Ari Morcos

Randall Balestriero, Mark Ibrahim, Vlad Sobal et al. · meta-ai

Self-supervised learning, dubbed the dark matter of intelligence, is a promising path to advance machine learning. Yet, much like cooking, training SSL methods is a delicate art with a high barrier to entry. While many components are familiar, successfully training a SSL method involves a dizzying set of choices from the pretext tasks to training hyper-parameters. Our goal is to lower the barrier to entry into SSL research by laying the foundations and latest SSL recipes in the style of a cookbook. We hope to empower the curious researcher to navigate the terrain of methods, understand the role of the various knobs, and gain the know-how required to explore how delicious SSL can be.

Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer et al. · uw

We introduce new methods for 1) accelerating and 2) stabilizing training for large language-vision models. 1) For acceleration, we introduce SwitchBack, a linear layer for int8 quantized training which provides a speed-up of 13-25% while matching the performance of bfloat16 training within 0.1 percentage points for the 1B parameter CLIP ViT-Huge -- the largest int8 training to date. Our main focus is int8 as GPU support for float8 is rare, though we also analyze float8 training through simulation. While SwitchBack proves effective for float8, we show that standard techniques are also successful if the network is trained and initialized so that large feature magnitudes are discouraged, which we accomplish via layer-scale initialized with zeros. 2) For stability, we analyze loss spikes and find they consistently occur 1-8 iterations after the squared gradients become under-estimated by their AdamW second moment estimator. As a result, we recommend an AdamW-Adafactor hybrid which avoids loss spikes when training a CLIP ViT-Huge model and outperforms gradient clipping at the scales we test.

Christina Baek, Ricardo Pio Monti, David Schwab et al.

Real-world model deployments demand strong performance on narrow domains where data is often scarce. Typically, practitioners finetune models to specialize them, but this risks overfitting to the domain and forgetting general knowledge. We study a simple strategy, specialized pretraining (SPT), where a small domain dataset, typically reserved for finetuning, is repeated starting from pretraining as a fraction of the total tokens. Across three specialized domains (ChemPile, MusicPile, and ProofPile), SPT improves domain performance and preserves general capabilities after finetuning compared to standard pretraining. In our experiments, SPT reduces the pretraining tokens needed to reach a given domain performance by up to 1.75x. These gains grow when the target domain is underrepresented in the pretraining corpus: on domains far from web text, a 1B SPT model outperforms a 3B standard pretrained model. Beyond these empirical gains, we derive overfitting scaling laws to guide practitioners in selecting the optimal domain-data repetition for a given pretraining compute budget. Our observations reveal the finetuner's fallacy: while finetuning may appear to be the cheapest path to domain adaptation, introducing specialized domain data during pretraining stretches its utility. SPT yields better specialized domain performance (via reduced overfitting across repeated exposures) and better general domain performance (via reduced forgetting during finetuning), ultimately achieving stronger results with fewer parameters and less total compute when amortized over inference. To get the most out of domain data, incorporate it as early in training as possible.

Shashank Shekhar, Florian Bordes, Pascal Vincent et al.

Joint-embedding based learning (e.g., SimCLR, MoCo, DINO) and reconstruction-based learning (e.g., BEiT, SimMIM, MAE) are the two leading paradigms for self-supervised learning of vision transformers, but they differ substantially in their transfer performance. Here, we aim to explain these differences by analyzing the impact of these objectives on the structure and transferability of the learned representations. Our analysis reveals that reconstruction-based learning features are significantly dissimilar to joint-embedding based learning features and that models trained with similar objectives learn similar features even across architectures. These differences arise early in the network and are primarily driven by attention and normalization layers. We find that joint-embedding features yield better linear probe transfer for classification because the different objectives drive different distributions of information and invariances in the learned representation. These differences explain opposite trends in transfer performance for downstream tasks that require spatial specificity in features. Finally, we address how fine-tuning changes reconstructive representations to enable better transfer, showing that fine-tuning re-organizes the information to be more similar to pre-trained joint embedding models.

