Scaling Laws for Data Filtering -- Data Curation cannot be Compute Agnostic

Vision-language models (VLMs) are trained for thousands of GPU hours on carefully curated web datasets. In recent times, data curation has gained prominence with several works developing strategies to retain 'high-quality' subsets of 'raw' scraped data. For instance, the LAION public dataset retained only 10% of the total crawled data. However, these strategies are typically developed agnostic of the available compute for training. In this paper, we first demonstrate that making filtering decisions independent of training compute is often suboptimal: the limited high-quality data rapidly loses its utility when repeated, eventually requiring the inclusion of 'unseen' but 'lower-quality' data. To address this quality-quantity tradeoff ($\texttt{QQT}$), we introduce neural scaling laws that account for the non-homogeneous nature of web data, an angle ignored in existing literature. Our scaling laws (i) characterize the $\textit{differing}$ 'utility' of various quality subsets of web data; (ii) account for how utility diminishes for a data point at its 'nth' repetition; and (iii) formulate the mutual interaction of various data pools when combined, enabling the estimation of model performance on a combination of multiple data pools without ever jointly training on them. Our key message is that data curation $\textit{cannot}$ be agnostic of the total compute that a model will be trained for. Our scaling laws allow us to curate the best possible pool for achieving top performance on Datacomp at various compute budgets, carving out a pareto-frontier for data curation. Code is available at https://github.com/locuslab/scaling_laws_data_filtering.

Think before you speak: Training Language Models With Pause Tokens

Language models generate responses by producing a series of tokens in immediate succession: the $(K+1)^{th}$ token is an outcome of manipulating $K$ hidden vectors per layer, one vector per preceding token. What if instead we were to let the model manipulate say, $K+10$ hidden vectors, before it outputs the $(K+1)^{th}$ token? We operationalize this idea by performing training and inference on language models with a (learnable) $\textit{pause}$ token, a sequence of which is appended to the input prefix. We then delay extracting the model's outputs until the last pause token is seen, thereby allowing the model to process extra computation before committing to an answer. We empirically evaluate $\textit{pause-training}$ on decoder-only models of 1B and 130M parameters with causal pretraining on C4, and on downstream tasks covering reasoning, question-answering, general understanding and fact recall. Our main finding is that inference-time delays show gains when the model is both pre-trained and finetuned with delays. For the 1B model, we witness gains on 8 of 9 tasks, most prominently, a gain of $18\%$ EM score on the QA task of SQuAD, $8\%$ on CommonSenseQA and $1\%$ accuracy on the reasoning task of GSM8k. Our work raises a range of conceptual and practical future research questions on making delayed next-token prediction a widely applicable new paradigm.

T-MARS: Improving Visual Representations by Circumventing Text Feature Learning

Large web-sourced multimodal datasets have powered a slew of new methods for learning general-purpose visual representations, advancing the state of the art in computer vision and revolutionizing zero- and few-shot recognition. One crucial decision facing practitioners is how, if at all, to curate these ever-larger datasets. For example, the creators of the LAION-5B dataset chose to retain only image-caption pairs whose CLIP similarity score exceeded a designated threshold. In this paper, we propose a new state-of-the-art data filtering approach motivated by our observation that nearly 40% of LAION's images contain text that overlaps significantly with the caption. Intuitively, such data could be wasteful as it incentivizes models to perform optical character recognition rather than learning visual features. However, naively removing all such data could also be wasteful, as it throws away images that contain visual features (in addition to overlapping text). Our simple and scalable approach, T-MARS (Text Masking and Re-Scoring), filters out only those pairs where the text dominates the remaining visual features -- by first masking out the text and then filtering out those with a low CLIP similarity score of the masked image. Experimentally, T-MARS outperforms the top-ranked method on the "medium scale" of DataComp (a data filtering benchmark) by a margin of 6.5% on ImageNet and 4.7% on VTAB. Additionally, our systematic evaluation on various data pool sizes from 2M to 64M shows that the accuracy gains enjoyed by T-MARS linearly increase as data and compute are scaled exponentially. Code is available at https://github.com/locuslab/T-MARS.

