Jeffrey Li, Mohammadreza Armandpour, Seyed Iman Mirzadeh, Sachin Mehta, Vaishaal Shankar, Raviteja Vemulapalli, Samy Bengio, Oncel Tuzel, Mehrdad Farajtabar, Hadi Pouransari, Fartash Faghri. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Vision foundation models pre-trained on massive data encode rich representations of real-world concepts, which can be adapted to downstream tasks by fine-tuning. However, fine-tuning foundation models on one task often leads to the issue of *concept forgetting* on other tasks. Recent methods of robust fine-tuning aim to mitigate forgetting of prior knowledge without affecting the fine-tuning performance. Knowledge is often preserved by matching the original and fine-tuned model weights or feature pairs. However, such point-wise matching can be too strong, without explicit awareness of the feature neighborhood structures that encode rich knowledge as well. We propose a novel regularization method **Proxy-FDA** that explicitly preserves the structural knowledge in feature space. Proxy-FDA performs Feature Distribution Alignment (using nearest neighbor graphs) between the pre-trained and fine-tuned feature spaces, and the alignment is further improved by informative proxies that are generated dynamically to increase data diversity. Experiments show that Proxy-FDA significantly reduces concept forgetting during fine-tuning, and we find a strong correlation between forgetting and a distributional distance metric (in comparison to L2 distance). We further demonstrate Proxy-FDA's benefits in various fine-tuning settings (end-to-end, few-shot and continual tuning) and across different tasks like image classification, captioning and VQA.

Defect segmentation is crucial for quality control in advanced manufacturing, yet data scarcity poses challenges for state-of-the-art supervised deep learning. Synthetic defect data generation is a popular approach for mitigating data challenges. However, many current methods simply generate defects following a fixed set of rules, which may not directly relate to downstream task performance. This can lead to suboptimal performance and may even hinder the downstream task. To solve this problem, we leverage a novel bi-level optimization-based synthetic defect data generation framework. We use an online synthetic defect generation module grounded in the commonly-used Cut&Paste framework, and adopt an efficient gradient-based optimization algorithm to solve the bi-level optimization problem. We achieve simultaneous training of the defect segmentation network, and learn various parameters of the data synthesis module by maximizing the validation performance of the trained defect segmentation network. Our experimental results on benchmark datasets under limited data settings show that the proposed bi-level optimization method can be used for learning the most effective locations for pasting synthetic defects thereby improving the segmentation performance by up to 21.3% when compared to pasting defects at random locations. We also demonstrate up to 2.7% performance gain by learning the importance weights for different augmentation-specific defect data sources when compared to giving equal importance to all the data sources.

Contrastive pretraining of image-text foundation models, such as CLIP, demonstrated excellent zero-shot performance and improved robustness on a wide range of downstream tasks. However, these models utilize large transformer-based encoders with significant memory and latency overhead which pose challenges for deployment on mobile devices. In this work, we introduce MobileCLIP - a new family of efficient image-text models optimized for runtime performance along with a novel and efficient training approach, namely multi-modal reinforced training. The proposed training approach leverages knowledge transfer from an image captioning model and an ensemble of strong CLIP encoders to improve the accuracy of efficient moels. Our approach avoids train-time compute overhead by storing the additional knowledge in a reinforced dataset. MobileCLIP sets a new state-of-the-art latency-accuracy tradeoff for zero-shot classification and retrieval tasks on several datasets. Our MobileCLIP-S2 variant is 2.3× faster while more accurate compared to previous best CLIP model based on ViT-B/16. We further demonstrate the effectiveness of our multi-modal reinforced training by training a CLIP model based on ViT-B/16 image backbone and achieving +2.9% average performance improvement on 38 evaluation benchmarks compared to the previous best. Moreover, we show that the proposed approach achieves 10×-1000 × improved learning efficiency when compared with non-reinforced CLIP training. Code and models are available at https://github.com/apple/ml-mobileclip

