Michael Maire

Yuhan Liu, Hanchen Li, Yihua Cheng et al. · stanford

As large language models (LLMs) take on complex tasks, their inputs are supplemented with longer contexts that incorporate domain knowledge. Yet using long contexts is challenging, as nothing can be generated until the whole context is processed by the LLM. While the context-processing delay can be reduced by reusing the KV cache of a context across different inputs, fetching the KV cache, which contains large tensors, over the network can cause high extra network delays. CacheGen is a fast context-loading module for LLM systems. First, CacheGen uses a custom tensor encoder, leveraging KV cache's distributional properties to encode a KV cache into more compact bitstream representations with negligible decoding overhead, to save bandwidth usage. Second, CacheGen adapts the compression level of different parts of a KV cache to cope with changes in available bandwidth, in order to maintain low context-loading delay and high generation quality. % When available bandwidth drops, CacheGen may raise the compression level for a part of the context or recompute its KV cache on the fly. We test CacheGen on popular LLMs and datasets. Compared to the recent systems that reuse the KV cache, CacheGen reduces the KV cache size by 3.5-4.3x and the total delay in fetching and processing contexts by 3.2-3.7x with negligible impact on the LLM response quality. Our code is at: https://github.com/UChi-JCL/CacheGen.

Tewodros Ayalew, Matthew Jeung, Samuel Wheeler et al.

We introduce CLAW, a fully end-to-end self-supervised framework for learning a world model jointly with continuous latent action representations directly from action-free videos. Our approach leverages adversarial latent regularization and diffusion-based video generation to capture structured and semantically meaningful action representations while modeling rich, predictive environment dynamics, without relying on any action labels or annotations. By simultaneously training the Latent Action Model and world model, CLAW learns to reason about how inferred actions induce environment transitions from visual observations alone. We show that the resulting latent action world model supports both imitation learning from observation and goal-directed planning. In imitation learning, latent actions extracted from raw videos enable behavior cloning. For planning, CLAW generates sequences of latent actions and maps them to executable actions to reach desired goals. Extensive experiments across diverse tasks and embodiments demonstrate that CLAW produces semantically meaningful latent action representations, supports effective action transfer, and enables planning and imitation from observation, outperforming existing methods.

Xin Yuan, Pedro Savarese, Michael Maire

We develop an approach to efficiently grow neural networks, within which parameterization and optimization strategies are designed by considering their effects on the training dynamics. Unlike existing growing methods, which follow simple replication heuristics or utilize auxiliary gradient-based local optimization, we craft a parameterization scheme which dynamically stabilizes weight, activation, and gradient scaling as the architecture evolves, and maintains the inference functionality of the network. To address the optimization difficulty resulting from imbalanced training effort distributed to subnetworks fading in at different growth phases, we propose a learning rate adaption mechanism that rebalances the gradient contribution of these separate subcomponents. Experimental results show that our method achieves comparable or better accuracy than training large fixed-size models, while saving a substantial portion of the original computation budget for training. We demonstrate that these gains translate into real wall-clock training speedups.

David Yunis, Kumar Kshitij Patel, Samuel Wheeler et al.

We propose an empirical approach centered on the spectral dynamics of weights -- the behavior of singular values and vectors during optimization -- to unify and clarify several phenomena in deep learning. We identify a consistent bias in optimization across various experiments, from small-scale ``grokking'' to large-scale tasks like image classification with ConvNets, image generation with UNets, speech recognition with LSTMs, and language modeling with Transformers. We also demonstrate that weight decay enhances this bias beyond its role as a norm regularizer, even in practical systems. Moreover, we show that these spectral dynamics distinguish memorizing networks from generalizing ones, offering a novel perspective on this longstanding conundrum. Additionally, we leverage spectral dynamics to explore the emergence of well-performing sparse subnetworks (lottery tickets) and the structure of the loss surface through linear mode connectivity. Our findings suggest that spectral dynamics provide a coherent framework to better understand the behavior of neural networks across diverse settings.

Yuhan Liu, Chengcheng Wan, Kuntai Du et al.

ML APIs have greatly relieved application developers of the burden to design and train their own neural network models -- classifying objects in an image can now be as simple as one line of Python code to call an API. However, these APIs offer the same pre-trained models regardless of how their output is used by different applications. This can be suboptimal as not all ML inference errors can cause application failures, and the distinction between inference errors that can or cannot cause failures varies greatly across applications. To tackle this problem, we first study 77 real-world applications, which collectively use six ML APIs from two providers, to reveal common patterns of how ML API output affects applications' decision processes. Inspired by the findings, we propose ChameleonAPI, an optimization framework for ML APIs, which takes effect without changing the application source code. ChameleonAPI provides application developers with a parser that automatically analyzes the application to produce an abstract of its decision process, which is then used to devise an application-specific loss function that only penalizes API output errors critical to the application. ChameleonAPI uses the loss function to efficiently train a neural network model customized for each application and deploys it to serve API invocations from the respective application via existing interface. Compared to a baseline that selects the best-of-all commercial ML API, we show that ChameleonAPI reduces incorrect application decisions by 43%.

