Search

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

The inference of transformer-based large language models consists of two sequential stages: 1) a prefilling stage to compute the KV cache of prompts and generate the first token, and 2) a decoding stage to generate subsequent tokens. For long prompts, the KV cache must be computed for all tokens during the prefilling stage, which can significantly increase the time needed to generate the first token. Consequently, the prefilling stage may become a bottleneck in the generation process. An open question remains whether all prompt tokens are essential for generating the first token. To answer this, we introduce a novel method, LazyLLM, that selectively computes the KV for tokens important for the next token prediction in both the prefilling and decoding stages. Contrary to static pruning approaches that prune the prompt at once, LazyLLM allows language models to dynamically select different subsets of tokens from the context in different generation steps, even though they might be pruned in previous steps. Extensive experiments on standard datasets across various tasks demonstrate that LazyLLM is a generic method that can be seamlessly integrated with existing language models to significantly accelerate the generation without fine-tuning. For instance, in the multi-document question-answering task, LazyLLM accelerates the prefilling stage of the LLama 2 7B model by 2.34x while maintaining accuracy.

Large Language Model or LLM inference has two phases, the prompt (or prefill) phase to output the first token and the extension (or decoding) phase to the generate subsequent tokens. In this work, we propose an efficient parallelization scheme, KV-Runahead to accelerate the prompt phase. The key observation is that the extension phase generates tokens faster than the prompt phase because of key-value cache (KV-cache). Hence, KV-Runahead parallelizes the prompt phase by orchestrating multiple processes to populate the KV-cache and minimizes the time-to-first-token (TTFT). Dual-purposing the KV-cache scheme has two main benefits. First, since KV-cache is designed to leverage the causal attention map, we minimize computation and computation automatically. Second, since it already exists for the extension phase, KV-Runahead is easy to implement. We further propose context-level load-balancing to handle uneven KV-cache generation (due to the causal attention) and to optimize TTFT. Compared with an existing parallelization scheme such as tensor or sequential parallelization where keys and values are locally generated and exchanged via all-gather collectives, our experimental results demonstrate that KV-Runahead can offer over 1.4x and 1.6x speedups for Llama 7B and Falcon 7B respectively., Comment: preprint for ICML 2024

Contrastive learning has emerged as a transformative method for learning effective visual representations through the alignment of image and text embeddings. However, pairwise similarity computation in contrastive loss between image and text pairs poses computational challenges. This paper presents a novel weakly supervised pre-training of vision models on web-scale image-text data. The proposed method reframes pre-training on image-text data as a classification task. Consequently, it eliminates the need for pairwise similarity computations in contrastive loss, achieving a remarkable $2.7\times$ acceleration in training speed compared to contrastive learning on web-scale data. Through extensive experiments spanning diverse vision tasks, including detection and segmentation, we demonstrate that the proposed method maintains high representation quality. Our source code along with pre-trained model weights and training recipes is available at \url{https://github.com/apple/corenet}.

The reproducibility and transparency of large language models are crucial for advancing open research, ensuring the trustworthiness of results, and enabling investigations into data and model biases, as well as potential risks. To this end, we release OpenELM, a state-of-the-art open language model. OpenELM uses a layer-wise scaling strategy to efficiently allocate parameters within each layer of the transformer model, leading to enhanced accuracy. For example, with a parameter budget of approximately one billion parameters, OpenELM exhibits a 2.36% improvement in accuracy compared to OLMo while requiring $2\times$ fewer pre-training tokens. Diverging from prior practices that only provide model weights and inference code, and pre-train on private datasets, our release includes the complete framework for training and evaluation of the language model on publicly available datasets, including training logs, multiple checkpoints, and pre-training configurations. We also release code to convert models to MLX library for inference and fine-tuning on Apple devices. This comprehensive release aims to empower and strengthen the open research community, paving the way for future open research endeavors. Our source code along with pre-trained model weights and training recipes is available at \url{https://github.com/apple/corenet}. Additionally, \model models can be found on HuggingFace at: \url{https://huggingface.co/apple/OpenELM}., Comment: Minor corrections

Despite the successes of large language models (LLMs), they exhibit significant drawbacks, particularly when processing long contexts. Their inference cost scales quadratically with respect to sequence length, making it expensive for deployment in some real-world text processing applications, such as retrieval-augmented generation (RAG). Additionally, LLMs also exhibit the "distraction phenomenon", where irrelevant context in the prompt degrades output quality. To address these drawbacks, we propose a novel RAG prompting methodology, *superposition prompting*, which can be directly applied to pre-trained transformer-based LLMs *without the need for fine-tuning*. At a high level, superposition prompting allows the LLM to process input documents in parallel *prompt paths*, discarding paths once they are deemed irrelevant. We demonstrate the capability of our method to simultaneously enhance time efficiency across a variety of question-answering benchmarks using multiple pre-trained LLMs. Furthermore, our technique significantly improves accuracy when the retrieved context is large relative the context the model was trained on. For example, our approach facilitates a 93x reduction in compute time while *improving* accuracy by 43% on the NaturalQuestions-Open dataset with the MPT-7B instruction-tuned model over naive RAG.

Speculative decoding is a prominent technique to speed up the inference of a large target language model based on predictions of an auxiliary draft model. While effective, in application-specific settings, it often involves fine-tuning both draft and target models to achieve high acceptance rates. As the number of downstream tasks grows, these draft models add significant complexity to inference systems. We propose Speculative Streaming, a single-model speculative decoding method that fuses drafting into the target model by changing the fine-tuning objective from next token prediction to future n-gram prediction. Speculative Streaming speeds up decoding by 1.8 - 3.1X in a diverse set of tasks, such as Summarization, Structured Queries, and Meaning Representation, without sacrificing generation quality. Additionally, Speculative Streaming is parameter-efficient. It achieves on-par/higher speed-ups than Medusa-style architectures while using ~10000X fewer extra parameters, making it well-suited for resource-constrained devices.