DatologyAI, Luke Merrick, Alex Fang et al.

Frontier language model quality increasingly hinges on our ability to organize web-scale text corpora for training. Today's dominant tools trade off speed and flexibility: lexical classifiers (e.g., FastText) are fast but limited to producing classification output scores, while the vector-valued outputs of transformer text embedding models flexibly support numerous workflows (e.g., clustering, classification, and retrieval) but are computationally expensive to produce. We introduce Luxical, a library for high-speed "lexical-dense" text embeddings that aims to recover the best properties of both approaches for web-scale text organization. Luxical combines sparse TF--IDF features, a small ReLU network, and a knowledge distillation training regimen to approximate large transformer embedding models at a fraction of their operational cost. In this technical report, we describe the Luxical architecture and training objective and evaluate a concrete Luxical model in two disparate applications: a targeted webcrawl document retrieval test and an end-to-end language model data curation task grounded in text classification. In these tasks we demonstrate speedups ranging from 3x to 100x over varying-sized neural baselines, and comparable to FastText model inference during the data curation task. On these evaluations, the tested Luxical model illustrates favorable compute/quality trade-offs for large-scale text organization, matching the quality of neural baselines. Luxical is available as open-source software at https://github.com/datologyai/luxical.

Mark Ibrahim, Quentin Garrido, Ari Morcos et al.

Recent state-of-the-art vision models introduced new architectures, learning paradigms, and larger pretraining data, leading to impressive performance on tasks such as classification. While previous generations of vision models were shown to lack robustness to factors such as pose, it's unclear the extent to which this next generation of models are more robust. To study this question, we develop a dataset of more than 7 million images with controlled changes in pose, position, background, lighting, and size. We study not only how robust recent state-of-the-art models are, but also the extent to which models can generalize variation in factors when they're present during training. We consider a catalog of recent vision models, including vision transformers (ViT), self-supervised models such as masked autoencoders (MAE), and models trained on larger datasets such as CLIP. We find out-of-the-box, even today's best models are not robust to common changes in pose, size, and background. When some samples varied during training, we found models required a significant portion of diversity to generalize -- though eventually robustness did improve. When diversity is only seen for some classes however, we found models did not generalize to other classes, unless the classes were very similar to those seen varying during training. We hope our work will shed further light on the blind spots of SoTA models and spur the development of more robust vision models.

Mark Ibrahim, Diane Bouchacourt, Ari Morcos

Deep learning has led to remarkable advances in computer vision. Even so, today's best models are brittle when presented with variations that differ even slightly from those seen during training. Minor shifts in the pose, color, or illumination of an object can lead to catastrophic misclassifications. State-of-the art models struggle to understand how a set of variations can affect different objects. We propose a framework for instilling a notion of how objects vary in more realistic settings. Our approach applies the formalism of Lie groups to capture continuous transformations to improve models' robustness to distributional shifts. We apply our framework on top of state-of-the-art self-supervised learning (SSL) models, finding that explicitly modeling transformations with Lie groups leads to substantial performance gains of greater than 10% for MAE on both known instances seen in typical poses now presented in new poses, and on unknown instances in any pose. We also apply our approach to ImageNet, finding that the Lie operator improves performance by almost 4%. These results demonstrate the promise of learning transformations to improve model robustness.

Anas Mahmoud, Mostafa Elhoushi, Amro Abbas et al.

Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning. We argue that this approach suffers from multiple limitations including: false positives and negatives due to CLIP's pretraining on noisy labels. We propose a pruning signal, Sieve, that employs synthetic captions generated by image-captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on unlabeled text corpus. Using DataComp, a multimodal dataset filtering benchmark, when evaluating on 38 downstream tasks, our pruning approach, surpasses CLIPScore by 2.6\% and 1.7\% on medium and large scale respectively. In addition, on retrieval tasks, Sieve leads to a significant improvement of 2.7% and 4.5% on medium and large scale respectively.