Sudarshan Babu, Richard Liu, Avery Zhou et al.

We introduce HyperFields, a method for generating text-conditioned Neural Radiance Fields (NeRFs) with a single forward pass and (optionally) some fine-tuning. Key to our approach are: (i) a dynamic hypernetwork, which learns a smooth mapping from text token embeddings to the space of NeRFs; (ii) NeRF distillation training, which distills scenes encoded in individual NeRFs into one dynamic hypernetwork. These techniques enable a single network to fit over a hundred unique scenes. We further demonstrate that HyperFields learns a more general map between text and NeRFs, and consequently is capable of predicting novel in-distribution and out-of-distribution scenes -- either zero-shot or with a few finetuning steps. Finetuning HyperFields benefits from accelerated convergence thanks to the learned general map, and is capable of synthesizing novel scenes 5 to 10 times faster than existing neural optimization-based methods. Our ablation experiments show that both the dynamic architecture and NeRF distillation are critical to the expressivity of HyperFields.

Cyrus Zhou, Pedro Savarese, Zack Hassman et al.

Ultra-low-precision inference can sharply reduce memory and latency but often degrades accuracy and relies on specialized hardware. We present SONIQ, a system-optimized, noise-injected quantization framework that learns per-channel mixed precision for both weights and activations while training under the same rules used at inference. By injecting hardware-calibrated quantization noise during training, SONIQ steers models toward the discrete arithmetic used at deployment -- without bespoke runtimes. Across CNNs and Transformers, SONIQ achieves up to 16x and 7x compression, respectively, while matching or exceeding full-precision accuracy. Measured end-to-end, SONIQ delivers up to 7.3x CPU speedup over strong INT8 baselines and up to 6.3x (vector units) / 2.8x (tensor cores) GPU speedup relative to FP16. A practical outcome is that two per-channel precision levels -- one in the 1--4-bit range and one in the 4--8-bit range -- suffice in practice; at inference, each channel selects one of the two, keeping kernels simple and fast. To our knowledge, SONIQ is the first framework to reach or surpass full-precision accuracy under ultra-low (1--4 bits per parameter) regimes while remaining deployable on commodity hardware, narrowing the gap between quantization theory and practical, high-throughput inference.

Xin Yuan, Michael Maire

We develop a neural network architecture which, trained in an unsupervised manner as a denoising diffusion model, simultaneously learns to both generate and segment images. Learning is driven entirely by the denoising diffusion objective, without any annotation or prior knowledge about regions during training. A computational bottleneck, built into the neural architecture, encourages the denoising network to partition an input into regions, denoise them in parallel, and combine the results. Our trained model generates both synthetic images and, by simple examination of its internal predicted partitions, a semantic segmentation of those images. Without any finetuning, we directly apply our unsupervised model to the downstream task of segmenting real images via noising and subsequently denoising them. Experiments demonstrate that our model achieves accurate unsupervised image segmentation and high-quality synthetic image generation across multiple datasets.

Xiao Zhang, Michael Maire

Within the framework of generative adversarial networks (GANs), we propose objectives that task the discriminator for self-supervised representation learning via additional structural modeling responsibilities. In combination with an efficient smoothness regularizer imposed on the network, these objectives guide the discriminator to learn to extract informative representations, while maintaining a generator capable of sampling from the domain. Specifically, our objectives encourage the discriminator to structure features at two levels of granularity: aligning distribution characteristics, such as mean and variance, at coarse scales, and grouping features into local clusters at finer scales. Operating as a feature learner within the GAN framework frees our self-supervised system from the reliance on hand-crafted data augmentation schemes that are prevalent across contrastive representation learning methods. Across CIFAR-10/100 and an ImageNet subset, experiments demonstrate that equipping GANs with our self-supervised objectives suffices to produce discriminators which, evaluated in terms of representation learning, compete with networks trained by contrastive learning approaches.

Tri Huynh, Simon Kornblith, Matthew R. Walter et al.

Self-supervised representation learning has made significant leaps fueled by progress in contrastive learning, which seeks to learn transformations that embed positive input pairs nearby, while pushing negative pairs far apart. While positive pairs can be generated reliably (e.g., as different views of the same image), it is difficult to accurately establish negative pairs, defined as samples from different images regardless of their semantic content or visual features. A fundamental problem in contrastive learning is mitigating the effects of false negatives. Contrasting false negatives induces two critical issues in representation learning: discarding semantic information and slow convergence. In this paper, we propose novel approaches to identify false negatives, as well as two strategies to mitigate their effect, i.e. false negative elimination and attraction, while systematically performing rigorous evaluations to study this problem in detail. Our method exhibits consistent improvements over existing contrastive learning-based methods. Without labels, we identify false negatives with 40% accuracy among 1000 semantic classes on ImageNet, and achieve 5.8% absolute improvement in top-1 accuracy over the previous state-of-the-art when finetuning with 1% labels. Our code is available at https://github.com/google-research/fnc.

