Hanlin Goh

Shuangfei Zhai, Navdeep Jaitly, Jason Ramapuram et al. · apple-ml

Transformers have gained increasing popularity in a wide range of applications, including Natural Language Processing (NLP), Computer Vision and Speech Recognition, because of their powerful representational capacity. However, harnessing this representational capacity effectively requires a large amount of data, strong regularization, or both, to mitigate overfitting. Recently, the power of the Transformer has been unlocked by self-supervised pretraining strategies based on masked autoencoders which rely on reconstructing masked inputs, directly, or contrastively from unmasked content. This pretraining strategy which has been used in BERT models in NLP, Wav2Vec models in Speech and, recently, in MAE models in Vision, forces the model to learn about relationships between the content in different parts of the input using autoencoding related objectives. In this paper, we propose a novel, but surprisingly simple alternative to content reconstruction~-- that of predicting locations from content, without providing positional information for it. Doing so requires the Transformer to understand the positional relationships between different parts of the input, from their content alone. This amounts to an efficient implementation where the pretext task is a classification problem among all possible positions for each input token. We experiment on both Vision and Speech benchmarks, where our approach brings improvements over strong supervised training baselines and is comparable to modern unsupervised/self-supervised pretraining methods. Our method also enables Transformers trained without position embeddings to outperform ones trained with full position information.

Miguel Angel Bautista, Pengsheng Guo, Samira Abnar et al. · apple-ml

We introduce GAUDI, a generative model capable of capturing the distribution of complex and realistic 3D scenes that can be rendered immersively from a moving camera. We tackle this challenging problem with a scalable yet powerful approach, where we first optimize a latent representation that disentangles radiance fields and camera poses. This latent representation is then used to learn a generative model that enables both unconditional and conditional generation of 3D scenes. Our model generalizes previous works that focus on single objects by removing the assumption that the camera pose distribution can be shared across samples. We show that GAUDI obtains state-of-the-art performance in the unconditional generative setting across multiple datasets and allows for conditional generation of 3D scenes given conditioning variables like sparse image observations or text that describes the scene.

Chen Huang, Hanlin Goh, Jiatao Gu et al. · meta-ai

Recent Self-Supervised Learning (SSL) methods are able to learn feature representations that are invariant to different data augmentations, which can then be transferred to downstream tasks of interest. However, different downstream tasks require different invariances for their best performance, so the optimal choice of augmentations for SSL depends on the target task. In this paper, we aim to learn self-supervised features that generalize well across a variety of downstream tasks (e.g., object classification, detection and instance segmentation) without knowing any task information beforehand. We do so by Masked Augmentation Subspace Training (or MAST) to encode in the single feature space the priors from different data augmentations in a factorized way. Specifically, we disentangle the feature space into separate subspaces, each induced by a learnable mask that selects relevant feature dimensions to model invariance to a specific augmentation. We show the success of MAST in jointly capturing generalizable priors from different augmentations, using both unique and shared features across the subspaces. We further show that MAST benefits from uncertainty modeling to reweight ambiguous samples from strong augmentations that may cause similarity mismatch in each subspace. Experiments demonstrate that MAST consistently improves generalization on various downstream tasks, while being task-agnostic and efficient during SSL. We also provide interesting insights about how different augmentations are related and how uncertainty reflects learning difficulty.

Ran Liu, Ellen L. Zippi, Hadi Pouransari et al.

Leveraging multimodal information from biosignals is vital for building a comprehensive representation of people's physical and mental states. However, multimodal biosignals often exhibit substantial distributional shifts between pretraining and inference datasets, stemming from changes in task specification or variations in modality compositions. To achieve effective pretraining in the presence of potential distributional shifts, we propose a frequency-aware masked autoencoder ($\texttt{bio}$FAME) that learns to parameterize the representation of biosignals in the frequency space. $\texttt{bio}$FAME incorporates a frequency-aware transformer, which leverages a fixed-size Fourier-based operator for global token mixing, independent of the length and sampling rate of inputs. To maintain the frequency components within each input channel, we further employ a frequency-maintain pretraining strategy that performs masked autoencoding in the latent space. The resulting architecture effectively utilizes multimodal information during pretraining, and can be seamlessly adapted to diverse tasks and modalities at test time, regardless of input size and order. We evaluated our approach on a diverse set of transfer experiments on unimodal time series, achieving an average of $\uparrow$5.5% improvement in classification accuracy over the previous state-of-the-art. Furthermore, we demonstrated that our architecture is robust in modality mismatch scenarios, including unpredicted modality dropout or substitution, proving its practical utility in real-world applications. Code is available at https://github.com/apple/ml-famae .