Training large transformer models from scratch for a target task requires lots of data and is computationally demanding. The usual practice of transfer learning overcomes this challenge by initializing the model with weights of a pretrained model of the same size and specification to increase the convergence and training speed. However, what if no pretrained model of the required size is available? In this paper, we introduce a simple yet effective technique to transfer the knowledge of a pretrained model to smaller variants. Our approach called weight subcloning expedites the training of scaled-down transformers by initializing their weights from larger pretrained models. Weight subcloning involves an operation on the pretrained model to obtain the equivalent initialized scaled-down model. It consists of two key steps: first, we introduce neuron importance ranking to decrease the embedding dimension per layer in the pretrained model. Then, we remove blocks from the transformer model to match the number of layers in the scaled-down network. The result is a network ready to undergo training, which gains significant improvements in training speed compared to random initialization. For instance, we achieve 4x faster training for vision transformers in image classification and language models designed for next token prediction.

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their substantial computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters in flash memory, but bringing them on demand to DRAM. Our method involves constructing an inference cost model that takes into account the characteristics of flash memory, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this hardware-informed framework, we introduce two principal techniques. First, "windowing" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory., Comment: ACL 2024

Vision Foundation Models (VFMs) pretrained on massive datasets exhibit impressive performance on various downstream tasks, especially with limited labeled target data. However, due to their high inference compute cost, these models cannot be deployed for many real-world applications. Motivated by this, we ask the following important question, "How can we leverage the knowledge from a large VFM to train a small task-specific model for a new target task with limited labeled training data?", and propose a simple task-oriented knowledge transfer approach as a highly effective solution to this problem. Our experimental results on five target tasks show that the proposed approach outperforms task-agnostic VFM distillation, web-scale CLIP pretraining, supervised ImageNet pretraining, and self-supervised DINO pretraining by up to 11.6%, 22.1%, 13.7%, and 29.8%, respectively. Furthermore, the proposed approach also demonstrates up to 9x, 4x and 15x reduction in pretraining compute cost when compared to task-agnostic VFM distillation, ImageNet pretraining and DINO pretraining, respectively, while outperforming them. We also show that the dataset used for transferring knowledge has a significant effect on the final target task performance, and introduce a retrieval-augmented knowledge transfer strategy that uses web-scale image retrieval to curate effective transfer sets., Comment: International Conference on Machine Learning, 2024

The landscape of publicly available vision foundation models (VFMs), such as CLIP and Segment Anything Model (SAM), is expanding rapidly. VFMs are endowed with distinct capabilities stemming from their pre-training objectives. For instance, CLIP excels in semantic understanding, while SAM specializes in spatial understanding for segmentation. In this work, we introduce a simple recipe to efficiently merge VFMs into a unified model that absorbs their expertise. Our method integrates techniques of multi-task learning, continual learning, and distillation. Further, it demands significantly less computational cost compared to traditional multi-task training from scratch, and it only needs a small fraction of the pre-training datasets that were initially used to train individual models. By applying our method to SAM and CLIP, we obtain SAM-CLIP: a unified model that combines the capabilities of SAM and CLIP into a single vision transformer. Compared with deploying SAM and CLIP independently, our merged model, SAM-CLIP, reduces storage and compute costs for inference, making it well-suited for edge device applications. We show that SAM-CLIP not only retains the foundational strengths of SAM and CLIP, but also introduces synergistic functionalities, notably in zero-shot semantic segmentation, where SAM-CLIP establishes new state-of-the-art results on 5 benchmarks. It outperforms previous models that are specifically designed for this task by a large margin, including +6.8% and +5.9% mean IoU improvement on Pascal-VOC and COCO-Stuff datasets, respectively.

Contrastive language image pretraining (CLIP) is a standard method for training vision-language models. While CLIP is scalable, promptable, and robust to distribution shifts on image classification tasks, it lacks object localization capabilities. This paper studies the following question: Can we augment CLIP training with task-specific vision models from model zoos to improve its visual representations? Towards this end, we leverage open-source task-specific vision models to generate pseudo-labels for an uncurated and noisy image-text dataset. Subsequently, we train CLIP models on these pseudo-labels in addition to the contrastive training on image and text pairs. This simple setup shows substantial improvements of up to 16.3% across different vision tasks, including segmentation, detection, depth estimation, and surface normal estimation. Importantly, these enhancements are achieved without compromising CLIP's existing capabilities, including its proficiency in promptable zero-shot classification.