Haochen Wang, Ruotian Luo, Michael Maire et al.

The core of our approach, Pixel Consensus Voting, is a framework for instance segmentation based on the Generalized Hough transform. Pixels cast discretized, probabilistic votes for the likely regions that contain instance centroids. At the detected peaks that emerge in the voting heatmap, backprojection is applied to collect pixels and produce instance masks. Unlike a sliding window detector that densely enumerates object proposals, our method detects instances as a result of the consensus among pixel-wise votes. We implement vote aggregation and backprojection using native operators of a convolutional neural network. The discretization of centroid voting reduces the training of instance segmentation to pixel labeling, analogous and complementary to FCN-style semantic segmentation, leading to an efficient and unified architecture that jointly models things and stuff. We demonstrate the effectiveness of our pipeline on COCO and Cityscapes Panoptic Segmentation and obtain competitive results. Code will be open-sourced.

Xiao Zhang, Ruoxi Jiang, William Gao et al.

We show that introducing a weighting factor to reduce the influence of identity shortcuts in residual networks significantly enhances semantic feature learning in generative representation learning frameworks, such as masked autoencoders (MAEs) and diffusion models. Our modification notably improves feature quality, raising ImageNet-1K K-Nearest Neighbor accuracy from 27.4% to 63.9% and linear probing accuracy from 67.8% to 72.7% for MAEs with a ViT-B/16 backbone, while also enhancing generation quality in diffusion models. This significant gap suggests that, while residual connection structure serves an essential role in facilitating gradient propagation, it may have a harmful side effect of reducing capacity for abstract learning by virtue of injecting an echo of shallower representations into deeper layers. We ameliorate this downside via a fixed formula for monotonically decreasing the contribution of identity connections as layer depth increases. Our design promotes the gradual development of feature abstractions, without impacting network trainability. Analyzing the representations learned by our modified residual networks, we find correlation between low effective feature rank and downstream task performance.

Xiao Zhang, David Yunis, Michael Maire

We present an approach for analyzing grouping information contained within a neural network's activations, permitting extraction of spatial layout and semantic segmentation from the behavior of large pre-trained vision models. Unlike prior work, our method conducts a holistic analysis of a network's activation state, leveraging features from all layers and obviating the need to guess which part of the model contains relevant information. Motivated by classic spectral clustering, we formulate this analysis in terms of an optimization objective involving a set of affinity matrices, each formed by comparing features within a different layer. Solving this optimization problem using gradient descent allows our technique to scale from single images to dataset-level analysis, including, in the latter, both intra- and inter-image relationships. Analyzing a pre-trained generative transformer provides insight into the computational strategy learned by such models. Equating affinity with key-query similarity across attention layers yields eigenvectors encoding scene spatial layout, whereas defining affinity by value vector similarity yields eigenvectors encoding object identity. This result suggests that key and query vectors coordinate attentional information flow according to spatial proximity (a `where' pathway), while value vectors refine a semantic category representation (a `what' pathway).

Ruoxi Jiang, Xiao Zhang, Karan Jakhar et al.

Neural PDE solvers offer a powerful tool for modeling complex dynamical systems, but often struggle with error accumulation over long time horizons and maintaining stability and physical consistency. We introduce a multiscale implicit neural emulator that enhances long-term prediction accuracy by conditioning on a hierarchy of lower-dimensional future state representations. Drawing inspiration from the stability properties of numerical implicit time-stepping methods, our approach leverages predictions several steps ahead in time at increasing compression rates for next-timestep refinements. By actively adjusting the temporal downsampling ratios, our design enables the model to capture dynamics across multiple granularities and enforce long-range temporal coherence. Experiments on turbulent fluid dynamics show that our method achieves high short-term accuracy and produces long-term stable forecasts, significantly outperforming autoregressive baselines while adding minimal computational overhead.

Tewodros Ayalew, Xiao Zhang, Kevin Yuanbo Wu et al.

We present PROGRESSOR, a novel framework that learns a task-agnostic reward function from videos, enabling policy training through goal-conditioned reinforcement learning (RL) without manual supervision. Underlying this reward is an estimate of the distribution over task progress as a function of the current, initial, and goal observations that is learned in a self-supervised fashion. Crucially, PROGRESSOR refines rewards adversarially during online RL training by pushing back predictions for out-of-distribution observations, to mitigate distribution shift inherent in non-expert observations. Utilizing this progress prediction as a dense reward together with an adversarial push-back, we show that PROGRESSOR enables robots to learn complex behaviors without any external supervision. Pretrained on large-scale egocentric human video from EPIC-KITCHENS, PROGRESSOR requires no fine-tuning on in-domain task-specific data for generalization to real-robot offline RL under noisy demonstrations, outperforming contemporary methods that provide dense visual reward for robotic learning. Our findings highlight the potential of PROGRESSOR for scalable robotic applications where direct action labels and task-specific rewards are not readily available.