DatologyAI, Aldo Gael Carranza, Kaleigh Mentzer et al.

Multilinguality is a core capability for modern foundation models, yet training high-quality multilingual models remains challenging due to uneven data availability across languages. A further challenge is the performance interference that can arise from joint multilingual training, commonly referred to as the "curse of multilinguality". We study multilingual data curation across thirteen languages and find that many reported regressions are not inherent to multilingual scaling but instead stem from correctable deficiencies in data quality and composition rather than fundamental capacity limits. In controlled bilingual experiments, improving data quality for any single language benefits others: curating English improves non-English performance in 12 of 13 languages, while curating non-English yields reciprocal improvements in English. Bespoke per-language curation produces substantially larger within-language improvements. Extending these findings to large-scale general-purpose training mixtures, we show that curated multilingual allocations comprising under 8% of total tokens remain remarkably effective. We operationalize this approach within an effort that produced a 20T-token pretraining corpus derived entirely from public sources. Models with 3B and 8B parameters trained on a 1T-token random subset achieve competitive multilingual accuracy with 4-10x fewer training FLOPs than strong public baselines, establishing a new Pareto frontier in multilingual performance versus compute. Moreover, these benefits extend to frontier model scale: the 20T-token corpus served as part of the pretraining dataset for Trinity Large (400B/A13B), which exhibits strong multilingual performance relative to its training FLOPs. These results show that targeted, per-language data curation mitigates multilingual interference and enables compute-efficient multilingual scaling.

DatologyAI, Siddharth Joshi, Haoli Yin et al.

Empirical evaluation serves as the primary compass guiding research progress in foundation models. Despite a large body of work focused on training frontier vision-language models (VLMs), approaches to their evaluation remain nascent. To guide their maturation, we propose three desiderata that evaluations should satisfy: (1) faithfulness to the modality and application, (2) discriminability between models of varying quality, and (3) efficiency in compute. Through this lens, we identify critical failure modes that violate faithfulness and discriminability, misrepresenting model capabilities: (i) multiple-choice formats reward guessing, poorly reflect downstream use cases, and saturate early as models improve; (ii) blindly solvable questions, which can be answered without images, constitute up to 70% of some evaluations; and (iii) mislabeled or ambiguous samples compromise up to 42% of examples in certain datasets. Regarding efficiency, the computational burden of evaluating frontier models has become prohibitive: by some accounts, nearly 20% of development compute is devoted to evaluation alone. Rather than discarding existing benchmarks, we curate them via transformation and filtering to maximize fidelity and discriminability. We find that converting multiple-choice questions to generative tasks reveals sharp capability drops of up to 35%. In addition, filtering blindly solvable and mislabeled samples improves discriminative power while simultaneously reducing computational cost. We release DatBench-Full, a cleaned evaluation suite of 33 datasets spanning nine VLM capabilities, and DatBench, a discriminative subset that achieves 13x average speedup (up to 50x) while closely matching the discriminative power of the original datasets. Our work outlines a path toward evaluation practices that are both rigorous and sustainable as VLMs continue to scale.

Siddharth Joshi, Haoli Yin, Rishabh Adiga et al.

