficiency and performance for pre-training on LLaMA 1B and 7B architectures with C4 dataset with up to 19.7B tokens, and on fine-tuning RoBERTa on GLUE tasks. Our 8-bit GaLore further reduces optimizer memory by up to 82.5% and total training memory by 63.3%, compared to a BF16 baseline. Notably, we demonstrate, for the first time, the feasibility of pre-training a 7B model on consumer GPUs with 24GB memory (e.g., NVIDIA RTX 4090) without model parallel, checkpointing, or offloading strategies.

InRank: Incremental Low-Rank Learning

The theory of greedy low-rank learning (GLRL) aims to explain the impressive generalization capabilities of deep learning. It proves that stochastic gradient-based training implicitly regularizes neural networks towards low-rank solutions through a gradual increase of the rank during training. However, there is a gap between theory and practice since GLRL requires an infinitesimal initialization of the weights, which is not practical due to the fact that it is a saddle point. In this work, we remove the assumption of infinitesimal initialization by focusing on cumulative weight updates. We prove the cumulative weight updates follow an incremental low-rank trajectory for arbitrary orthogonal initialization of weights in a three-layer linear network. Empirically, we demonstrate that our theory holds on a broad range of neural networks (e.g., transformers) and standard training algorithms (e.g., SGD, Adam). However, existing training algorithms do not exploit the low-rank property to improve computational efficiency as the networks are not parameterized in low-rank. To remedy this, we design a new training algorithm Incremental Low-Rank Learning (InRank), which explicitly expresses cumulative weight updates as low-rank matrices while incrementally augmenting their ranks during training. We evaluate InRank on GPT-2, and our results indicate that InRank achieves comparable prediction performance as the full-rank counterpart while requiring at most 33% of the total ranks throughout training. We also propose an efficient version of InRank that achieves a reduction of 20% in total training time and 37% in memory usage when training GPT-medium on WikiText-103 from scratch.

UMIRobot: An Open-{Software, Hardware} Low-Cost Robotic Manipulator for Education

Robot teleoperation has been studied for the past 70 years and is relevant in many contexts, such as in the handling of hazardous materials and telesurgery. The COVID19 pandemic has rekindled interest in this topic, but the existing robotic education kits fail short of being suitable for teleoperated robotic manipulator learning. In addition, the global restrictions of motion motivated large investments in online/hybrid education. In this work, a newly developed robotics education kit and its ecosystem are presented which is used as the backbone of an online/hybrid course in teleoperated robots. The students are assembled into teams, design, fabricate, and control a master device and gripper, and compete in a teleoperation challenge. The kit is low cost (< 100USD), which allows higher-learning institutions to provide one kit per student and they can learn in a risk-free environment. As of now, 53 such kits have been assembled and sent to course participants in eight countries. As major success stories, we show an example of gripper and master designed for the proposed course. In addition, we show a teleoperated task between Japan and Bangladesh executed by course participants. Design files, videos, and more information available at https://mmmarinho.github.io/UMIRobot/

Incremental Fourier Neural Operator

Recently, neural networks have proven their impressive ability to solve partial differential equations (PDEs). Among them, Fourier neural operator (FNO) has shown success in learning solution operators for highly non-linear problems such as turbulence flow. FNO is discretization-invariant, where it can be trained on low-resolution data and generalizes to problems with high-resolution. This property is related to the low-pass filters in FNO, where only a limited number of frequency modes are selected to propagate information. However, it is still a challenge to select an appropriate number of frequency modes and training resolution for different PDEs. Too few frequency modes and low-resolution data hurt generalization, while too many frequency modes and high-resolution data are computationally expensive and lead to over-fitting. To this end, we propose Incremental Fourier Neural Operator (IFNO), which augments both the frequency modes and data resolution incrementally during training. We show that IFNO achieves better generalization (around 15% reduction on testing L2 loss) while reducing the computational cost by 35%, compared to the standard FNO. In addition, we observe that IFNO follows the behavior of implicit regularization in FNO, which explains its excellent generalization ability.

ZerO Initialization: Initializing Residual Networks with only Zeros and Ones

Deep neural networks are usually initialized with random weights, with adequately selected initial variance to ensure stable signal propagation during training. However, there is no consensus on how to select the variance, and this becomes challenging especially as the number of layers grows. In this work, we replace the widely used random weight initialization with a fully deterministic initialization scheme ZerO, which initializes residual networks with only zeros and ones. By augmenting the standard ResNet architectures with a few extra skip connections and Hadamard transforms, ZerO allows us to start the training from zeros and ones entirely. This has many benefits such as improving reproducibility (by reducing the variance over different experimental runs) and allowing network training without batch normalization. Surprisingly, we find that ZerO achieves state-of-the-art performance over various image classification datasets, including ImageNet, which suggests random weights may be unnecessary for modern network initialization.

