Scaling laws are useful guides for derisking expensive training runs, as they predict performance of large models using cheaper, small-scale experiments. However, there remain gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime). In contrast, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but models are usually compared on downstream task performance. To address both shortcomings, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we fit scaling laws that extrapolate in both the amount of over-training and the number of model parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32$\times$ over-trained) and a 6.9B parameter, 138B token run (i.e., a compute-optimal run)––each from experiments that take 300$\times$ less compute. Second, we relate the perplexity of a language model to its downstream task performance by proposing a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models, using experiments that take 20$\times$ less compute.

The successful training of deep neural networks requires addressing challenges such as overfitting, numerical instabilities leading to divergence, and increasing variance in the residual stream. A common solution is to apply regularization and normalization techniques that usually require tuning additional hyperparameters. An alternative is to force all parameters and representations to lie on a hypersphere. This removes the need for regularization and increases convergence speed, but comes with additional costs. In this work, we propose a more holistic, approximate normalization via simple scalar multiplications motivated by the tight concentration of the norms of high-dimensional random vectors. Additionally, instead of applying strict normalization for the parameters, we constrain their norms. These modifications remove the need for weight decay and learning rate warm-up as well, but do not increase the total number of normalization layers. Our experiments with transformer architectures show up to 40% faster convergence compared to GPT models with QK normalization, with only 3% additional runtime cost. When deriving scaling laws, we found that our method enables training with larger batch sizes while preserving the favorable scaling characteristics of classic GPT architectures.

In studies of transferable learning, scaling laws are obtained for various important foundation models to predict their properties and performance at larger scales. Taking language-vision learning as example, we show here how scaling law derivation can also be used for model and dataset comparison, allowing to decide which procedure is to be preferred for pre-training. Full scaling laws based on dense measurements across a wide span of model and samples seen scales are derived for two important language-vision learning procedures, CLIP and MaMMUT, that use either contrastive only or contrastive and captioning text generative loss. For the first time, we use derived scaling laws to compare both models and three open datasets, DataComp-1.4B, Re-LAION-1.4B and DFN-1.4B, while ensuring sufficient prediction accuracy on held out points. From comparison, we obtain evidence for (i) MaMMUT's stronger improvement with scale and better sample efficiency than standard CLIP (ii) DFN-1.4B outperforming other open datasets. To strengthen validity of the comparison, we show scaling laws for various downstream tasks, classification, retrieval, and segmentation, observing consistently the same scaling trends for models and datasets across tasks. We show that comparison can also be performed when deriving scaling laws with a constant learning rate schedule, reducing compute cost. Accurate derivation of scaling laws provides thus means to perform model and dataset comparison on aligned common compute axis across large scale span, avoiding misleading conclusions based on measurements from few isolated single reference scales only. This paves road for guided collective improvement of open foundation models and training datasets, as scaling law based comparisons from various studies executed in common frame can be combined to identify overall better procedures. We release all the pre-trained models with their intermediate checkpoints, including openMaMMUT-L/14, which achieves 80.3% zero-shot ImageNet-1k accuracy, trained on 12.8B samples from DataComp-1.4B. Code for reproducing experiments in the paper and raw experiments data can be found at https://github.com/LAION-AI/scaling-laws-for-comparison.

We introduce DataComp for Language Models, a testbed for controlled dataset experiments with the goal of improving language models.As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations.Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing atmodel scales ranging from 412M to 7B parameters.As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set.The resulting dataset, DCLM-Baseline, enables training a 7B parameter language model from scratch to 63% 5-shot accuracy on MMLU with 2T training tokens.Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6 percentage point improvement on MMLU while being trained with half the compute.Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation. We release the \dclm benchmark, framework, models, and datasets at https://www.datacomp.ai/dclm/

Kaplan et al. and Hoffmann et al. developed influential scaling laws for the optimal model size as a function of the compute budget, but these laws yield substantially different predictions. We explain the discrepancy by reproducing the Kaplan scaling law on two datasets (OpenWebText2 and RefinedWeb) and identifying three factors causing the difference: last layer computational cost, warmup duration, and scale-dependent optimizer tuning. With these factors corrected, we obtain excellent agreement with the Hoffmann et al. (i.e., "Chinchilla") scaling law. Counter to a hypothesis of Hoffmann et al., we find that careful learning rate decay is not essential for the validity of their scaling law. As a secondary result, we derive scaling laws for the optimal learning rate and batch size, finding that tuning the AdamW $\beta_2$ parameter is essential at lower batch sizes.

