From Attention to Activation: Unravelling the Enigmas of Large Language Models

We study two strange phenomena in auto-regressive Transformers: (1) the dominance of the first token in attention heads; (2) the occurrence of large outlier activations in the hidden states. We find that popular large language models, such as Llama attend maximally to the first token in 98% of attention heads, a behaviour we attribute to the softmax function. To mitigate this issue, we propose a reformulation of softmax to softmax-1. Furthermore, we identify adaptive optimisers, e.g. Adam, as the primary contributor to the large outlier activations and introduce OrthoAdam, a novel optimiser that utilises orthogonal matrices to transform gradients, to address this issue. Finally, not only do our methods prevent these phenomena from occurring, but additionally, they enable Transformers to sustain their performance when quantised using basic algorithms, something that standard methods are unable to do. In summary, our methods reduce the attention proportion on the first token from 65% to 3.3%, the activation kurtosis in the hidden states from 1657 to 3.1, and perplexity penalty under 4-bit weight quantisation from 3565 to 0.3.

Fractal Calibration for long-tailed object detection

Real-world datasets follow an imbalanced distribution, which poses significant challenges in rare-category object detection. Recent studies tackle this problem by developing re-weighting and re-sampling methods, that utilise the class frequencies of the dataset. However, these techniques focus solely on the frequency statistics and ignore the distribution of the classes in image space, missing important information. In contrast to them, we propose FRActal CALibration (FRACAL): a novel post-calibration method for long-tailed object detection. FRACAL devises a logit adjustment method that utilises the fractal dimension to estimate how uniformly classes are distributed in image space. During inference, it uses the fractal dimension to inversely downweight the probabilities of uniformly spaced class predictions achieving balance in two axes: between frequent and rare categories, and between uniformly spaced and sparsely spaced classes. FRACAL is a post-processing method and it does not require any training, also it can be combined with many off-the-shelf models such as one-stage sigmoid detectors and two-stage instance segmentation models. FRACAL boosts the rare class performance by up to 8.6% and surpasses all previous methods on LVIS dataset, while showing good generalisation to other datasets such as COCO, V3Det and OpenImages. The code will be released.

VeLoRA: Memory Efficient Training using Rank-1 Sub-Token Projections

Large language models (LLMs) have recently emerged as powerful tools for tackling many language-processing tasks. Despite their success, training and fine-tuning these models is still far too computationally and memory intensive. In this paper, we identify and characterise the important components needed for effective model convergence using gradient descent. In doing so we find that the intermediate activations used to implement backpropagation can be excessively compressed without incurring any degradation in performance. This result leads us to a cheap and memory-efficient algorithm for both fine-tuning and pre-training LLMs. The proposed algorithm simply divides the tokens up into smaller sub-tokens before projecting them onto a fixed 1-dimensional subspace during the forward pass. These features are then coarsely reconstructed during the backward pass to implement the update rules. We confirm the effectiveness of our algorithm as being complimentary to many state-of-the-art PEFT methods on the VTAB-1k fine-tuning benchmark. Furthermore, we outperform QLoRA for fine-tuning LLaMA and show competitive performance against other memory-efficient pre-training methods on the large-scale C4 dataset.

Deep Active Learning: A Reality Check

We conduct a comprehensive evaluation of state-of-the-art deep active learning methods. Surprisingly, under general settings, no single-model method decisively outperforms entropy-based active learning, and some even fall short of random sampling. We delve into overlooked aspects like starting budget, budget step, and pretraining's impact, revealing their significance in achieving superior results. Additionally, we extend our evaluation to other tasks, exploring the active learning effectiveness in combination with semi-supervised learning, and object detection. Our experiments provide valuable insights and concrete recommendations for future active learning studies. By uncovering the limitations of current methods and understanding the impact of different experimental settings, we aim to inspire more efficient training of deep learning models in real-world scenarios with limited annotation budgets. This work contributes to advancing active learning's efficacy in deep learning and empowers researchers to make informed decisions when applying active learning to their tasks.

$V_kD:$ Improving Knowledge Distillation using Orthogonal Projections

Knowledge distillation is an effective method for training small and efficient deep learning models. However, the efficacy of a single method can degenerate when transferring to other tasks, modalities, or even other architectures. To address this limitation, we propose a novel constrained feature distillation method. This method is derived from a small set of core principles, which results in two emerging components: an orthogonal projection and a task-specific normalisation. Equipped with both of these components, our transformer models can outperform all previous methods on ImageNet and reach up to a 4.4% relative improvement over the previous state-of-the-art methods. To further demonstrate the generality of our method, we apply it to object detection and image generation, whereby we obtain consistent and substantial performance improvements over state-of-the-art. Code and models are publicly available: https://github.com/roymiles/vkd