Finetune like you pretrain: Improved finetuning of zero-shot vision models

Finetuning image-text models such as CLIP achieves state-of-the-art accuracies on a variety of benchmarks. However, recent works like WiseFT (Wortsman et al., 2021) and LP-FT (Kumar et al., 2022) have shown that even subtle differences in the finetuning process can lead to surprisingly large differences in the final performance, both for in-distribution (ID) and out-of-distribution (OOD) data. In this work, we show that a natural and simple approach of mimicking contrastive pretraining consistently outperforms alternative finetuning approaches. Specifically, we cast downstream class labels as text prompts and continue optimizing the contrastive loss between image embeddings and class-descriptive prompt embeddings (contrastive finetuning). Our method consistently outperforms baselines across 7 distribution shifts, 6 transfer learning, and 3 few-shot learning benchmarks. On WILDS-iWILDCam, our proposed approach FLYP outperforms the top of the leaderboard by $2.3\%$ ID and $2.7\%$ OOD, giving the highest reported accuracy. Averaged across 7 OOD datasets (2 WILDS and 5 ImageNet associated shifts), FLYP gives gains of $4.2\%$ OOD over standard finetuning and outperforms the current state of the art (LP-FT) by more than $1\%$ both ID and OOD. Similarly, on 3 few-shot learning benchmarks, our approach gives gains up to $4.6\%$ over standard finetuning and $4.4\%$ over the state of the art. In total, these benchmarks establish contrastive finetuning as a simple, intuitive, and state-of-the-art approach for supervised finetuning of image-text models like CLIP. Code is available at https://github.com/locuslab/FLYP.

Test-Time Adaptation via Conjugate Pseudo-labels

Test-time adaptation (TTA) refers to adapting neural networks to distribution shifts, with access to only the unlabeled test samples from the new domain at test-time. Prior TTA methods optimize over unsupervised objectives such as the entropy of model predictions in TENT [Wang et al., 2021], but it is unclear what exactly makes a good TTA loss. In this paper, we start by presenting a surprising phenomenon: if we attempt to meta-learn the best possible TTA loss over a wide class of functions, then we recover a function that is remarkably similar to (a temperature-scaled version of) the softmax-entropy employed by TENT. This only holds, however, if the classifier we are adapting is trained via cross-entropy; if trained via squared loss, a different best TTA loss emerges. To explain this phenomenon, we analyze TTA through the lens of the training losses's convex conjugate. We show that under natural conditions, this (unsupervised) conjugate function can be viewed as a good local approximation to the original supervised loss and indeed, it recovers the best losses found by meta-learning. This leads to a generic recipe that can be used to find a good TTA loss for any given supervised training loss function of a general class. Empirically, our approach consistently dominates other baselines over a wide range of benchmarks. Our approach is particularly of interest when applied to classifiers trained with novel loss functions, e.g., the recently-proposed PolyLoss, where it differs substantially from (and outperforms) an entropy-based loss. Further, we show that our approach can also be interpreted as a kind of self-training using a very specific soft label, which we refer to as the conjugate pseudolabel. Overall, our method provides a broad framework for better understanding and improving test-time adaptation. Code is available at https://github.com/locuslab/tta_conjugate.