Yao-Hung Hubert Tsai, Hanlin Goh, Ali Farhadi et al.

The perception system in personalized mobile agents requires developing indoor scene understanding models, which can understand 3D geometries, capture objectiveness, analyze human behaviors, etc. Nonetheless, this direction has not been well-explored in comparison with models for outdoor environments (e.g., the autonomous driving system that includes pedestrian prediction, car detection, traffic sign recognition, etc.). In this paper, we first discuss the main challenge: insufficient, or even no, labeled data for real-world indoor environments, and other challenges such as fusion between heterogeneous sources of information (e.g., RGB images and Lidar point clouds), modeling relationships between a diverse set of outputs (e.g., 3D object locations, depth estimation, and human poses), and computational efficiency. Then, we describe MMISM (Multi-modality input Multi-task output Indoor Scene understanding Model) to tackle the above challenges. MMISM considers RGB images as well as sparse Lidar points as inputs and 3D object detection, depth completion, human pose estimation, and semantic segmentation as output tasks. We show that MMISM performs on par or even better than single-task models; e.g., we improve the baseline 3D object detection results by 11.7% on the benchmark ARKitScenes dataset.

Hsiang-Yun Sherry Chien, Hanlin Goh, Christopher M. Sandino et al.

Decoding information from bio-signals such as EEG, using machine learning has been a challenge due to the small data-sets and difficulty to obtain labels. We propose a reconstruction-based self-supervised learning model, the masked auto-encoder for EEG (MAEEG), for learning EEG representations by learning to reconstruct the masked EEG features using a transformer architecture. We found that MAEEG can learn representations that significantly improve sleep stage classification (~5% accuracy increase) when only a small number of labels are given. We also found that input sample lengths and different ways of masking during reconstruction-based SSL pretraining have a huge effect on downstream model performance. Specifically, learning to reconstruct a larger proportion and more concentrated masked signal results in better performance on sleep classification. Our findings provide insight into how reconstruction-based SSL could help representation learning for EEG.

Yuhang Zang, Hanlin Goh, Josh Susskind et al.

Existing vision-language models exhibit strong generalization on a variety of visual domains and tasks. However, such models mainly perform zero-shot recognition in a closed-set manner, and thus struggle to handle open-domain visual concepts by design. There are recent finetuning methods, such as prompt learning, that not only study the discrimination between in-distribution (ID) and out-of-distribution (OOD) samples, but also show some improvements in both ID and OOD accuracies. In this paper, we first demonstrate that vision-language models, after long enough finetuning but without proper regularization, tend to overfit the known classes in the given dataset, with degraded performance on unknown classes. Then we propose a novel approach OGEN to address this pitfall, with the main focus on improving the OOD GENeralization of finetuned models. Specifically, a class-conditional feature generator is introduced to synthesize OOD features using just the class name of any unknown class. Such synthesized features will provide useful knowledge about unknowns and help regularize the decision boundary between ID and OOD data when optimized jointly. Equally important is our adaptive self-distillation mechanism to regularize our feature generation model during joint optimization, i.e., adaptively transferring knowledge between model states to further prevent overfitting. Experiments validate that our method yields convincing gains in OOD generalization performance in different settings. Code: https://github.com/apple/ml-ogen.

Christopher Sandino, Sayeri Lala, Geeling Chau et al.