Xiao Zhang, Ruoxi Jiang, Rebecca Willett et al.

We introduce nested diffusion models, an efficient and powerful hierarchical generative framework that substantially enhances the generation quality of diffusion models, particularly for images of complex scenes. Our approach employs a series of diffusion models to progressively generate latent variables at different semantic levels. Each model in this series is conditioned on the output of the preceding higher-level models, culminating in image generation. Hierarchical latent variables guide the generation process along predefined semantic pathways, allowing our approach to capture intricate structural details while significantly improving image quality. To construct these latent variables, we leverage a pre-trained visual encoder, which learns strong semantic visual representations, and modulate its capacity via dimensionality reduction and noise injection. Across multiple datasets, our system demonstrates significant enhancements in image quality for both unconditional and class/text conditional generation. Moreover, our unconditional generation system substantially outperforms the baseline conditional system. These advancements incur minimal computational overhead as the more abstract levels of our hierarchy work with lower-dimensional representations.

Sudarshan Babu, Phillip Lo, Xiao Zhang et al.

We introduce HyperDiffusionFields (HyDiF), a framework that models 3D molecular conformers as continuous fields rather than discrete atomic coordinates or graphs. At the core of our approach is the Molecular Directional Field (MDF), a vector field that maps any point in space to the direction of the nearest atom of a particular type. We represent MDFs using molecule-specific neural implicit fields, which we call Molecular Neural Fields (MNFs). To enable learning across molecules and facilitate generalization, we adopt an approach where a shared hypernetwork, conditioned on a molecule, generates the weights of the given molecule's MNF. To endow the model with generative capabilities, we train the hypernetwork as a denoising diffusion model, enabling sampling in the function space of molecular fields. Our design naturally extends to a masked diffusion mechanism to support structure-conditioned generation tasks, such as molecular inpainting, by selectively noising regions of the field. Beyond generation, the localized and continuous nature of MDFs enables spatially fine-grained feature extraction for molecular property prediction, something not easily achievable with graph or point cloud based methods. Furthermore, we demonstrate that our approach scales to larger biomolecules, illustrating a promising direction for field-based molecular modeling.

Xin Yuan, Rana Hanocka, Michael Maire

We cast multiview reconstruction from unknown pose as a generative modeling problem. From a collection of unannotated 2D images of a scene, our approach simultaneously learns both a network to predict camera pose from 2D image input, as well as the parameters of a Neural Radiance Field (NeRF) for the 3D scene. To drive learning, we wrap both the pose prediction network and NeRF inside a Denoising Diffusion Probabilistic Model (DDPM) and train the system via the standard denoising objective. Our framework requires the system accomplish the task of denoising an input 2D image by predicting its pose and rendering the NeRF from that pose. Learning to denoise thus forces the system to concurrently learn the underlying 3D NeRF representation and a mapping from images to camera extrinsic parameters. To facilitate the latter, we design a custom network architecture to represent pose as a distribution, granting implicit capacity for discovering view correspondences when trained end-to-end for denoising alone. This technique allows our system to successfully build NeRFs, without pose knowledge, for challenging scenes where competing methods fail. At the conclusion of training, our learned NeRF can be extracted and used as a 3D scene model; our full system can be used to sample novel camera poses and generate novel-view images.

Zhuokai Zhao, Takumi Matsuzawa, William Irvine et al.

Proper evaluations are crucial for better understanding, troubleshooting, interpreting model behaviors and further improving model performance. While using scalar-based error metrics provides a fast way to overview model performance, they are often too abstract to display certain weak spots and lack information regarding important model properties, such as robustness. This not only hinders machine learning models from being more interpretable and gaining trust, but also can be misleading to both model developers and users. Additionally, conventional evaluation procedures often leave researchers unclear about where and how model fails, which complicates model comparisons and further developments. To address these issues, we propose a novel evaluation workflow, named Non-Equivariance Revealed on Orbits (NERO) Evaluation. The goal of NERO evaluation is to turn focus from traditional scalar-based metrics onto evaluating and visualizing models equivariance, closely capturing model robustness, as well as to allow researchers quickly investigating interesting or unexpected model behaviors. NERO evaluation is consist of a task-agnostic interactive interface and a set of visualizations, called NERO plots, which reveals the equivariance property of the model. Case studies on how NERO evaluation can be applied to multiple research areas, including 2D digit recognition, object detection, particle image velocimetry (PIV), and 3D point cloud classification, demonstrate that NERO evaluation can quickly illustrate different model equivariance, and effectively explain model behaviors through interactive visualizations of the model outputs. In addition, we propose consensus, an alternative to ground truths, to be used in NERO evaluation so that model equivariance can still be evaluated with new, unlabeled datasets.

Xin Yuan, Zhe Lin, Jason Kuen et al.