Benchmarks are essential in the design of modern HPC installations, as they define key aspects of system components. Beyond synthetic workloads, it is crucial to include real applications that represent user requirements into benchmark suites, to guarantee high usability and widespread adoption of a new system. Given the significant investments in leadership-class supercomputers of the exascale era, this is even more important and necessitates alignment with a vision of Open Science and reproducibility. In this work, we present the JUPITER Benchmark Suite, which incorporates 16 applications from various domains. It was designed for and used in the procurement of JUPITER, the first European exascale supercomputer. We identify requirements and challenges and outline the project and software infrastructure setup. We provide descriptions and scalability studies of selected applications and a set of key takeaways. The JUPITER Benchmark Suite is released as open source software with this work at github.com/FZJ-JSC/jubench

Scaling up neural networks has led to remarkable performance across a wide range of tasks. Moreover, performance often follows reliable scaling laws as a function of training set size, model size, and compute, which offers valuable guidance as large-scale experiments are becoming increasingly expensive. However, previous work on scaling laws has primarily used private data & models or focused on uni-modal language or vision learning. To address these limitations, we investigate scaling laws for contrastive language-image pre-training (CLIP) with the public LAION dataset and the open-source OpenCLIP repository. Our large-scale experiments involve models trained on up to two billion image-text pairs and identify power law scaling for multiple downstream tasks including zero-shot classification, retrieval, linear probing, and end-to-end fine-tuning. We find that the training distribution plays a key role in scaling laws as the OpenAI and OpenCLIP models exhibit different scaling behavior despite identical model architectures and similar training recipes. We open-source our evaluation workflow and all models, including the largest public CLIP models, to ensure reproducibility and make scaling laws research more accessible. Source code and instructions to reproduce this study is available at https://github.eom/LAION-AI/sealing-laws-openelip.

Multimodal datasets are a critical component in recent breakthroughs such as CLIP, Stable Diffusion and GPT-4, yet their design does not receive the same research attention as model architectures or training algorithms. To address this shortcoming in the machine learning ecosystem, we introduce DataComp, a testbed for dataset experiments centered around a new candidate pool of 12.8 billion image-text pairs from Common Crawl. Participants in our benchmark design new filtering techniques or curate new data sources and then evaluate their new dataset by running our standardized CLIP training code and testing the resulting model on 38 downstream test sets. Our benchmark consists of multiple compute scales spanning four orders of magnitude, which enables the study of scaling trends and makes the benchmark accessible to researchers with varying resources. Our baseline experiments show that the DataComp workflow leads to better training sets. Our best baseline, DataComp-1B, enables training a CLIP ViT-L/14 from scratch to 79.2% zero-shot accuracy on ImageNet, outperforming OpenAI's CLIP ViT-L/14 by 3.7 percentage points while using the same training procedure and compute. We release \datanet and all accompanying code at www.datacomp.ai.

Increasing model, data and compute budget scale in the pre-training has been shown to strongly improve model generalization and transfer learning in vast line of work done in language modeling and natural image recognition. However, most studies on the positive effect of larger scale were done in scope of in-domain setting, with source and target data being in close proximity. To study effect of larger scale for both in-domain and out-of-domain setting when performing full and few-shot transfer, we combine here for the first time large, openly available medical X-Ray chest imaging datasets to reach a scale for medical imaging domain comparable to ImageNet-1k, routinely used for pre-training in natural image domain. We then conduct supervised pre-training, while varying network size and source data scale and domain, being either large natural (ImageNet-1k/21k) or large medical chest X-Ray datasets, and transfer pre-trained models to different natural or medical targets11Repository: https://github.com/SLAMPAI/large-scale-pretraining-transfer, We observe strong improvement due to larger pre-training scale for intra-domain natural-natural and medical-medical transfer. For inter-domain natural-medical transfer, we find improvements due to larger pre-training scale on larger X-Ray targets in full shot regime, while for smaller targets and for few-shot regime the improvement is not visible. Remarkably, large networks pre-trained on very large natural ImageNet-21k are as good or better than networks pre-trained on largest available medical X-Ray data when performing transfer to large X-Ray targets. We conclude that substantially increasing model and generic, medical domain-agnostic natural image source data scale in the pre-training can enable high quality out-of-domain transfer to medical domain specific targets, removing dependency on large medical domain-specific source data often not available in the practice.