Data curation has shifted the quality-compute frontier for language-model and contrastive image-text pretraining, but its role for vision-language models (VLMs) is far less established. We ask how far data curation alone can take VLM performance, holding architecture, training recipe, and compute fixed and varying only the training data. Our pipeline, applied to the MAmmoTH-VL single-image subset, lifts performance by +11.7pp on average across 20 public VLM benchmarks (spanning grounding, VQA, OCR/documents, captioning, spatial/3D, counting, charts, math, brand-ID, and multi-image reasoning) and by +11.3pp on average across all nine capability axes of DatBench, our high-fidelity VLM eval suite. At 2B, our curated model surpasses InternVL3.5-2B by 9.9pp at ~17x less training compute and closes the gap to Qwen3-VL-2B to within 1.8pp at ~87x less compute, from pretraining alone. Beyond accuracy, curation delivers four further properties: (1) Reliability: per-capability std across training seeds drops by ~67% and the lift survives a 4k-to-16k context-length sweep; (2) OOD generalization: the 9-eval OOD average rises by +7.2pp, and multi-image BLINK rises by +3.09pp despite single-image-only training, with Visual Correspondence gaining +11.8pp; (3) Behavioral gains beyond benchmarks: across ~1,100 open-ended queries the curated 2B is more honest and more specific than the matched-compute baseline, and more concise and less refusal-prone than a frontier 2B reference; (4) Pareto-dominance on inference cost: at every scale (1B, 2B, 4B) the curated model raises accuracy while lowering response FLOPs vs. the matched-compute baseline, and the curated 4B matches near-frontier accuracy at 3.3x lower response FLOPs than Qwen3-VL-4B. Data curation is a high-leverage tool for building better VLMs, reaching near-frontier accuracy at up to ~150x less training compute.

Stéphane d'Ascoli, Hugo Touvron, Matthew Leavitt et al.

Convolutional architectures have proven extremely successful for vision tasks. Their hard inductive biases enable sample-efficient learning, but come at the cost of a potentially lower performance ceiling. Vision Transformers (ViTs) rely on more flexible self-attention layers, and have recently outperformed CNNs for image classification. However, they require costly pre-training on large external datasets or distillation from pre-trained convolutional networks. In this paper, we ask the following question: is it possible to combine the strengths of these two architectures while avoiding their respective limitations? To this end, we introduce gated positional self-attention (GPSA), a form of positional self-attention which can be equipped with a ``soft" convolutional inductive bias. We initialise the GPSA layers to mimic the locality of convolutional layers, then give each attention head the freedom to escape locality by adjusting a gating parameter regulating the attention paid to position versus content information. The resulting convolutional-like ViT architecture, ConViT, outperforms the DeiT on ImageNet, while offering a much improved sample efficiency. We further investigate the role of locality in learning by first quantifying how it is encouraged in vanilla self-attention layers, then analysing how it is escaped in GPSA layers. We conclude by presenting various ablations to better understand the success of the ConViT. Our code and models are released publicly at https://github.com/facebookresearch/convit.

DatologyAI, Pratyush Maini, Vineeth Dorna et al.

Recent advances in large language model (LLM) pretraining have shown that simply scaling data quantity eventually leads to diminishing returns, hitting a data wall. In response, the use of synthetic data for pretraining has emerged as a promising paradigm for pushing the frontier of performance. Despite this, the factors affecting synthetic data quality remain poorly understood. In this work, we introduce BeyondWeb, a synthetic data generation framework that produces high-quality synthetic data for pretraining. BeyondWeb significantly extends the capabilities of traditional web-scale datasets, outperforming state-of-the-art synthetic pretraining datasets such as Cosmopedia and Nemotron-CC's high-quality synthetic subset (Nemotron-Synth) by up to 5.1 percentage points (pp) and 2.6pp, respectively, when averaged across a suite of 14 benchmark evaluations. It delivers up to 7.7x faster training than open web data and 2.7x faster than Nemotron-Synth. Remarkably, a 3B model trained for 180B tokens on BeyondWeb outperforms an 8B model trained for the same token budget on Cosmopedia. We also present several insights from BeyondWeb on synthetic data for pretraining: what drives its benefits, which data to rephrase and how, and the impact of model size and family on data quality. Overall, our work shows that there's no silver bullet for generating high-quality synthetic pretraining data. The best outcomes require jointly optimizing many factors, a challenging task that requires rigorous science and practical expertise. Naive approaches can yield modest improvements, potentially at great cost, while well-executed methods can yield transformative improvements, as exemplified by BeyondWeb.