Self-supervised learning (SSL) offers a promising approach for learning electroencephalography (EEG) representations from unlabeled data, reducing the need for expensive annotations for clinical applications like sleep staging and seizure detection. While current EEG SSL methods predominantly use masked reconstruction strategies like masked autoencoders (MAE) that capture local temporal patterns, position prediction pretraining remains underexplored despite its potential to learn long-range dependencies in neural signals. We introduce PAirwise Relative Shift or PARS pretraining, a novel pretext task that predicts relative temporal shifts between randomly sampled EEG window pairs. Unlike reconstruction-based methods that focus on local pattern recovery, PARS encourages encoders to capture relative temporal composition and long-range dependencies inherent in neural signals. Through comprehensive evaluation on various EEG decoding tasks, we demonstrate that PARS-pretrained transformers consistently outperform existing pretraining strategies in label-efficient and transfer learning settings, establishing a new paradigm for self-supervised EEG representation learning.

Yao-Hung Hubert Tsai, Nitish Srivastava, Hanlin Goh et al.

We introduce a new routing algorithm for capsule networks, in which a child capsule is routed to a parent based only on agreement between the parent's state and the child's vote. The new mechanism 1) designs routing via inverted dot-product attention; 2) imposes Layer Normalization as normalization; and 3) replaces sequential iterative routing with concurrent iterative routing. When compared to previously proposed routing algorithms, our method improves performance on benchmark datasets such as CIFAR-10 and CIFAR-100, and it performs at-par with a powerful CNN (ResNet-18) with 4x fewer parameters. On a different task of recognizing digits from overlayed digit images, the proposed capsule model performs favorably against CNNs given the same number of layers and neurons per layer. We believe that our work raises the possibility of applying capsule networks to complex real-world tasks. Our code is publicly available at: https://github.com/apple/ml-capsules-inverted-attention-routing An alternative implementation is available at: https://github.com/yaohungt/Capsules-Inverted-Attention-Routing/blob/master/README.md

Vimal Thilak, Chen Huang, Omid Saremi et al. · apple-ml

Joint embedding (JE) architectures have emerged as a promising avenue for acquiring transferable data representations. A key obstacle to using JE methods, however, is the inherent challenge of evaluating learned representations without access to a downstream task, and an annotated dataset. Without efficient and reliable evaluation, it is difficult to iterate on architectural and training choices for JE methods. In this paper, we introduce LiDAR (Linear Discriminant Analysis Rank), a metric designed to measure the quality of representations within JE architectures. Our metric addresses several shortcomings of recent approaches based on feature covariance rank by discriminating between informative and uninformative features. In essence, LiDAR quantifies the rank of the Linear Discriminant Analysis (LDA) matrix associated with the surrogate SSL task -- a measure that intuitively captures the information content as it pertains to solving the SSL task. We empirically demonstrate that LiDAR significantly surpasses naive rank based approaches in its predictive power of optimal hyperparameters. Our proposed criterion presents a more robust and intuitive means of assessing the quality of representations within JE architectures, which we hope facilitates broader adoption of these powerful techniques in various domains.

Melika Ayoughi, Samira Abnar, Chen Huang et al. · apple-ml

The composition of objects and their parts, along with object-object positional relationships, provides a rich source of information for representation learning. Hence, spatial-aware pretext tasks have been actively explored in self-supervised learning. Existing works commonly start from a grid structure, where the goal of the pretext task involves predicting the absolute position index of patches within a fixed grid. However, grid-based approaches fall short of capturing the fluid and continuous nature of real-world object compositions. We introduce PART, a self-supervised learning approach that leverages continuous relative transformations between off-grid patches to overcome these limitations. By modeling how parts relate to each other in a continuous space, PART learns the relative composition of images-an off-grid structural relative positioning process that generalizes beyond occlusions and deformations. In tasks requiring precise spatial understanding such as object detection and time series prediction, PART outperforms strong grid-based methods like MAE and DropPos, while also maintaining competitive performance on global classification tasks with minimal hyperparameter tuning. By breaking free from grid constraints, PART opens up an exciting new trajectory for universal self-supervised pretraining across diverse datatypes-from natural images to EEG signals-with promising potential in video, medical imaging, and audio.