We consider the task of animating 3D facial geometry from speech signal. Existing works are primarily deterministic, focusing on learning a one-to-one mapping from speech signal to 3D face meshes on small datasets with limited speakers. While these models can achieve high-quality lip articulation for speakers in the training set, they are unable to capture the full and diverse distribution of 3D facial motions that accompany speech in the real world. Importantly, the relationship between speech and facial motion is one-to-many, containing both inter-speaker and intra-speaker variations and necessitating a probabilistic approach. In this paper, we identify and address key challenges that have so far limited the development of probabilistic models: lack of datasets and metrics that are suitable for training and evaluating them, as well as the difficulty of designing a model that generates diverse results while remaining faithful to a strong conditioning signal as speech. We first propose large-scale benchmark datasets and metrics suitable for probabilistic modeling. Then, we demonstrate a probabilistic model that achieves both diversity and fidelity to speech, outperforming other methods across the proposed benchmarks. Finally, we showcase useful applications of probabilistic models trained on these large-scale datasets: we can generate diverse speech-driven 3D facial motion that matches unseen speaker styles extracted from reference clips; and our synthetic meshes can be used to improve the performance of downstream audio-visual models.

While Automatic Speech Recognition (ASR) systems are widely used in many real-world applications, they often do not generalize well to new domains and need to be fine-tuned on data from these domains. However, target-domain data usually are not readily available in many scenarios. In this paper, we propose a new strategy for adapting ASR models to new target domains without any text or speech from those domains. To accomplish this, we propose a novel data synthesis pipeline that uses a Large Language Model (LLM) to generate a target domain text corpus, and a state-of-the-art controllable speech synthesis model to generate the corresponding speech. We propose a simple yet effective in-context instruction fine-tuning strategy to increase the effectiveness of LLM in generating text corpora for new domains. Experiments on the SLURP dataset show that the proposed method achieves an average relative word error rate improvement of 28% on unseen target domains without any performance drop in source domains.

Keeping large foundation models up to date on latest data is inherently expensive. To avoid the prohibitive costs of constantly retraining, it is imperative to continually train these models. This problem is exacerbated by the lack of any large scale continual learning benchmarks or baselines. We introduce the first set of web-scale Time-Continual (TiC) benchmarks for training vision-language models: TiC-DataComp, TiC-YFCC, and TiC-Redcaps. TiC-DataComp, our largest dataset, contains over 12.7B timestamped image-text pairs spanning 9 years (2014-2022). We first use our benchmarks to curate various dynamic evaluations to measure temporal robustness of existing models. We show OpenAI's CLIP (trained on data up to 2020) loses $\approx 8\%$ zero-shot accuracy on our curated retrieval task from 2021-2022 compared with more recently trained models in OpenCLIP repository. We then study how to efficiently train models on time-continuous data. We demonstrate that a simple rehearsal-based approach that continues training from the last checkpoint and replays old data reduces compute by $2.5\times$ when compared to the standard practice of retraining from scratch. Code is available at https://github.com/apple/ml-tic-clip.

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.

Mutual gaze detection, i.e., predicting whether or not two people are looking at each other, plays an important role in understanding human interactions. In this work, we focus on the task of image-based mutual gaze detection, and propose a simple and effective approach to boost the performance by using an auxiliary 3D gaze estimation task during the training phase. We achieve the performance boost without additional labeling cost by training the 3D gaze estimation branch using pseudo 3D gaze labels deduced from mutual gaze labels. By sharing the head image encoder between the 3D gaze estimation and the mutual gaze detection branches, we achieve better head features than learned by training the mutual gaze detection branch alone. Experimental results on three image datasets show that the proposed approach improves the detection performance significantly without additional annotations. This work also introduces a new image dataset that consists of 33.1K pairs of humans annotated with mutual gaze labels in 29.2K images.