We develop an approach to learning visual representations that embraces multimodal data, driven by a combination of intra- and inter-modal similarity preservation objectives. Unlike existing visual pre-training methods, which solve a proxy prediction task in a single domain, our method exploits intrinsic data properties within each modality and semantic information from cross-modal correlation simultaneously, hence improving the quality of learned visual representations. By including multimodal training in a unified framework with different types of contrastive losses, our method can learn more powerful and generic visual features. We first train our model on COCO and evaluate the learned visual representations on various downstream tasks including image classification, object detection, and instance segmentation. For example, the visual representations pre-trained on COCO by our method achieve state-of-the-art top-1 validation accuracy of $55.3\%$ on ImageNet classification, under the common transfer protocol. We also evaluate our method on the large-scale Stock images dataset and show its effectiveness on multi-label image tagging, and cross-modal retrieval tasks.

Pedro Savarese, Sunnie S. Y. Kim, Michael Maire et al.

We study image segmentation from an information-theoretic perspective, proposing a novel adversarial method that performs unsupervised segmentation by partitioning images into maximally independent sets. More specifically, we group image pixels into foreground and background, with the goal of minimizing predictability of one set from the other. An easily computed loss drives a greedy search process to maximize inpainting error over these partitions. Our method does not involve training deep networks, is computationally cheap, class-agnostic, and even applicable in isolation to a single unlabeled image. Experiments demonstrate that it achieves a new state-of-the-art in unsupervised segmentation quality, while being substantially faster and more general than competing approaches.

Xiao Zhang, Michael Maire

We create a framework for bootstrapping visual representation learning from a primitive visual grouping capability. We operationalize grouping via a contour detector that partitions an image into regions, followed by merging of those regions into a tree hierarchy. A small supervised dataset suffices for training this grouping primitive. Across a large unlabeled dataset, we apply this learned primitive to automatically predict hierarchical region structure. These predictions serve as guidance for self-supervised contrastive feature learning: we task a deep network with producing per-pixel embeddings whose pairwise distances respect the region hierarchy. Experiments demonstrate that our approach can serve as state-of-the-art generic pre-training, benefiting downstream tasks. We additionally explore applications to semantic region search and video-based object instance tracking.

Chengcheng Wan, Henry Hoffmann, Shan Lu et al.

We propose a novel variant of SGD customized for training network architectures that support anytime behavior: such networks produce a series of increasingly accurate outputs over time. Efficient architectural designs for these networks focus on re-using internal state; subnetworks must produce representations relevant for both immediate prediction as well as refinement by subsequent network stages. We consider traditional branched networks as well as a new class of recursively nested networks. Our new optimizer, Orthogonalized SGD, dynamically re-balances task-specific gradients when training a multitask network. In the context of anytime architectures, this optimizer projects gradients from later outputs onto a parameter subspace that does not interfere with those from earlier outputs. Experiments demonstrate that training with Orthogonalized SGD significantly improves generalization accuracy of anytime networks.

Xin Yuan, Pedro Savarese, Michael Maire

We develop an approach to growing deep network architectures over the course of training, driven by a principled combination of accuracy and sparsity objectives. Unlike existing pruning or architecture search techniques that operate on full-sized models or supernet architectures, our method can start from a small, simple seed architecture and dynamically grow and prune both layers and filters. By combining a continuous relaxation of discrete network structure optimization with a scheme for sampling sparse subnetworks, we produce compact, pruned networks, while also drastically reducing the computational expense of training. For example, we achieve $49.7\%$ inference FLOPs and $47.4\%$ training FLOPs savings compared to a baseline ResNet-50 on ImageNet, while maintaining $75.2\%$ top-1 accuracy -- all without any dedicated fine-tuning stage. Experiments across CIFAR, ImageNet, PASCAL VOC, and Penn Treebank, with convolutional networks for image classification and semantic segmentation, and recurrent networks for language modeling, demonstrate that we both train faster and produce more efficient networks than competing architecture pruning or search methods.

Pedro Savarese, Hugo Silva, Michael Maire

The search for efficient, sparse deep neural network models is most prominently performed by pruning: training a dense, overparameterized network and removing parameters, usually via following a manually-crafted heuristic. Additionally, the recent Lottery Ticket Hypothesis conjectures that, for a typically-sized neural network, it is possible to find small sub-networks which, when trained from scratch on a comparable budget, match the performance of the original dense counterpart. We revisit fundamental aspects of pruning algorithms, pointing out missing ingredients in previous approaches, and develop a method, Continuous Sparsification, which searches for sparse networks based on a novel approximation of an intractable $\ell_0$ regularization. We compare against dominant heuristic-based methods on pruning as well as ticket search -- finding sparse subnetworks that can be successfully re-trained from an early iterate. Empirical results show that we surpass the state-of-the-art for both objectives, across models and datasets, including VGG trained on CIFAR-10 and ResNet-50 trained on ImageNet. In addition to setting a new standard for pruning, Continuous Sparsification also offers fast parallel ticket search, opening doors to new applications of the Lottery Ticket Hypothesis.

Pedro Savarese, David McAllester, Sudarshan Babu et al.