Transformer-based Dual Relation Graph for Multi-label Image Recognition

The simultaneous recognition of multiple objects in one image remains a challenging task, spanning multiple events in the recognition field such as various object scales, inconsistent appearances, and confused inter-class relationships. Recent research efforts mainly resort to the statistic label co-occurrences and linguistic word embedding to enhance the unclear semantics. Different from these researches, in this paper, we propose a novel Transformer-based Dual Relation learning framework, constructing complementary relationships by exploring two aspects of correlation, i.e., structural relation graph and semantic relation graph. The structural relation graph aims to capture long-range correlations from object context, by developing a cross-scale transformer-based architecture. The semantic graph dynamically models the semantic meanings of image objects with explicit semantic-aware constraints. In addition, we also incorporate the learnt structural relationship into the semantic graph, constructing a joint relation graph for robust representations. With the collaborative learning of these two effective relation graphs, our approach achieves new state-of-the-art on two popular multi-label recognition benchmarks, i.e., MS-COCO and VOC 2007 dataset.

RGB-D Salient Object Detection with Ubiquitous Target Awareness

Conventional RGB-D salient object detection methods aim to leverage depth as complementary information to find the salient regions in both modalities. However, the salient object detection results heavily rely on the quality of captured depth data which sometimes are unavailable. In this work, we make the first attempt to solve the RGB-D salient object detection problem with a novel depth-awareness framework. This framework only relies on RGB data in the testing phase, utilizing captured depth data as supervision for representation learning. To construct our framework as well as achieving accurate salient detection results, we propose a Ubiquitous Target Awareness (UTA) network to solve three important challenges in RGB-D SOD task: 1) a depth awareness module to excavate depth information and to mine ambiguous regions via adaptive depth-error weights, 2) a spatial-aware cross-modal interaction and a channel-aware cross-level interaction, exploiting the low-level boundary cues and amplifying high-level salient channels, and 3) a gated multi-scale predictor module to perceive the object saliency in different contextual scales. Besides its high performance, our proposed UTA network is depth-free for inference and runs in real-time with 43 FPS. Experimental evidence demonstrates that our proposed network not only surpasses the state-of-the-art methods on five public RGB-D SOD benchmarks by a large margin, but also verifies its extensibility on five public RGB SOD benchmarks.

Low-Precision Training in Logarithmic Number System using Multiplicative Weight Update

Training large-scale deep neural networks (DNNs) currently requires a significant amount of energy, leading to serious environmental impacts. One promising approach to reduce the energy costs is representing DNNs with low-precision numbers. While it is common to train DNNs with forward and backward propagation in low-precision, training directly over low-precision weights, without keeping a copy of weights in high-precision, still remains to be an unsolved problem. This is due to complex interactions between learning algorithms and low-precision number systems. To address this, we jointly design a low-precision training framework involving a logarithmic number system (LNS) and a multiplicative weight update training method, termed LNS-Madam. LNS has a high dynamic range even in a low-bitwidth setting, leading to high energy efficiency and making it relevant for on-board training in energy-constrained edge devices. We design LNS to have the flexibility of choosing different bases for weights and gradients, as they usually require different quantization gaps and dynamic ranges during training. By drawing the connection between LNS and multiplicative update, LNS-Madam ensures low quantization error during weight update, leading to a stable convergence even if the bitwidth is limited. Compared to using a fixed-point or floating-point number system and training with popular learning algorithms such as SGD and Adam, our joint design with LNS and LNS-Madam optimizer achieves better accuracy while requiring smaller bitwidth. Notably, with only 5-bit for gradients, the proposed training framework achieves accuracy comparable to full-precision state-of-the-art models such as ResNet-50 and BERT. After conducting energy estimations by analyzing the math datapath units during training, the results show that our design achieves over 60x energy reduction compared to FP32 on BERT models.

A Knowledge Graph based Approach for Mobile Application Recommendation

With the rapid prevalence of mobile devices and the dramatic proliferation of mobile applications (apps), app recommendation becomes an emergent task that would benefit both app users and stockholders. How to effectively organize and make full use of rich side information of users and apps is a key challenge to address the sparsity issue for traditional approaches. To meet this challenge, we proposed a novel end-to-end Knowledge Graph Convolutional Embedding Propagation Model (KGEP) for app recommendation. Specifically, we first designed a knowledge graph construction method to model the user and app side information, then adopted KG embedding techniques to capture the factual triplet-focused semantics of the side information related to the first-order structure of the KG, and finally proposed a relation-weighted convolutional embedding propagation model to capture the recommendation-focused semantics related to high-order structure of the KG. Extensive experiments conducted on a real-world dataset validate the effectiveness of the proposed approach compared to the state-of-the-art recommendation approaches.

Learning compositional functions via multiplicative weight updates

Compositionality is a basic structural feature of both biological and artificial neural networks. Learning compositional functions via gradient descent incurs well known problems like vanishing and exploding gradients, making careful learning rate tuning essential for real-world applications. This paper proves that multiplicative weight updates satisfy a descent lemma tailored to compositional functions. Based on this lemma, we derive Madam---a multiplicative version of the Adam optimiser---and show that it can train state of the art neural network architectures without learning rate tuning. We further show that Madam is easily adapted to train natively compressed neural networks by representing their weights in a logarithmic number system. We conclude by drawing connections between multiplicative weight updates and recent findings about synapses in biology.