Large Language Models (LLMs) with billions of parameters have drastically transformed AI applications. However, their demanding computation during inference has raised significant challenges for deployment on resource-constrained devices. Despite recent trends favoring alternative activation functions such as GELU or SiLU, known for increased computation, this study strongly advocates for reinstating ReLU activation in LLMs. We demonstrate that using the ReLU activation function has a negligible impact on convergence and performance while significantly reducing computation and weight transfer. This reduction is particularly valuable during the memory-bound inference step, where efficiency is paramount. Exploring sparsity patterns in ReLU-based LLMs, we unveil the reutilization of activated neurons for generating new tokens and leveraging these insights, we propose practical strategies to substantially reduce LLM inference computation up to three times, using ReLU activations with minimal performance trade-offs., Comment: preprint

Over the past several years, the synchronization between audio and visual signals has been leveraged to learn richer audio-visual representations. Aided by the large availability of unlabeled videos, many unsupervised training frameworks have demonstrated impressive results in various downstream audio and video tasks. Recently, Masked Audio-Video Learners (MAViL) has emerged as a state-of-the-art audio-video pre-training framework. MAViL couples contrastive learning with masked autoencoding to jointly reconstruct audio spectrograms and video frames by fusing information from both modalities. In this paper, we study the potential synergy between diffusion models and MAViL, seeking to derive mutual benefits from these two frameworks. The incorporation of diffusion into MAViL, combined with various training efficiency methodologies that include the utilization of a masking ratio curriculum and adaptive batch sizing, results in a notable 32% reduction in pre-training Floating-Point Operations (FLOPS) and an 18% decrease in pre-training wall clock time. Crucially, this enhanced efficiency does not compromise the model's performance in downstream audio-classification tasks when compared to MAViL's performance., Comment: Camera-ready version for the Machine Learning for Audio Workshop at NeurIPS 2023

Compressing Large Language Models (LLMs) often leads to reduced performance, especially for knowledge-intensive tasks. In this work, we dive into how compression damages LLMs' inherent knowledge and the possible remedies. We start by proposing two conjectures on the nature of the damage: one is certain knowledge being forgotten (or erased) after LLM compression, hence necessitating the compressed model to (re)learn from data with additional parameters; the other presumes that knowledge is internally displaced and hence one requires merely "inference re-direction" with input-side augmentation such as prompting, to recover the knowledge-related performance. Extensive experiments are then designed to (in)validate the two conjectures. We observe the promise of prompting in comparison to model tuning; we further unlock prompting's potential by introducing a variant called Inference-time Dynamic Prompting (IDP), that can effectively increase prompt diversity without incurring any inference overhead. Our experiments consistently suggest that compared to the classical re-training alternatives such as LoRA, prompting with IDP leads to better or comparable post-compression performance recovery, while saving the extra parameter size by 21x and reducing inference latency by 60%. Our experiments hence strongly endorse the conjecture of "knowledge displaced" over "knowledge forgotten", and shed light on a new efficient mechanism to restore compressed LLM performance. We additionally visualize and analyze the different attention and activation patterns between prompted and re-trained models, demonstrating they achieve performance recovery in two different regimes.

Multi-scale resolution training has seen an increased adoption across multiple vision tasks, including classification and detection. Training with smaller resolutions enables faster training at the expense of a drop in accuracy. Conversely, training with larger resolutions has been shown to improve performance, but memory constraints often make this infeasible. In this paper, we empirically study the properties of multi-scale training procedures. We focus on variable batch size multi-scale data samplers that randomly sample an input resolution at each training iteration and dynamically adjust their batch size according to the resolution. Such samplers have been shown to improve model accuracy beyond standard training with a fixed batch size and resolution, though it is not clear why this is the case. We explore the properties of these data samplers by performing extensive experiments on ResNet-101 and validate our conclusions across multiple architectures, tasks, and datasets. We show that multi-scale samplers behave as implicit data regularizers and accelerate training speed. Compared to models trained with single-scale samplers, we show that models trained with multi-scale samplers retain or improve accuracy, while being better-calibrated and more robust to scaling and data distribution shifts. We additionally extend a multi-scale variable batch sampler with a simple curriculum that progressively grows resolutions throughout training, allowing for a compute reduction of more than 30%. We show that the benefits of multi-scale training extend to detection and instance segmentation tasks, where we observe a 37% reduction in training FLOPs along with a 3-4% mAP increase on MS-COCO using a Mask R-CNN model.

Since Large Language Models or LLMs have demonstrated high-quality performance on many complex language tasks, there is a great interest in bringing these LLMs to mobile devices for faster responses and better privacy protection. However, the size of LLMs (i.e., billions of parameters) requires highly effective compression to fit into storage-limited devices. Among many compression techniques, weight-clustering, a form of non-linear quantization, is one of the leading candidates for LLM compression, and supported by modern smartphones. Yet, its training overhead is prohibitively significant for LLM fine-tuning. Especially, Differentiable KMeans Clustering, or DKM, has shown the state-of-the-art trade-off between compression ratio and accuracy regression, but its large memory complexity makes it nearly impossible to apply to train-time LLM compression. In this paper, we propose a memory-efficient DKM implementation, eDKM powered by novel techniques to reduce the memory footprint of DKM by orders of magnitudes. For a given tensor to be saved on CPU for the backward pass of DKM, we compressed the tensor by applying uniquification and sharding after checking if there is no duplicated tensor previously copied to CPU. Our experimental results demonstrate that \prjname can fine-tune and compress a pretrained LLaMA 7B model from 12.6 GB to 2.5 GB (3bit/weight) with the Alpaca dataset by reducing the train-time memory footprint of a decoder layer by 130$\times$, while delivering good accuracy on broader LLM benchmarks (i.e., 77.7% for PIQA, 66.1% for Winograde, and so on)., Comment: preprint

Modern deep learning approaches usually utilize modality-specific processing. For example, the most common deep learning approach to image classification involves decoding image file bytes into an RGB tensor which is passed into a neural network. Instead, we investigate modality-independent representation learning by performing classification directly on file bytes, without the need for decoding files at inference time. This enables models to operate on various modalities without any hand-designed, modality-specific processing. Our model, ByteFormer, improves ImageNet Top-1 classification accuracy by $5\%$ (from $72.2\%$ to $77.33\%$) relative to DeIT models of similar size. Compared to Perceiver IO, our model requires absolutely no modality-specific processing at inference time, and uses an order of magnitude fewer parameters at equivalent accuracy on ImageNet. We demonstrate that the same ByteFormer architecture can perform audio classification without modifications or modality-specific preprocessing. We achieve $95.42\%$ classification accuracy on the Speech Commands V2 dataset (comparable to the state-of-the-art accuracy of $98.7\%$). Additionally, we demonstrate that ByteFormer can operate jointly on images and audio, handling joint classification without explicit knowledge of the input modality. We release our code at https://github.com/apple/corenet/tree/main/projects/byteformer.