Jose Javier Gonzalez Ortiz, Jonathan Frankle, Mike Rabbat et al.

As datasets and models become increasingly large, distributed training has become a necessary component to allow deep neural networks to train in reasonable amounts of time. However, distributed training can have substantial communication overhead that hinders its scalability. One strategy for reducing this overhead is to perform multiple unsynchronized SGD steps independently on each worker between synchronization steps, a technique known as local SGD. We conduct a comprehensive empirical study of local SGD and related methods on a large-scale image classification task. We find that performing local SGD comes at a price: lower communication costs (and thereby faster training) are accompanied by lower accuracy. This finding is in contrast from the smaller-scale experiments in prior work, suggesting that local SGD encounters challenges at scale. We further show that incorporating the slow momentum framework of Wang et al. (2020) consistently improves accuracy without requiring additional communication, hinting at future directions for potentially escaping this trade-off.

Stéphane d'Ascoli, Levent Sagun, Giulio Biroli et al.

Vision Transformers (ViT) have recently emerged as a powerful alternative to convolutional networks (CNNs). Although hybrid models attempt to bridge the gap between these two architectures, the self-attention layers they rely on induce a strong computational bottleneck, especially at large spatial resolutions. In this work, we explore the idea of reducing the time spent training these layers by initializing them as convolutional layers. This enables us to transition smoothly from any pre-trained CNN to its functionally identical hybrid model, called Transformed CNN (T-CNN). With only 50 epochs of fine-tuning, the resulting T-CNNs demonstrate significant performance gains over the CNN (+2.2% top-1 on ImageNet-1k for a ResNet50-RS) as well as substantially improved robustness (+11% top-1 on ImageNet-C). We analyze the representations learnt by the T-CNN, providing deeper insights into the fruitful interplay between convolutions and self-attention. Finally, we experiment initializing the T-CNN from a partially trained CNN, and find that it reaches better performance than the corresponding hybrid model trained from scratch, while reducing training time.

Matthew L. Leavitt, Ari Morcos

Methods for understanding the decisions of and mechanisms underlying deep neural networks (DNNs) typically rely on building intuition by emphasizing sensory or semantic features of individual examples. For instance, methods aim to visualize the components of an input which are "important" to a network's decision, or to measure the semantic properties of single neurons. Here, we argue that interpretability research suffers from an over-reliance on intuition-based approaches that risk-and in some cases have caused-illusory progress and misleading conclusions. We identify a set of limitations that we argue impede meaningful progress in interpretability research, and examine two popular classes of interpretability methods-saliency and single-neuron-based approaches-that serve as case studies for how overreliance on intuition and lack of falsifiability can undermine interpretability research. To address these concerns, we propose a strategy to address these impediments in the form of a framework for strongly falsifiable interpretability research. We encourage researchers to use their intuitions as a starting point to develop and test clear, falsifiable hypotheses, and hope that our framework yields robust, evidence-based interpretability methods that generate meaningful advances in our understanding of DNNs.

Matthew L. Leavitt, Ari Morcos

Representational sparsity is known to affect robustness to input perturbations in deep neural networks (DNNs), but less is known about how the semantic content of representations affects robustness. Class selectivity-the variability of a unit's responses across data classes or dimensions-is one way of quantifying the sparsity of semantic representations. Given recent evidence that class selectivity may not be necessary for, and in some cases can impair generalization, we investigate whether it also confers robustness (or vulnerability) to perturbations of input data. We found that networks regularized to have lower levels of class selectivity were more robust to average-case (naturalistic) perturbations, while networks with higher class selectivity are more vulnerable. In contrast, class selectivity increases robustness to multiple types of worst-case (i.e. white box adversarial) perturbations, suggesting that while decreasing class selectivity is helpful for average-case perturbations, it is harmful for worst-case perturbations. To explain this difference, we studied the dimensionality of the networks' representations: we found that the dimensionality of early-layer representations is inversely proportional to a network's class selectivity, and that adversarial samples cause a larger increase in early-layer dimensionality than corrupted samples. Furthermore, the input-unit gradient is more variable across samples and units in high-selectivity networks compared to low-selectivity networks. These results lead to the conclusion that units participate more consistently in low-selectivity regimes compared to high-selectivity regimes, effectively creating a larger attack surface and hence vulnerability to worst-case perturbations.