From a simplified analysis of adaptive methods, we derive AvaGrad, a new optimizer which outperforms SGD on vision tasks when its adaptability is properly tuned. We observe that the power of our method is partially explained by a decoupling of learning rate and adaptability, greatly simplifying hyperparameter search. In light of this observation, we demonstrate that, against conventional wisdom, Adam can also outperform SGD on vision tasks, as long as the coupling between its learning rate and adaptability is taken into account. In practice, AvaGrad matches the best results, as measured by generalization accuracy, delivered by any existing optimizer (SGD or adaptive) across image classification (CIFAR, ImageNet) and character-level language modelling (Penn Treebank) tasks.

Chengcheng Wan, Muhammad Santriaji, Eri Rogers et al.

An increasing number of software applications incorporate runtime Deep Neural Networks (DNNs) to process sensor data and return inference results to humans. Effective deployment of DNNs in these interactive scenarios requires meeting latency and accuracy constraints while minimizing energy, a problem exacerbated by common system dynamics. Prior approaches handle dynamics through either (1) system-oblivious DNN adaptation, which adjusts DNN latency/accuracy tradeoffs, or (2) application-oblivious system adaptation, which adjusts resources to change latency/energy tradeoffs. In contrast, this paper improves on the state-of-the-art by coordinating application- and system-level adaptation. ALERT, our runtime scheduler, uses a probabilistic model to detect environmental volatility and then simultaneously select both a DNN and a system resource configuration to meet latency, accuracy, and energy constraints. We evaluate ALERT on CPU and GPU platforms for image and speech tasks in dynamic environments. ALERT's holistic approach achieves more than 13% energy reduction, and 27% error reduction over prior approaches that adapt solely at the application or system level. Furthermore, ALERT incurs only 3% more energy consumption and 2% higher DNN-inference error than an oracle scheme with perfect application and system knowledge.

Tri Huynh, Michael Maire, Matthew R. Walter

We introduce a novel approach to endowing neural networks with emergent, long-term, large-scale memory. Distinct from strategies that connect neural networks to external memory banks via intricately crafted controllers and hand-designed attentional mechanisms, our memory is internal, distributed, co-located alongside computation, and implicitly addressed, while being drastically simpler than prior efforts. Architecting networks with multigrid structure and connectivity, while distributing memory cells alongside computation throughout this topology, we observe the emergence of coherent memory subsystems. Our hierarchical spatial organization, parameterized convolutionally, permits efficient instantiation of large-capacity memories, while multigrid topology provides short internal routing pathways, allowing convolutional networks to efficiently approximate the behavior of fully connected networks. Such networks have an implicit capacity for internal attention; augmented with memory, they learn to read and write specific memory locations in a dynamic data-dependent manner. We demonstrate these capabilities on exploration and mapping tasks, where our network is able to self-organize and retain long-term memory for trajectories of thousands of time steps. On tasks decoupled from any notion of spatial geometry: sorting, associative recall, and question answering, our design functions as a truly generic memory and yields excellent results.

Pedro Savarese, Michael Maire

We introduce a parameter sharing scheme, in which different layers of a convolutional neural network (CNN) are defined by a learned linear combination of parameter tensors from a global bank of templates. Restricting the number of templates yields a flexible hybridization of traditional CNNs and recurrent networks. Compared to traditional CNNs, we demonstrate substantial parameter savings on standard image classification tasks, while maintaining accuracy. Our simple parameter sharing scheme, though defined via soft weights, in practice often yields trained networks with near strict recurrent structure; with negligible side effects, they convert into networks with actual loops. Training these networks thus implicitly involves discovery of suitable recurrent architectures. Though considering only the design aspect of recurrent links, our trained networks achieve accuracy competitive with those built using state-of-the-art neural architecture search (NAS) procedures. Our hybridization of recurrent and convolutional networks may also represent a beneficial architectural bias. Specifically, on synthetic tasks which are algorithmic in nature, our hybrid networks both train faster and extrapolate better to test examples outside the span of the training set.

Mohammadreza Mostajabi, Michael Maire, Gregory Shakhnarovich

We construct custom regularization functions for use in supervised training of deep neural networks. Our technique is applicable when the ground-truth labels themselves exhibit internal structure; we derive a regularizer by learning an autoencoder over the set of annotations. Training thereby becomes a two-phase procedure. The first phase models labels with an autoencoder. The second phase trains the actual network of interest by attaching an auxiliary branch that must predict output via a hidden layer of the autoencoder. After training, we discard this auxiliary branch. We experiment in the context of semantic segmentation, demonstrating this regularization strategy leads to consistent accuracy boosts over baselines, both when training from scratch, or in combination with ImageNet pretraining. Gains are also consistent over different choices of convolutional network architecture. As our regularizer is discarded after training, our method has zero cost at test time; the performance improvements are essentially free. We are simply able to learn better network weights by building an abstract model of the label space, and then training the network to understand this abstraction alongside the original task.

Ligeng Zhu, Ruizhi Deng, Michael Maire et al.