Sayeri Lala, Hanlin Goh, Christopher Sandino

Sleep staging is a clinically important task for diagnosing various sleep disorders, but remains challenging to deploy at scale because it because it is both labor-intensive and time-consuming. Supervised deep learning-based approaches can automate sleep staging but at the expense of large labeled datasets, which can be unfeasible to procure for various settings, e.g., uncommon sleep disorders. While self-supervised learning (SSL) can mitigate this need, recent studies on SSL for sleep staging have shown performance gains saturate after training with labeled data from only tens of subjects, hence are unable to match peak performance attained with larger datasets. We hypothesize that the rapid saturation stems from applying a sub-optimal pretraining scheme that pretrains only a portion of the architecture, i.e., the feature encoder, but not the temporal encoder; therefore, we propose adopting an architecture that seamlessly couples the feature and temporal encoding and a suitable pretraining scheme that pretrains the entire model. On a sample sleep staging dataset, we find that the proposed scheme offers performance gains that do not saturate with amount of labeled training data (e.g., 3-5\% improvement in balanced sleep staging accuracy across low- to high-labeled data settings), reducing the amount of labeled training data needed for high performance (e.g., by 800 subjects). Based on our findings, we recommend adopting this SSL paradigm for subsequent work on SSL for sleep staging.

Etai Littwin, Omid Saremi, Shuangfei Zhai et al.

We analyze the learning dynamics of infinitely wide neural networks with a finite sized bottle-neck. Unlike the neural tangent kernel limit, a bottleneck in an otherwise infinite width network al-lows data dependent feature learning in its bottle-neck representation. We empirically show that a single bottleneck in infinite networks dramatically accelerates training when compared to purely in-finite networks, with an improved overall performance. We discuss the acceleration phenomena by drawing similarities to infinitely wide deep linear models, where the acceleration effect of a bottleneck can be understood theoretically.

Shuangfei Zhai, Walter Talbott, Nitish Srivastava et al.

We introduce Attention Free Transformer (AFT), an efficient variant of Transformers that eliminates the need for dot product self attention. In an AFT layer, the key and value are first combined with a set of learned position biases, the result of which is multiplied with the query in an element-wise fashion. This new operation has a memory complexity linear w.r.t. both the context size and the dimension of features, making it compatible to both large input and model sizes. We also introduce AFT-local and AFT-conv, two model variants that take advantage of the idea of locality and spatial weight sharing while maintaining global connectivity. We conduct extensive experiments on two autoregressive modeling tasks (CIFAR10 and Enwik8) as well as an image recognition task (ImageNet-1K classification). We show that AFT demonstrates competitive performance on all the benchmarks, while providing excellent efficiency at the same time.

Yue Wu, Shuangfei Zhai, Nitish Srivastava et al.

Offline Reinforcement Learning promises to learn effective policies from previously-collected, static datasets without the need for exploration. However, existing Q-learning and actor-critic based off-policy RL algorithms fail when bootstrapping from out-of-distribution (OOD) actions or states. We hypothesize that a key missing ingredient from the existing methods is a proper treatment of uncertainty in the offline setting. We propose Uncertainty Weighted Actor-Critic (UWAC), an algorithm that detects OOD state-action pairs and down-weights their contribution in the training objectives accordingly. Implementation-wise, we adopt a practical and effective dropout-based uncertainty estimation method that introduces very little overhead over existing RL algorithms. Empirically, we observe that UWAC substantially improves model stability during training. In addition, UWAC out-performs existing offline RL methods on a variety of competitive tasks, and achieves significant performance gains over the state-of-the-art baseline on datasets with sparse demonstrations collected from human experts.

Joseph Y. Cheng, Hanlin Goh, Kaan Dogrusoz et al.

Datasets for biosignals, such as electroencephalogram (EEG) and electrocardiogram (ECG), often have noisy labels and have limited number of subjects (<100). To handle these challenges, we propose a self-supervised approach based on contrastive learning to model biosignals with a reduced reliance on labeled data and with fewer subjects. In this regime of limited labels and subjects, intersubject variability negatively impacts model performance. Thus, we introduce subject-aware learning through (1) a subject-specific contrastive loss, and (2) an adversarial training to promote subject-invariance during the self-supervised learning. We also develop a number of time-series data augmentation techniques to be used with the contrastive loss for biosignals. Our method is evaluated on publicly available datasets of two different biosignals with different tasks: EEG decoding and ECG anomaly detection. The embeddings learned using self-supervision yield competitive classification results compared to entirely supervised methods. We show that subject-invariance improves representation quality for these tasks, and observe that subject-specific loss increases performance when fine-tuning with supervised labels.