We propose Dataset Reinforcement, a strategy to improve a dataset once such that the accuracy of any model architecture trained on the reinforced dataset is improved at no additional training cost for users. We propose a Dataset Reinforcement strategy based on data augmentation and knowledge distillation. Our generic strategy is designed based on extensive analysis across CNN- and transformer-based models and performing large-scale study of distillation with state-of-the-art models with various data augmentations. We create a reinforced version of the ImageNet training dataset, called ImageNet+, as well as reinforced datasets CIFAR-100+, Flowers-102+, and Food-101+. Models trained with ImageNet+ are more accurate, robust, and calibrated, and transfer well to downstream tasks (e.g., segmentation and detection). As an example, the accuracy of ResNet-50 improves by 1.7% on the ImageNet validation set, 3.5% on ImageNetV2, and 10.0% on ImageNet-R. Expected Calibration Error (ECE) on the ImageNet validation set is also reduced by 9.9%. Using this backbone with Mask-RCNN for object detection on MS-COCO, the mean average precision improves by 0.8%. We reach similar gains for MobileNets, ViTs, and Swin-Transformers. For MobileNetV3 and Swin-Tiny, we observe significant improvements on ImageNet-R/A/C of up to 20% improved robustness. Models pretrained on ImageNet+ and fine-tuned on CIFAR-100+, Flowers-102+, and Food-101+, reach up to 3.4% improved accuracy. The code, datasets, and pretrained models are available at https://github.com/apple/ml-dr., Comment: Accepted at International Conference on Computer Vision (ICCV) 2023. v2: Camera-ready version with new Tables 9 and 10. v3: Correction to Table 7-Avg. column

State-of-the-art automatic augmentation methods (e.g., AutoAugment and RandAugment) for visual recognition tasks diversify training data using a large set of augmentation operations. The range of magnitudes of many augmentation operations (e.g., brightness and contrast) is continuous. Therefore, to make search computationally tractable, these methods use fixed and manually-defined magnitude ranges for each operation, which may lead to sub-optimal policies. To answer the open question on the importance of magnitude ranges for each augmentation operation, we introduce RangeAugment that allows us to efficiently learn the range of magnitudes for individual as well as composite augmentation operations. RangeAugment uses an auxiliary loss based on image similarity as a measure to control the range of magnitudes of augmentation operations. As a result, RangeAugment has a single scalar parameter for search, image similarity, which we simply optimize via linear search. RangeAugment integrates seamlessly with any model and learns model- and task-specific augmentation policies. With extensive experiments on the ImageNet dataset across different networks, we show that RangeAugment achieves competitive performance to state-of-the-art automatic augmentation methods with 4-5 times fewer augmentation operations. Experimental results on semantic segmentation, object detection, foundation models, and knowledge distillation further shows RangeAugment's effectiveness., Comment: Technical report (22 pages including references and appendix)

Recent isotropic networks, such as ConvMixer and vision transformers, have found significant success across visual recognition tasks, matching or outperforming non-isotropic convolutional neural networks (CNNs). Isotropic architectures are particularly well-suited to cross-layer weight sharing, an effective neural network compression technique. In this paper, we perform an empirical evaluation on methods for sharing parameters in isotropic networks (SPIN). We present a framework to formalize major weight sharing design decisions and perform a comprehensive empirical evaluation of this design space. Guided by our experimental results, we propose a weight sharing strategy to generate a family of models with better overall efficiency, in terms of FLOPs and parameters versus accuracy, compared to traditional scaling methods alone, for example compressing ConvMixer by 1.9x while improving accuracy on ImageNet. Finally, we perform a qualitative study to further understand the behavior of weight sharing in isotropic architectures. The code is available at https://github.com/apple/ml-spin., Comment: Accepted at ECCV 2022

Mobile vision transformers (MobileViT) can achieve state-of-the-art performance across several mobile vision tasks, including classification and detection. Though these models have fewer parameters, they have high latency as compared to convolutional neural network-based models. The main efficiency bottleneck in MobileViT is the multi-headed self-attention (MHA) in transformers, which requires $O(k^2)$ time complexity with respect to the number of tokens (or patches) $k$. Moreover, MHA requires costly operations (e.g., batch-wise matrix multiplication) for computing self-attention, impacting latency on resource-constrained devices. This paper introduces a separable self-attention method with linear complexity, i.e. $O(k)$. A simple yet effective characteristic of the proposed method is that it uses element-wise operations for computing self-attention, making it a good choice for resource-constrained devices. The improved model, MobileViTv2, is state-of-the-art on several mobile vision tasks, including ImageNet object classification and MS-COCO object detection. With about three million parameters, MobileViTv2 achieves a top-1 accuracy of 75.6% on the ImageNet dataset, outperforming MobileViT by about 1% while running $3.2\times$ faster on a mobile device. Our source code is available at: \url{https://github.com/apple/ml-cvnets}, Comment: Technical report