Ramakrishna Vedantam, Arthur Szlam, Maximilian Nickel et al.

Humans can learn and reason under substantial uncertainty in a space of infinitely many concepts, including structured relational concepts ("a scene with objects that have the same color") and ad-hoc categories defined through goals ("objects that could fall on one's head"). In contrast, standard classification benchmarks: 1) consider only a fixed set of category labels, 2) do not evaluate compositional concept learning and 3) do not explicitly capture a notion of reasoning under uncertainty. We introduce a new few-shot, meta-learning benchmark, Compositional Reasoning Under Uncertainty (CURI) to bridge this gap. CURI evaluates different aspects of productive and systematic generalization, including abstract understandings of disentangling, productive generalization, learning boolean operations, variable binding, etc. Importantly, it also defines a model-independent "compositionality gap" to evaluate the difficulty of generalizing out-of-distribution along each of these axes. Extensive evaluations across a range of modeling choices spanning different modalities (image, schemas, and sounds), splits, privileged auxiliary concept information, and choices of negatives reveal substantial scope for modeling advances on the proposed task. All code and datasets will be available online.

Erik Wijmans, Julian Straub, Dhruv Batra et al.

Recent advances in deep reinforcement learning require a large amount of training data and generally result in representations that are often over specialized to the target task. In this work, we present a methodology to study the underlying potential causes for this specialization. We use the recently proposed projection weighted Canonical Correlation Analysis (PWCCA) to measure the similarity of visual representations learned in the same environment by performing different tasks. We then leverage our proposed methodology to examine the task dependence of visual representations learned on related but distinct embodied navigation tasks. Surprisingly, we find that slight differences in task have no measurable effect on the visual representation for both SqueezeNet and ResNet architectures. We then empirically demonstrate that visual representations learned on one task can be effectively transferred to a different task.

Matthew L. Leavitt, Ari Morcos

The properties of individual neurons are often analyzed in order to understand the biological and artificial neural networks in which they're embedded. Class selectivity-typically defined as how different a neuron's responses are across different classes of stimuli or data samples-is commonly used for this purpose. However, it remains an open question whether it is necessary and/or sufficient for deep neural networks (DNNs) to learn class selectivity in individual units. We investigated the causal impact of class selectivity on network function by directly regularizing for or against class selectivity. Using this regularizer to reduce class selectivity across units in convolutional neural networks increased test accuracy by over 2% for ResNet18 trained on Tiny ImageNet. For ResNet20 trained on CIFAR10 we could reduce class selectivity by a factor of 2.5 with no impact on test accuracy, and reduce it nearly to zero with only a small ($\sim$2%) drop in test accuracy. In contrast, regularizing to increase class selectivity significantly decreased test accuracy across all models and datasets. These results indicate that class selectivity in individual units is neither sufficient nor strictly necessary, and can even impair DNN performance. They also encourage caution when focusing on the properties of single units as representative of the mechanisms by which DNNs function.

Or Litany, Ari Morcos, Srinath Sridhar et al.