Groundbreaking language-vision architectures like CLIP and DALL-E proved the utility of training on large amounts of noisy image-text data, without relying on expensive accurate labels used in standard vision unimodal supervised learning. The resulting models showed capabilities of strong text-guided image generation and transfer to downstream tasks, while performing remarkably at zero-shot classification with noteworthy out-of-distribution robustness. Since then, large-scale language-vision models like ALIGN, BASIC, GLIDE, Flamingo and Imagen made further improvements. Studying the training and capabilities of such models requires datasets containing billions of image-text pairs. Until now, no datasets of this size have been made openly available for the broader research community. To address this problem and democratize research on large-scale multi-modal models, we present LAION-5B - a dataset consisting of 5.85 billion CLIP-filtered image-text pairs, of which 2.32B contain English language. We show successful replication and fine-tuning of foundational models like CLIP, GLIDE and Stable Diffusion using the dataset, and discuss further experiments enabled with an openly available dataset of this scale. Additionally we provide several nearest neighbor indices, an improved web-interface for dataset exploration and subset generation, and detection scores for watermark, NSFW, and toxic content detection.

The Obstacle Tower Challenge is the task to master a procedurally generated chain of levels that subsequently get harder to complete. Whereas the most top performing entries of last year's competition used human demonstrations or reward shaping to learn how to cope with the challenge, we present an approach that performed competitively (placed 7th) but starts completely from scratch by means of Deep Reinforcement Learning with a relatively simple feed-forward deep network structure. We especially look at the generalization performance of the taken approach concerning different seeds and various visual themes that have become available after the competition, and investigate where the agent fails and why. Note that our approach does not possess a short-term memory like employing recurrent hidden states. With this work, we hope to contribute to a better understanding of what is possible with a relatively simple, flexible solution that can be applied to learning in environments featuring complex 3D visual input where the abstract task structure itself is still fairly simple.

Similarly to other scientific domains, Deep Learning (DL) holds great promises to fulfil the challenging needs of Remote Sensing (RS) applications. However, the increase in volume, variety and complexity of acquisitions that are carried out on a daily basis by Earth Observation (EO) missions generates new processing and storage challenges within operational processing pipelines. The aim of this work is to show that High-Performance Computing (HPC) systems can speed up the training time of Convolutional Neural Networks (CNNs). Particular attention is put on the monitoring of the classification accuracy that usually degrades when using large batch sizes. The experimental results of this work show that the training of the model scales up to a batch size of 8,000, obtaining classification performances in terms of accuracy in line with those using smaller batch sizes.

This work proposes a novel distributed deep learning model for Remote Sensing (RS) images super-resolution. High Performance Computing (HPC) systems with GPUs are used to accelerate the learning of the unknown low to high resolution mapping from large volumes of Sentinel-2 data. The proposed deep learning model is based on self-attention mechanism and residual learning. The results demonstrate that state-of-the-art performance can be achieved by keeping the size of the model relatively small. Synchronous data parallelism is applied to scale up the training process without severe performance loss. Distributed training is thus shown to speed up learning substantially while keeping performance intact.

Despite the vast amount of experimental findings on the role of the basal ganglia in reinforcement learning, there is still general lack of network models that use spiking neurons and plausible plasticity mechanisms to demonstrate network-level reward-based learning. In this work we extend a recent spiking actor-critic network model of the basal ganglia, aiming to create a minimal realistic model of learning from both positive and negative rewards. We hypothesize and implement in the model segregation of not only the dorsal striatum, but also of the ventral striatum into populations of medium spiny neurons (MSNs) that carry either D1 or D2 dopamine (DA) receptor type. This segregation allows explicit representation of both positive and negative expected reward within respective population. In line with recent experiments, we further assume that D1 and D2 MSN populations have distinct, opposing DA-modulated bidirectional synaptic plasticity. We implement the spiking network model in the simulator NEST and conduct experiments involving application of delayed rewards in a grid world setting, where a moving agent has to reach a goal state while maximizing the total obtained reward. We demonstrate that the network can learn not only to approach the positive rewards, but also to consequently avoid punishments as opposed to the original model. The spiking network model highlights thus functional role of D1-D2 MSN segregation within striatum and explains necessity for reversed direction of DA-dependent plasticity found at synapses converging on different types of striatal MSNs.

Recently, experience-driven unsupervised learning was shown to create combinatorial parts-based representations in a model of hierarchical visual memory. Examining the memory's ability to recognize persons from a database of natural face images, we show that an off-line, sleep-like operating regime of the memory domain results in a significant improvement of the system's ability to generalize over novel face views. Surprisingly, the positive effect turns out to be independent of synapse-specific plasticity, relying entirely on a homeostatic mechanism equalizing the intrinsic excitability levels of the units within the memory network. We show that this excitability equalization is the main cause for the improvement of memory function. A possible relation to cortical off-line memory reprocessing during certain sleep stages is discussed.