Nitish Srivastava, Hanlin Goh, Ruslan Salakhutdinov

We propose a method to learn object representations from 3D point clouds using bundles of geometrically interpretable hidden units, which we call geometric capsules. Each geometric capsule represents a visual entity, such as an object or a part, and consists of two components: a pose and a feature. The pose encodes where the entity is, while the feature encodes what it is. We use these capsules to construct a Geometric Capsule Autoencoder that learns to group 3D points into parts (small local surfaces), and these parts into the whole object, in an unsupervised manner. Our novel Multi-View Agreement voting mechanism is used to discover an object's canonical pose and its pose-invariant feature vector. Using the ShapeNet and ModelNet40 datasets, we analyze the properties of the learned representations and show the benefits of having multiple votes agree. We perform alignment and retrieval of arbitrarily rotated objects -- tasks that evaluate our model's object identification and canonical pose recovery capabilities -- and obtained insightful results.

Vijay Chandrasekhar, Jie Lin, Olivier Morère et al.

With deep learning becoming the dominant approach in computer vision, the use of representations extracted from Convolutional Neural Nets (CNNs) is quickly gaining ground on Fisher Vectors (FVs) as favoured state-of-the-art global image descriptors for image instance retrieval. While the good performance of CNNs for image classification are unambiguously recognised, which of the two has the upper hand in the image retrieval context is not entirely clear yet. In this work, we propose a comprehensive study that systematically evaluates FVs and CNNs for image retrieval. The first part compares the performances of FVs and CNNs on multiple publicly available data sets. We investigate a number of details specific to each method. For FVs, we compare sparse descriptors based on interest point detectors with dense single-scale and multi-scale variants. For CNNs, we focus on understanding the impact of depth, architecture and training data on retrieval results. Our study shows that no descriptor is systematically better than the other and that performance gains can usually be obtained by using both types together. The second part of the study focuses on the impact of geometrical transformations such as rotations and scale changes. FVs based on interest point detectors are intrinsically resilient to such transformations while CNNs do not have a built-in mechanism to ensure such invariance. We show that performance of CNNs can quickly degrade in presence of rotations while they are far less affected by changes in scale. We then propose a number of ways to incorporate the required invariances in the CNN pipeline. Overall, our work is intended as a reference guide offering practically useful and simply implementable guidelines to anyone looking for state-of-the-art global descriptors best suited to their specific image instance retrieval problem.

Olivier Morère, Hanlin Goh, Antoine Veillard et al.

Compact keyframe-based video summaries are a popular way of generating viewership on video sharing platforms. Yet, creating relevant and compelling summaries for arbitrarily long videos with a small number of keyframes is a challenging task. We propose a comprehensive keyframe-based summarization framework combining deep convolutional neural networks and restricted Boltzmann machines. An original co-regularization scheme is used to discover meaningful subject-scene associations. The resulting multimodal representations are then used to select highly-relevant keyframes. A comprehensive user study is conducted comparing our proposed method to a variety of schemes, including the summarization currently in use by one of the most popular video sharing websites. The results show that our method consistently outperforms the baseline schemes for any given amount of keyframes both in terms of attractiveness and informativeness. The lead is even more significant for smaller summaries.

Jie Lin, Olivier Morere, Vijay Chandrasekhar et al.

This work focuses on representing very high-dimensional global image descriptors using very compact 64-1024 bit binary hashes for instance retrieval. We propose DeepHash: a hashing scheme based on deep networks. Key to making DeepHash work at extremely low bitrates are three important considerations -- regularization, depth and fine-tuning -- each requiring solutions specific to the hashing problem. In-depth evaluation shows that our scheme consistently outperforms state-of-the-art methods across all data sets for both Fisher Vectors and Deep Convolutional Neural Network features, by up to 20 percent over other schemes. The retrieval performance with 256-bit hashes is close to that of the uncompressed floating point features -- a remarkable 512 times compression.