We explore a key architectural aspect of deep convolutional neural networks: the pattern of internal skip connections used to aggregate outputs of earlier layers for consumption by deeper layers. Such aggregation is critical to facilitate training of very deep networks in an end-to-end manner. This is a primary reason for the widespread adoption of residual networks, which aggregate outputs via cumulative summation. While subsequent works investigate alternative aggregation operations (e.g. concatenation), we focus on an orthogonal question: which outputs to aggregate at a particular point in the network. We propose a new internal connection structure which aggregates only a sparse set of previous outputs at any given depth. Our experiments demonstrate this simple design change offers superior performance with fewer parameters and lower computational requirements. Moreover, we show that sparse aggregation allows networks to scale more robustly to 1000+ layers, thereby opening future avenues for training long-running visual processes.

Huaizu Jiang, Erik Learned-Miller, Gustav Larsson et al.

As an agent moves through the world, the apparent motion of scene elements is (usually) inversely proportional to their depth. It is natural for a learning agent to associate image patterns with the magnitude of their displacement over time: as the agent moves, faraway mountains don't move much; nearby trees move a lot. This natural relationship between the appearance of objects and their motion is a rich source of information about the world. In this work, we start by training a deep network, using fully automatic supervision, to predict relative scene depth from single images. The relative depth training images are automatically derived from simple videos of cars moving through a scene, using recent motion segmentation techniques, and no human-provided labels. This proxy task of predicting relative depth from a single image induces features in the network that result in large improvements in a set of downstream tasks including semantic segmentation, joint road segmentation and car detection, and monocular (absolute) depth estimation, over a network trained from scratch. The improvement on the semantic segmentation task is greater than those produced by any other automatically supervised methods. Moreover, for monocular depth estimation, our unsupervised pre-training method even outperforms supervised pre-training with ImageNet. In addition, we demonstrate benefits from learning to predict (unsupervised) relative depth in the specific videos associated with various downstream tasks. We adapt to the specific scenes in those tasks in an unsupervised manner to improve performance. In summary, for semantic segmentation, we present state-of-the-art results among methods that do not use supervised pre-training, and we even exceed the performance of supervised ImageNet pre-trained models for monocular depth estimation, achieving results that are comparable with state-of-the-art methods.

Gustav Larsson, Michael Maire, Gregory Shakhnarovich

We investigate and improve self-supervision as a drop-in replacement for ImageNet pretraining, focusing on automatic colorization as the proxy task. Self-supervised training has been shown to be more promising for utilizing unlabeled data than other, traditional unsupervised learning methods. We build on this success and evaluate the ability of our self-supervised network in several contexts. On VOC segmentation and classification tasks, we present results that are state-of-the-art among methods not using ImageNet labels for pretraining representations. Moreover, we present the first in-depth analysis of self-supervision via colorization, concluding that formulation of the loss, training details and network architecture play important roles in its effectiveness. This investigation is further expanded by revisiting the ImageNet pretraining paradigm, asking questions such as: How much training data is needed? How many labels are needed? How much do features change when fine-tuned? We relate these questions back to self-supervision by showing that colorization provides a similarly powerful supervisory signal as various flavors of ImageNet pretraining.

Tsung-Wei Ke, Michael Maire, Stella X. Yu

We propose a multigrid extension of convolutional neural networks (CNNs). Rather than manipulating representations living on a single spatial grid, our network layers operate across scale space, on a pyramid of grids. They consume multigrid inputs and produce multigrid outputs; convolutional filters themselves have both within-scale and cross-scale extent. This aspect is distinct from simple multiscale designs, which only process the input at different scales. Viewed in terms of information flow, a multigrid network passes messages across a spatial pyramid. As a consequence, receptive field size grows exponentially with depth, facilitating rapid integration of context. Most critically, multigrid structure enables networks to learn internal attention and dynamic routing mechanisms, and use them to accomplish tasks on which modern CNNs fail. Experiments demonstrate wide-ranging performance advantages of multigrid. On CIFAR and ImageNet classification tasks, flipping from a single grid to multigrid within the standard CNN paradigm improves accuracy, while being compute and parameter efficient. Multigrid is independent of other architectural choices; we show synergy in combination with residual connections. Multigrid yields dramatic improvement on a synthetic semantic segmentation dataset. Most strikingly, relatively shallow multigrid networks can learn to directly perform spatial transformation tasks, where, in contrast, current CNNs fail. Together, our results suggest that continuous evolution of features on a multigrid pyramid is a more powerful alternative to existing CNN designs on a flat grid.

Gustav Larsson, Michael Maire, Gregory Shakhnarovich

We introduce a design strategy for neural network macro-architecture based on self-similarity. Repeated application of a simple expansion rule generates deep networks whose structural layouts are precisely truncated fractals. These networks contain interacting subpaths of different lengths, but do not include any pass-through or residual connections; every internal signal is transformed by a filter and nonlinearity before being seen by subsequent layers. In experiments, fractal networks match the excellent performance of standard residual networks on both CIFAR and ImageNet classification tasks, thereby demonstrating that residual representations may not be fundamental to the success of extremely deep convolutional neural networks. Rather, the key may be the ability to transition, during training, from effectively shallow to deep. We note similarities with student-teacher behavior and develop drop-path, a natural extension of dropout, to regularize co-adaptation of subpaths in fractal architectures. Such regularization allows extraction of high-performance fixed-depth subnetworks. Additionally, fractal networks exhibit an anytime property: shallow subnetworks provide a quick answer, while deeper subnetworks, with higher latency, provide a more accurate answer.