We seek to learn a representation on a large annotated data source that generalizes to a target domain using limited new supervision. Many prior approaches to this problem have focused on learning "disentangled" representations so that as individual factors vary in a new domain, only a portion of the representation need be updated. In this work, we seek the generalization power of disentangled representations, but relax the requirement of explicit latent disentanglement and instead encourage linearity of individual factors of variation by requiring them to be manipulable by learned linear transformations. We dub these transformations latent canonicalizers, as they aim to modify the value of a factor to a pre-determined (but arbitrary) canonical value (e.g., recoloring the image foreground to black). Assuming a source domain with access to meta-labels specifying the factors of variation within an image, we demonstrate experimentally that our method helps reduce the number of observations needed to generalize to a similar target domain when compared to a number of supervised baselines.

Mathilde Caron, Ari Morcos, Piotr Bojanowski et al.

Convolutional neural networks trained without supervision come close to matching performance with supervised pre-training, but sometimes at the cost of an even higher number of parameters. Extracting subnetworks from these large unsupervised convnets with preserved performance is of particular interest to make them less computationally intensive. Typical pruning methods operate during training on a task while trying to maintain the performance of the pruned network on the same task. However, in self-supervised feature learning, the training objective is agnostic on the representation transferability to downstream tasks. Thus, preserving performance for this objective does not ensure that the pruned subnetwork remains effective for solving downstream tasks. In this work, we investigate the use of standard pruning methods, developed primarily for supervised learning, for networks trained without labels (i.e. on self-supervised tasks). We show that pruned masks obtained with or without labels reach comparable performance when re-trained on labels, suggesting that pruning operates similarly for self-supervised and supervised learning. Interestingly, we also find that pruning preserves the transfer performance of self-supervised subnetwork representations.

Erik Wijmans, Abhishek Kadian, Ari Morcos et al.

We present Decentralized Distributed Proximal Policy Optimization (DD-PPO), a method for distributed reinforcement learning in resource-intensive simulated environments. DD-PPO is distributed (uses multiple machines), decentralized (lacks a centralized server), and synchronous (no computation is ever stale), making it conceptually simple and easy to implement. In our experiments on training virtual robots to navigate in Habitat-Sim, DD-PPO exhibits near-linear scaling -- achieving a speedup of 107x on 128 GPUs over a serial implementation. We leverage this scaling to train an agent for 2.5 Billion steps of experience (the equivalent of 80 years of human experience) -- over 6 months of GPU-time training in under 3 days of wall-clock time with 64 GPUs. This massive-scale training not only sets the state of art on Habitat Autonomous Navigation Challenge 2019, but essentially solves the task --near-perfect autonomous navigation in an unseen environment without access to a map, directly from an RGB-D camera and a GPS+Compass sensor. Fortuitously, error vs computation exhibits a power-law-like distribution; thus, 90% of peak performance is obtained relatively early (at 100 million steps) and relatively cheaply (under 1 day with 8 GPUs). Finally, we show that the scene understanding and navigation policies learned can be transferred to other navigation tasks -- the analog of ImageNet pre-training + task-specific fine-tuning for embodied AI. Our model outperforms ImageNet pre-trained CNNs on these transfer tasks and can serve as a universal resource (all models and code are publicly available).

Yuandong Tian, Tina Jiang, Qucheng Gong et al.

We analyze the dynamics of training deep ReLU networks and their implications on generalization capability. Using a teacher-student setting, we discovered a novel relationship between the gradient received by hidden student nodes and the activations of teacher nodes for deep ReLU networks. With this relationship and the assumption of small overlapping teacher node activations, we prove that (1) student nodes whose weights are initialized to be close to teacher nodes converge to them at a faster rate, and (2) in over-parameterized regimes and 2-layer case, while a small set of lucky nodes do converge to the teacher nodes, the fan-out weights of other nodes converge to zero. This framework provides insight into multiple puzzling phenomena in deep learning like over-parameterization, implicit regularization, lottery tickets, etc. We verify our assumption by showing that the majority of BatchNorm biases of pre-trained VGG11/16 models are negative. Experiments on (1) random deep teacher networks with Gaussian inputs, (2) teacher network pre-trained on CIFAR-10 and (3) extensive ablation studies validate our multiple theoretical predictions.