We introduce CVNets, a high-performance open-source library for training deep neural networks for visual recognition tasks, including classification, detection, and segmentation. CVNets supports image and video understanding tools, including data loading, data transformations, novel data sampling methods, and implementations of several standard networks with similar or better performance than previous studies. Our source code is available at: \url{https://github.com/apple/ml-cvnets}., Comment: Technical report

When deploying deep learning models to a device, it is traditionally assumed that available computational resources (compute, memory, and power) remain static. However, real-world computing systems do not always provide stable resource guarantees. Computational resources need to be conserved when load from other processes is high or battery power is low. Inspired by recent works on neural network subspaces, we propose a method for training a "compressible subspace" of neural networks that contains a fine-grained spectrum of models that range from highly efficient to highly accurate. Our models require no retraining, thus our subspace of models can be deployed entirely on-device to allow adaptive network compression at inference time. We present results for achieving arbitrarily fine-grained accuracy-efficiency trade-offs at inference time for structured and unstructured sparsity. We achieve accuracies on-par with standard models when testing our uncompressed models, and maintain high accuracy for sparsity rates above 90% when testing our compressed models. We also demonstrate that our algorithm extends to quantization at variable bit widths, achieving accuracy on par with individually trained networks.

Despite the recent success in many applications, the high computational requirements of vision transformers limit their use in resource-constrained settings. While many existing methods improve the quadratic complexity of attention, in most vision transformers, self-attention is not the major computation bottleneck, e.g., more than 80% of the computation is spent on fully-connected layers. To improve the computational complexity of all layers, we propose a novel token downsampling method, called Token Pooling, efficiently exploiting redundancies in the images and intermediate token representations. We show that, under mild assumptions, softmax-attention acts as a high-dimensional low-pass (smoothing) filter. Thus, its output contains redundancy that can be pruned to achieve a better trade-off between the computational cost and accuracy. Our new technique accurately approximates a set of tokens by minimizing the reconstruction error caused by downsampling. We solve this optimization problem via cost-efficient clustering. We rigorously analyze and compare to prior downsampling methods. Our experiments show that Token Pooling significantly improves the cost-accuracy trade-off over the state-of-the-art downsampling. Token Pooling is a simple and effective operator that can benefit many architectures. Applied to DeiT, it achieves the same ImageNet top-1 accuracy using 42% fewer computations.

Light-weight convolutional neural networks (CNNs) are the de-facto for mobile vision tasks. Their spatial inductive biases allow them to learn representations with fewer parameters across different vision tasks. However, these networks are spatially local. To learn global representations, self-attention-based vision trans-formers (ViTs) have been adopted. Unlike CNNs, ViTs are heavy-weight. In this paper, we ask the following question: is it possible to combine the strengths of CNNs and ViTs to build a light-weight and low latency network for mobile vision tasks? Towards this end, we introduce MobileViT, a light-weight and general-purpose vision transformer for mobile devices. MobileViT presents a different perspective for the global processing of information with transformers, i.e., transformers as convolutions. Our results show that MobileViT significantly outperforms CNN- and ViT-based networks across different tasks and datasets. On the ImageNet-1k dataset, MobileViT achieves top-1 accuracy of 78.4% with about 6 million parameters, which is 3.2% and 6.2% more accurate than MobileNetv3 (CNN-based) and DeIT (ViT-based) for a similar number of parameters. On the MS-COCO object detection task, MobileViT is 5.7% more accurate than MobileNetv3 for a similar number of parameters. Our source code is open-source and available at: https://github.com/apple/ml-cvnets, Comment: Accepted at ICLR'22

Deep neural network (DNN) model compression for efficient on-device inference is becoming increasingly important to reduce memory requirements and keep user data on-device. To this end, we propose a novel differentiable k-means clustering layer (DKM) and its application to train-time weight clustering-based DNN model compression. DKM casts k-means clustering as an attention problem and enables joint optimization of the DNN parameters and clustering centroids. Unlike prior works that rely on additional regularizers and parameters, DKM-based compression keeps the original loss function and model architecture fixed. We evaluated DKM-based compression on various DNN models for computer vision and natural language processing (NLP) tasks. Our results demonstrate that DKM delivers superior compression and accuracy trade-off on ImageNet1k and GLUE benchmarks. For example, DKM-based compression can offer 74.5% top-1 ImageNet1k accuracy on ResNet50 DNN model with 3.3MB model size (29.4x model compression factor). For MobileNet-v1, which is a challenging DNN to compress, DKM delivers 63.9% top-1 ImageNet1k accuracy with 0.72 MB model size (22.4x model compression factor). This result is 6.8% higher top-1accuracy and 33% relatively smaller model size than the current state-of-the-art DNN compression algorithms. Additionally, DKM enables compression of DistilBERT model by 11.8x with minimal (1.1%) accuracy loss on GLUE NLP benchmarks., Comment: ICLR 2022

Recent observations have advanced our understanding of the neural network optimization landscape, revealing the existence of (1) paths of high accuracy containing diverse solutions and (2) wider minima offering improved performance. Previous methods observing diverse paths require multiple training runs. In contrast we aim to leverage both property (1) and (2) with a single method and in a single training run. With a similar computational cost as training one model, we learn lines, curves, and simplexes of high-accuracy neural networks. These neural network subspaces contain diverse solutions that can be ensembled, approaching the ensemble performance of independently trained networks without the training cost. Moreover, using the subspace midpoint boosts accuracy, calibration, and robustness to label noise, outperforming Stochastic Weight Averaging.