MET: Masked Encoding for Tabular Data

We consider the task of self-supervised representation learning (SSL) for tabular data: tabular-SSL. Typical contrastive learning based SSL methods require instance-wise data augmentations which are difficult to design for unstructured tabular data. Existing tabular-SSL methods design such augmentations in a relatively ad-hoc fashion and can fail to capture the underlying data manifold. Instead of augmentations based approaches for tabular-SSL, we propose a new reconstruction based method, called Masked Encoding for Tabular Data (MET), that does not require augmentations. MET is based on the popular MAE approach for vision-SSL [He et al., 2021] and uses two key ideas: (i) since each coordinate in a tabular dataset has a distinct meaning, we need to use separate representations for all coordinates, and (ii) using an adversarial reconstruction loss in addition to the standard one. Empirical results on five diverse tabular datasets show that MET achieves a new state of the art (SOTA) on all of these datasets and improves up to 9% over current SOTA methods. We shed more light on the working of MET via experiments on carefully designed simple datasets.

PAL : Pretext-based Active Learning

When obtaining labels is expensive, the requirement of a large labeled training data set for deep learning can be mitigated by active learning. Active learning refers to the development of algorithms to judiciously pick limited subsets of unlabeled samples that can be sent for labeling by an oracle. We propose an intuitive active learning technique that, in addition to the task neural network (e.g., for classification), uses an auxiliary self-supervised neural network that assesses the utility of an unlabeled sample for inclusion in the labeled set. Our core idea is that the difficulty of the auxiliary network trained on labeled samples to solve a self-supervision task on an unlabeled sample represents the utility of obtaining the label of that unlabeled sample. Specifically, we assume that an unlabeled image on which the precision of predicting a random applied geometric transform is low must be out of the distribution represented by the current set of labeled images. These images will therefore maximize the relative information gain when labeled by the oracle. We also demonstrate that augmenting the auxiliary network with task specific training further improves the results. We demonstrate strong performance on a range of widely used datasets and establish a new state of the art for active learning. We also make our code publicly available to encourage further research.

Indoor Distance Estimation using LSTMs over WLAN Network

The Global Navigation Satellite Systems (GNSS) like GPS suffer from accuracy degradation and are almost unavailable in indoor environments. Indoor positioning systems (IPS) based on WiFi signals have been gaining popularity. However, owing to the strong spatial and temporal variations of wireless communication channels in the indoor environment, the achieved accuracy of existing IPS is around several tens of centimeters. We present the detailed design and implementation of a self-adaptive WiFi-based indoor distance estimation system using LSTMs. The system is novel in its method of estimating with high accuracy the distance of an object by overcoming possible causes of channel variations and is self-adaptive to the changing environmental and surrounding conditions. The proposed design has been developed and physically realized over a WiFi network consisting of ESP8266 (NodeMCU) devices. The experiment were conducted in a real indoor environment while changing the surroundings in order to establish the adaptability of the system. We introduce and compare different architectures for this task based on LSTMs, CNNs, and fully connected networks (FCNs). We show that the LSTM based model performs better among all the above-mentioned architectures by achieving an accuracy of 5.85 cm with a confidence interval of 93% on the scale of (4.14 m * 2.86 m). To the best of our knowledge, the proposed method outperforms other methods reported in the literature by a significant margin.

DROCC: Deep Robust One-Class Classification

Classical approaches for one-class problems such as one-class SVM (Scholkopf et al., 1999) and isolation forest (Liu et al., 2008) require careful feature engineering when applied to structured domains like images. To alleviate this concern, state-of-the-art methods like DeepSVDD (Ruff et al., 2018) consider the natural alternative of minimizing a classical one-class loss applied to the learned final layer representations. However, such an approach suffers from the fundamental drawback that a representation that simply collapses all the inputs minimizes the one class loss; heuristics to mitigate collapsed representations provide limited benefits. In this work, we propose Deep Robust One Class Classification (DROCC) method that is robust to such a collapse by training the network to distinguish the training points from their perturbations, generated adversarially. DROCC is motivated by the assumption that the interesting class lies on a locally linear low dimensional manifold. Empirical evaluation demonstrates DROCC's effectiveness on two different one-class problem settings and on a range of real-world datasets across different domains - images(CIFAR and ImageNet), audio and timeseries, offering up to 20% increase in accuracy over the state-of-the-art in anomaly detection.