G3DR: Generative 3D Reconstruction in ImageNet

We introduce a novel 3D generative method, Generative 3D Reconstruction (G3DR) in ImageNet, capable of generating diverse and high-quality 3D objects from single images, addressing the limitations of existing methods. At the heart of our framework is a novel depth regularization technique that enables the generation of scenes with high-geometric fidelity. G3DR also leverages a pretrained language-vision model, such as CLIP, to enable reconstruction in novel views and improve the visual realism of generations. Additionally, G3DR designs a simple but effective sampling procedure to further improve the quality of generations. G3DR offers diverse and efficient 3D asset generation based on class or text conditioning. Despite its simplicity, G3DR is able to beat state-of-theart methods, improving over them by up to 22% in perceptual metrics and 90% in geometry scores, while needing only half of the training time. Code is available at https://github.com/preddy5/G3DR

Three Heads Are Better Than One: Complementary Experts for Long-Tailed Semi-supervised Learning

We address the challenging problem of Long-Tailed Semi-Supervised Learning (LTSSL) where labeled data exhibit imbalanced class distribution and unlabeled data follow an unknown distribution. Unlike in balanced SSL, the generated pseudo-labels are skewed towards head classes, intensifying the training bias. Such a phenomenon is even amplified as more unlabeled data will be mislabeled as head classes when the class distribution of labeled and unlabeled datasets are mismatched. To solve this problem, we propose a novel method named ComPlementary Experts (CPE). Specifically, we train multiple experts to model various class distributions, each of them yielding high-quality pseudo-labels within one form of class distribution. Besides, we introduce Classwise Batch Normalization for CPE to avoid performance degradation caused by feature distribution mismatch between head and non-head classes. CPE achieves state-of-the-art performances on CIFAR-10-LT, CIFAR-100-LT, and STL-10-LT dataset benchmarks. For instance, on CIFAR-10-LT, CPE improves test accuracy by over >2.22% compared to baselines. Code is available at https://github.com/machengcheng2016/CPE-LTSSL.

Data-Driven but Privacy-Conscious: Pedestrian Dataset De-identification via Full-Body Person Synthesis

The advent of data-driven technology solutions is accompanied by an increasing concern with data privacy. This is of particular importance for human-centered image recognition tasks, such as pedestrian detection, re-identification, and tracking. To highlight the importance of privacy issues and motivate future research, we motivate and introduce the Pedestrian Dataset De-Identification (PDI) task. PDI evaluates the degree of de-identification and downstream task training performance for a given de-identification method. As a first baseline, we propose IncogniMOT, a two-stage full-body de-identification pipeline based on image synthesis via generative adversarial networks. The first stage replaces target pedestrians with synthetic identities. To improve downstream task performance, we then apply stage two, which blends and adapts the synthetic image parts into the data. To demonstrate the effectiveness of IncogniMOT, we generate a fully de-identified version of the MOT17 pedestrian tracking dataset and analyze its application as training data for pedestrian re-identification, detection, and tracking models. Furthermore, we show how our data is able to narrow the synthetic-to-real performance gap in a privacy-conscious manner.

Learning to Discover and Detect Objects

We tackle the problem of novel class discovery, detection, and localization (NCDL). In this setting, we assume a source dataset with labels for objects of commonly observed classes. Instances of other classes need to be discovered, classified, and localized automatically based on visual similarity, without human supervision. To this end, we propose a two-stage object detection network Region-based NCDL (RNCDL), that uses a region proposal network to localize object candidates and is trained to classify each candidate, either as one of the known classes, seen in the source dataset, or one of the extended set of novel classes, with a long-tail distribution constraint on the class assignments, reflecting the natural frequency of classes in the real world. By training our detection network with this objective in an end-to-end manner, it learns to classify all region proposals for a large variety of classes, including those that are not part of the labeled object class vocabulary. Our experiments conducted using COCO and LVIS datasets reveal that our method is significantly more effective compared to multi-stage pipelines that rely on traditional clustering algorithms or use pre-extracted crops. Furthermore, we demonstrate the generality of our approach by applying our method to a large-scale Visual Genome dataset, where our network successfully learns to detect various semantic classes without explicit supervision.

The Unreasonable Effectiveness of Fully-Connected Layers for Low-Data Regimes

Convolutional neural networks were the standard for solving many computer vision tasks until recently, when Transformers of MLP-based architectures have started to show competitive performance. These architectures typically have a vast number of weights and need to be trained on massive datasets; hence, they are not suitable for their use in low-data regimes. In this work, we propose a simple yet effective framework to improve generalization from small amounts of data. We augment modern CNNs with fully-connected (FC) layers and show the massive impact this architectural change has in low-data regimes. We further present an online joint knowledge-distillation method to utilize the extra FC layers at train time but avoid them during test time. This allows us to improve the generalization of a CNN-based model without any increase in the number of weights at test time. We perform classification experiments for a large range of network backbones and several standard datasets on supervised learning and active learning. Our experiments significantly outperform the networks without fully-connected layers, reaching a relative improvement of up to $16\%$ validation accuracy in the supervised setting without adding any extra parameters during inference.