We present a computationally efficient method for compressing a trained neural network without using real data. We break the problem of data-free network compression into independent layer-wise compressions. We show how to efficiently generate layer-wise training data using only a pretrained network. We use this data to perform independent layer-wise compressions on the pretrained network. We also show how to precondition the network to improve the accuracy of our layer-wise compression method. We present results for layer-wise compression using quantization and pruning. When quantizing, we compress with higher accuracy than related works while using orders of magnitude less compute. When compressing MobileNetV2 and evaluating on ImageNet, our method outperforms existing methods for quantization at all bit-widths, achieving a $+0.34\%$ improvement in $8$-bit quantization, and a stronger improvement at lower bit-widths (up to a $+28.50\%$ improvement at $5$ bits). When pruning, we outperform baselines of a similar compute envelope, achieving $1.5$ times the sparsity rate at the same accuracy. We also show how to combine our efficient method with high-compute generative methods to improve upon their results.

We present the Supermasks in Superposition (SupSup) model, capable of sequentially learning thousands of tasks without catastrophic forgetting. Our approach uses a randomly initialized, fixed base network and for each task finds a subnetwork (supermask) that achieves good performance. If task identity is given at test time, the correct subnetwork can be retrieved with minimal memory usage. If not provided, SupSup can infer the task using gradient-based optimization to find a linear superposition of learned supermasks which minimizes the output entropy. In practice we find that a single gradient step is often sufficient to identify the correct mask, even among 2500 tasks. We also showcase two promising extensions. First, SupSup models can be trained entirely without task identity information, as they may detect when they are uncertain about new data and allocate an additional supermask for the new training distribution. Finally the entire, growing set of supermasks can be stored in a constant-sized reservoir by implicitly storing them as attractors in a fixed-sized Hopfield network., Comment: NeurIPS 2020 Camera Ready

Training a neural network is synonymous with learning the values of the weights. By contrast, we demonstrate that randomly weighted neural networks contain subnetworks which achieve impressive performance without ever training the weight values. Hidden in a randomly weighted Wide ResNet-50 we show that there is a subnetwork (with random weights) that is smaller than, but matches the performance of a ResNet-34 trained on ImageNet. Not only do these "untrained subnetworks" exist, but we provide an algorithm to effectively find them. We empirically show that as randomly weighted neural networks with fixed weights grow wider and deeper, an "untrained subnetwork" approaches a network with learned weights in accuracy. Our code and pretrained models are available at https://github.com/allenai/hidden-networks., Comment: Accepted to CVPR 2020

For sequence models with large vocabularies, a majority of network parameters lie in the input and output layers. In this work, we describe a new method, DeFINE, for learning deep token representations efficiently. Our architecture uses a hierarchical structure with novel skip-connections which allows for the use of low dimensional input and output layers, reducing total parameters and training time while delivering similar or better performance versus existing methods. DeFINE can be incorporated easily in new or existing sequence models. Compared to state-of-the-art methods including adaptive input representations, this technique results in a 6% to 20% drop in perplexity. On WikiText-103, DeFINE reduces the total parameters of Transformer-XL by half with minimal impact on performance. On the Penn Treebank, DeFINE improves AWD-LSTM by 4 points with a 17% reduction in parameters, achieving comparable performance to state-of-the-art methods with fewer parameters. For machine translation, DeFINE improves the efficiency of the Transformer model by about 1.4 times while delivering similar performance., Comment: Accepted at ICLR 2020

We present a simple and effective learning technique that significantly improves mAP of YOLO object detectors without compromising their speed. During network training, we carefully feed in localization information. We excite certain activations in order to help the network learn to better localize. In the later stages of training, we gradually reduce our assisted excitation to zero. We reached a new state-of-the-art in the speed-accuracy trade-off. Our technique improves the mAP of YOLOv2 by 3.8% and mAP of YOLOv3 by 2.2% on MSCOCO dataset.This technique is inspired from curriculum learning. It is simple and effective and it is applicable to most single-stage object detectors.

We introduce a novel and generic convolutional unit, DiCE unit, that is built using dimension-wise convolutions and dimension-wise fusion. The dimension-wise convolutions apply light-weight convolutional filtering across each dimension of the input tensor while dimension-wise fusion efficiently combines these dimension-wise representations; allowing the DiCE unit to efficiently encode spatial and channel-wise information contained in the input tensor. The DiCE unit is simple and can be seamlessly integrated with any architecture to improve its efficiency and performance. Compared to depth-wise separable convolutions, the DiCE unit shows significant improvements across different architectures. When DiCE units are stacked to build the DiCENet model, we observe significant improvements over state-of-the-art models across various computer vision tasks including image classification, object detection, and semantic segmentation. On the ImageNet dataset, the DiCENet delivers 2-4% higher accuracy than state-of-the-art manually designed models (e.g., MobileNetv2 and ShuffleNetv2). Also, DiCENet generalizes better to tasks (e.g., object detection) that are often used in resource-constrained devices in comparison to state-of-the-art separable convolution-based efficient networks, including neural search-based methods (e.g., MobileNetv3 and MixNet. Our source code in PyTorch is open-source and is available at https://github.com/sacmehta/EdgeNets/, Comment: Accepted at IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

In this paper, we show that extending the butterfly operations from the FFT algorithm to a general Butterfly Transform (BFT) can be beneficial in building an efficient block structure for CNN designs. Pointwise convolutions, which we refer to as channel fusions, are the main computational bottleneck in the state-of-the-art efficient CNNs (e.g. MobileNets ). We introduce a set of criteria for channel fusion and prove that BFT yields an asymptotically optimal FLOP count with respect to these criteria. By replacing pointwise convolutions with BFT, we reduce the computational complexity of these layers from O(n^2) to O(n\log n) with respect to the number of channels. Our experimental evaluations show that our method results in significant accuracy gains across a wide range of network architectures, especially at low FLOP ranges. For example, BFT results in up to a 6.75% absolute Top-1 improvement for MobileNetV1, 4.4 \% for ShuffleNet V2 and 5.4% for MobileNetV3 on ImageNet under a similar number of FLOPS. Notably, ShuffleNet-V2+BFT outperforms state-of-the-art architecture search methods MNasNet, FBNet and MobilenetV3 in the low FLOP regime.