Gustav Larsson, Michael Maire, Gregory Shakhnarovich

We develop a fully automatic image colorization system. Our approach leverages recent advances in deep networks, exploiting both low-level and semantic representations. As many scene elements naturally appear according to multimodal color distributions, we train our model to predict per-pixel color histograms. This intermediate output can be used to automatically generate a color image, or further manipulated prior to image formation. On both fully and partially automatic colorization tasks, we outperform existing methods. We also explore colorization as a vehicle for self-supervised visual representation learning.

Michael Maire, Takuya Narihira, Stella X. Yu

Spectral embedding provides a framework for solving perceptual organization problems, including image segmentation and figure/ground organization. From an affinity matrix describing pairwise relationships between pixels, it clusters pixels into regions, and, using a complex-valued extension, orders pixels according to layer. We train a convolutional neural network (CNN) to directly predict the pairwise relationships that define this affinity matrix. Spectral embedding then resolves these predictions into a globally-consistent segmentation and figure/ground organization of the scene. Experiments demonstrate significant benefit to this direct coupling compared to prior works which use explicit intermediate stages, such as edge detection, on the pathway from image to affinities. Our results suggest spectral embedding as a powerful alternative to the conditional random field (CRF)-based globalization schemes typically coupled to deep neural networks.

Takuya Narihira, Michael Maire, Stella X. Yu

We introduce a new approach to intrinsic image decomposition, the task of decomposing a single image into albedo and shading components. Our strategy, which we term direct intrinsics, is to learn a convolutional neural network (CNN) that directly predicts output albedo and shading channels from an input RGB image patch. Direct intrinsics is a departure from classical techniques for intrinsic image decomposition, which typically rely on physically-motivated priors and graph-based inference algorithms. The large-scale synthetic ground-truth of the MPI Sintel dataset plays a key role in training direct intrinsics. We demonstrate results on both the synthetic images of Sintel and the real images of the classic MIT intrinsic image dataset. On Sintel, direct intrinsics, using only RGB input, outperforms all prior work, including methods that rely on RGB+Depth input. Direct intrinsics also generalizes across modalities; it produces quite reasonable decompositions on the real images of the MIT dataset. Our results indicate that the marriage of CNNs with synthetic training data may be a powerful new technique for tackling classic problems in computer vision.

Michael Maire, Stella X. Yu, Pietro Perona

We frame the task of predicting a semantic labeling as a sparse reconstruction procedure that applies a target-specific learned transfer function to a generic deep sparse code representation of an image. This strategy partitions training into two distinct stages. First, in an unsupervised manner, we learn a set of generic dictionaries optimized for sparse coding of image patches. We train a multilayer representation via recursive sparse dictionary learning on pooled codes output by earlier layers. Second, we encode all training images with the generic dictionaries and learn a transfer function that optimizes reconstruction of patches extracted from annotated ground-truth given the sparse codes of their corresponding image patches. At test time, we encode a novel image using the generic dictionaries and then reconstruct using the transfer function. The output reconstruction is a semantic labeling of the test image. Applying this strategy to the task of contour detection, we demonstrate performance competitive with state-of-the-art systems. Unlike almost all prior work, our approach obviates the need for any form of hand-designed features or filters. To illustrate general applicability, we also show initial results on semantic part labeling of human faces. The effectiveness of our approach opens new avenues for research on deep sparse representations. Our classifiers utilize this representation in a novel manner. Rather than acting on nodes in the deepest layer, they attach to nodes along a slice through multiple layers of the network in order to make predictions about local patches. Our flexible combination of a generatively learned sparse representation with discriminatively trained transfer classifiers extends the notion of sparse reconstruction to encompass arbitrary semantic labeling tasks.

Tsung-Yi Lin, Michael Maire, Serge Belongie et al.

We present a new dataset with the goal of advancing the state-of-the-art in object recognition by placing the question of object recognition in the context of the broader question of scene understanding. This is achieved by gathering images of complex everyday scenes containing common objects in their natural context. Objects are labeled using per-instance segmentations to aid in precise object localization. Our dataset contains photos of 91 objects types that would be easily recognizable by a 4 year old. With a total of 2.5 million labeled instances in 328k images, the creation of our dataset drew upon extensive crowd worker involvement via novel user interfaces for category detection, instance spotting and instance segmentation. We present a detailed statistical analysis of the dataset in comparison to PASCAL, ImageNet, and SUN. Finally, we provide baseline performance analysis for bounding box and segmentation detection results using a Deformable Parts Model.