The success of neural networks has driven a shift in focus from feature engineering to architecture engineering. However, successful networks today are constructed using a small and manually defined set of building blocks. Even in methods of neural architecture search (NAS) the network connectivity patterns are largely constrained. In this work we propose a method for discovering neural wirings. We relax the typical notion of layers and instead enable channels to form connections independent of each other. This allows for a much larger space of possible networks. The wiring of our network is not fixed during training -- as we learn the network parameters we also learn the structure itself. Our experiments demonstrate that our learned connectivity outperforms hand engineered and randomly wired networks. By learning the connectivity of MobileNetV1we boost the ImageNet accuracy by 10% at ~41M FLOPs. Moreover, we show that our method generalizes to recurrent and continuous time networks. Our work may also be regarded as unifying core aspects of the neural architecture search problem with sparse neural network learning. As NAS becomes more fine grained, finding a good architecture is akin to finding a sparse subnetwork of the complete graph. Accordingly, DNW provides an effective mechanism for discovering sparse subnetworks of predefined architectures in a single training run. Though we only ever use a small percentage of the weights during the forward pass, we still play the so-called initialization lottery with a combinatorial number of subnetworks. Code and pretrained models are available at https://github.com/allenai/dnw while additional visualizations may be found at https://mitchellnw.github.io/blog/2019/dnw/., Comment: NeurIPS 2019 Camera Ready

Visual Question Answering (VQA) in its ideal form lets us study reasoning in the joint space of vision and language and serves as a proxy for the AI task of scene understanding. However, most VQA benchmarks to date are focused on questions such as simple counting, visual attributes, and object detection that do not require reasoning or knowledge beyond what is in the image. In this paper, we address the task of knowledge-based visual question answering and provide a benchmark, called OK-VQA, where the image content is not sufficient to answer the questions, encouraging methods that rely on external knowledge resources. Our new dataset includes more than 14,000 questions that require external knowledge to answer. We show that the performance of the state-of-the-art VQA models degrades drastically in this new setting. Our analysis shows that our knowledge-based VQA task is diverse, difficult, and large compared to previous knowledge-based VQA datasets. We hope that this dataset enables researchers to open up new avenues for research in this domain. See http://okvqa.allenai.org to download and browse the dataset., Comment: CVPR 2019

Collaboration is a necessary skill to perform tasks that are beyond one agent's capabilities. Addressed extensively in both conventional and modern AI, multi-agent collaboration has often been studied in the context of simple grid worlds. We argue that there are inherently visual aspects to collaboration which should be studied in visually rich environments. A key element in collaboration is communication that can be either explicit, through messages, or implicit, through perception of the other agents and the visual world. Learning to collaborate in a visual environment entails learning (1) to perform the task, (2) when and what to communicate, and (3) how to act based on these communications and the perception of the visual world. In this paper we study the problem of learning to collaborate directly from pixels in AI2-THOR and demonstrate the benefits of explicit and implicit modes of communication to perform visual tasks. Refer to our project page for more details: https://prior.allenai.org/projects/two-body-problem, Comment: Accepted to CVPR 2019

Scale variation has been a challenge from traditional to modern approaches in computer vision. Most solutions to scale issues have a similar theme: a set of intuitive and manually designed policies that are generic and fixed (e.g. SIFT or feature pyramid). We argue that the scaling policy should be learned from data. In this paper, we introduce ELASTIC, a simple, efficient and yet very effective approach to learn a dynamic scale policy from data. We formulate the scaling policy as a non-linear function inside the network's structure that (a) is learned from data, (b) is instance specific, (c) does not add extra computation, and (d) can be applied on any network architecture. We applied ELASTIC to several state-of-the-art network architectures and showed consistent improvement without extra (sometimes even lower) computation on ImageNet classification, MSCOCO multi-label classification, and PASCAL VOC semantic segmentation. Our results show major improvement for images with scale challenges. Our code is available here: https://github.com/allenai/elastic, Comment: CVPR 2019 oral, code available https://github.com/allenai/elastic

Learning is an inherently continuous phenomenon. When humans learn a new task there is no explicit distinction between training and inference. As we learn a task, we keep learning about it while performing the task. What we learn and how we learn it varies during different stages of learning. Learning how to learn and adapt is a key property that enables us to generalize effortlessly to new settings. This is in contrast with conventional settings in machine learning where a trained model is frozen during inference. In this paper we study the problem of learning to learn at both training and test time in the context of visual navigation. A fundamental challenge in navigation is generalization to unseen scenes. In this paper we propose a self-adaptive visual navigation method (SAVN) which learns to adapt to new environments without any explicit supervision. Our solution is a meta-reinforcement learning approach where an agent learns a self-supervised interaction loss that encourages effective navigation. Our experiments, performed in the AI2-THOR framework, show major improvements in both success rate and SPL for visual navigation in novel scenes. Our code and data are available at: https://github.com/allenai/savn .

We introduce a light-weight, power efficient, and general purpose convolutional neural network, ESPNetv2, for modeling visual and sequential data. Our network uses group point-wise and depth-wise dilated separable convolutions to learn representations from a large effective receptive field with fewer FLOPs and parameters. The performance of our network is evaluated on four different tasks: (1) object classification, (2) semantic segmentation, (3) object detection, and (4) language modeling. Experiments on these tasks, including image classification on the ImageNet and language modeling on the PenTree bank dataset, demonstrate the superior performance of our method over the state-of-the-art methods. Our network outperforms ESPNet by 4-5% and has 2-4x fewer FLOPs on the PASCAL VOC and the Cityscapes dataset. Compared to YOLOv2 on the MS-COCO object detection, ESPNetv2 delivers 4.4% higher accuracy with 6x fewer FLOPs. Our experiments show that ESPNetv2 is much more power efficient than existing state-of-the-art efficient methods including ShuffleNets and MobileNets. Our code is open-source and available at https://github.com/sacmehta/ESPNetv2, Comment: Accepted at CVPR'19

LSTMs are powerful tools for modeling contextual information, as evidenced by their success at the task of language modeling. However, modeling contexts in very high dimensional space can lead to poor generalizability. We introduce the Pyramidal Recurrent Unit (PRU), which enables learning representations in high dimensional space with more generalization power and fewer parameters. PRUs replace the linear transformation in LSTMs with more sophisticated interactions including pyramidal and grouped linear transformations. This architecture gives strong results on word-level language modeling while reducing the number of parameters significantly. In particular, PRU improves the perplexity of a recent state-of-the-art language model Merity et al. (2018) by up to 1.3 points while learning 15-20% fewer parameters. For similar number of model parameters, PRU outperforms all previous RNN models that exploit different gating mechanisms and transformations. We provide a detailed examination of the PRU and its behavior on the language modeling tasks. Our code is open-source and available at https://sacmehta.github.io/PRU/, Comment: Accepted as a long paper in EMNLP 2018

Among the three main components (data, labels, and models) of any supervised learning system, data and models have been the main subjects of active research. However, studying labels and their properties has received very little attention. Current principles and paradigms of labeling impose several challenges to machine learning algorithms. Labels are often incomplete, ambiguous, and redundant. In this paper we study the effects of various properties of labels and introduce the Label Refinery: an iterative procedure that updates the ground truth labels after examining the entire dataset. We show significant gain using refined labels across a wide range of models. Using a Label Refinery improves the state-of-the-art top-1 accuracy of (1) AlexNet from 59.3 to 67.2, (2) MobileNet from 70.6 to 73.39, (3) MobileNet-0.25 from 50.6 to 55.59, (4) VGG19 from 72.7 to 75.46, and (5) Darknet19 from 72.9 to 74.47.

We introduce a fast and efficient convolutional neural network, ESPNet, for semantic segmentation of high resolution images under resource constraints. ESPNet is based on a new convolutional module, efficient spatial pyramid (ESP), which is efficient in terms of computation, memory, and power. ESPNet is 22 times faster (on a standard GPU) and 180 times smaller than the state-of-the-art semantic segmentation network PSPNet, while its category-wise accuracy is only 8% less. We evaluated ESPNet on a variety of semantic segmentation datasets including Cityscapes, PASCAL VOC, and a breast biopsy whole slide image dataset. Under the same constraints on memory and computation, ESPNet outperforms all the current efficient CNN networks such as MobileNet, ShuffleNet, and ENet on both standard metrics and our newly introduced performance metrics that measure efficiency on edge devices. Our network can process high resolution images at a rate of 112 and 9 frames per second on a standard GPU and edge device, respectively., Comment: Accepted at ECCV'18

We introduce Interactive Question Answering (IQA), the task of answering questions that require an autonomous agent to interact with a dynamic visual environment. IQA presents the agent with a scene and a question, like: "Are there any apples in the fridge?" The agent must navigate around the scene, acquire visual understanding of scene elements, interact with objects (e.g. open refrigerators) and plan for a series of actions conditioned on the question. Popular reinforcement learning approaches with a single controller perform poorly on IQA owing to the large and diverse state space. We propose the Hierarchical Interactive Memory Network (HIMN), consisting of a factorized set of controllers, allowing the system to operate at multiple levels of temporal abstraction. To evaluate HIMN, we introduce IQUAD V1, a new dataset built upon AI2-THOR, a simulated photo-realistic environment of configurable indoor scenes with interactive objects (code and dataset available at https://github.com/danielgordon10/thor-iqa-cvpr-2018). IQUAD V1 has 75,000 questions, each paired with a unique scene configuration. Our experiments show that our proposed model outperforms popular single controller based methods on IQUAD V1. For sample questions and results, please view our video: https://youtu.be/pXd3C-1jr98, Comment: Published in CVPR 2018

Porting state of the art deep learning algorithms to resource constrained compute platforms (e.g. VR, AR, wearables) is extremely challenging. We propose a fast, compact, and accurate model for convolutional neural networks that enables efficient learning and inference. We introduce LCNN, a lookup-based convolutional neural network that encodes convolutions by few lookups to a dictionary that is trained to cover the space of weights in CNNs. Training LCNN involves jointly learning a dictionary and a small set of linear combinations. The size of the dictionary naturally traces a spectrum of trade-offs between efficiency and accuracy. Our experimental results on ImageNet challenge show that LCNN can offer 3.2x speedup while achieving 55.1% top-1 accuracy using AlexNet architecture. Our fastest LCNN offers 37.6x speed up over AlexNet while maintaining 44.3% top-1 accuracy. LCNN not only offers dramatic speed ups at inference, but it also enables efficient training. In this paper, we show the benefits of LCNN in few-shot learning and few-iteration learning, two crucial aspects of on-device training of deep learning models., Comment: CVPR 17