Collecting labeled data for the task of semantic segmentation is expensive and time-consuming, as it requires dense pixel-level annotations. While recent Convolutional Neural Network (CNN) based semantic segmentation approaches have achieved impressive results by using large amounts of labeled training data, their performance drops significantly as the amount of labeled data decreases. This happens because deep CNNs trained with the de facto cross-entropy loss can easily overfit to small amounts of labeled data. To address this issue, we propose a simple and effective contrastive learning-based training strategy in which we first pretrain the network using a pixel-wise, label-based contrastive loss, and then fine-tune it using the cross-entropy loss. This approach increases intra-class compactness and inter-class separability, thereby resulting in a better pixel classifier. We demonstrate the effectiveness of the proposed training strategy using the Cityscapes and PASCAL VOC 2012 segmentation datasets. Our results show that pretraining with the proposed contrastive loss results in large performance gains (more than 20% absolute improvement in some settings) when the amount of labeled data is limited. In many settings, the proposed contrastive pretraining strategy, which does not use any additional data, is able to match or outperform the widely-used ImageNet pretraining strategy that uses more than a million additional labeled images.

Standard Knowledge Distillation (KD) approaches distill the knowledge of a cumbersome teacher model into the parameters of a student model with a pre-defined architecture. However, the knowledge of a neural network, which is represented by the network's output distribution conditioned on its input, depends not only on its parameters but also on its architecture. Hence, a more generalized approach for KD is to distill the teacher's knowledge into both the parameters and architecture of the student. To achieve this, we present a new \textit{Architecture-aware Knowledge Distillation (AKD)} approach that finds student models (pearls for the teacher) that are best for distilling the given teacher model. In particular, we leverage Neural Architecture Search (NAS), equipped with our KD-guided reward, to search for the best student architectures for a given teacher. Experimental results show our proposed AKD consistently outperforms the conventional NAS plus KD approach, and achieves state-of-the-art results on the ImageNet classification task under various latency settings. Furthermore, the best AKD student architecture for the ImageNet classification task also transfers well to other tasks such as million level face recognition and ensemble learning.

Most of the existing work on automatic facial expression analysis focuses on discrete emotion recognition, or facial action unit detection. However, facial expressions do not always fall neatly into pre-defined semantic categories. Also, the similarity between expressions measured in the action unit space need not correspond to how humans perceive expression similarity. Different from previous work, our goal is to describe facial expressions in a continuous fashion using a compact embedding space that mimics human visual preferences. To achieve this goal, we collect a large-scale faces-in-the-wild dataset with human annotations in the form: Expressions A and B are visually more similar when compared to expression C, and use this dataset to train a neural network that produces a compact (16-dimensional) expression embedding. We experimentally demonstrate that the learned embedding can be successfully used for various applications such as expression retrieval, photo album summarization, and emotion recognition. We also show that the embedding learned using the proposed dataset performs better than several other embeddings learned using existing emotion or action unit datasets.

Recent advances in video super-resolution have shown that convolutional neural networks combined with motion compensation are able to merge information from multiple low-resolution (LR) frames to generate high-quality images. Current state-of-the-art methods process a batch of LR frames to generate a single high-resolution (HR) frame and run this scheme in a sliding window fashion over the entire video, effectively treating the problem as a large number of separate multi-frame super-resolution tasks. This approach has two main weaknesses: 1) Each input frame is processed and warped multiple times, increasing the computational cost, and 2) each output frame is estimated independently conditioned on the input frames, limiting the system's ability to produce temporally consistent results. In this work, we propose an end-to-end trainable frame-recurrent video super-resolution framework that uses the previously inferred HR estimate to super-resolve the subsequent frame. This naturally encourages temporally consistent results and reduces the computational cost by warping only one image in each step. Furthermore, due to its recurrent nature, the proposed method has the ability to assimilate a large number of previous frames without increased computational demands. Extensive evaluations and comparisons with previous methods validate the strengths of our approach and demonstrate that the proposed framework is able to significantly outperform the current state of the art.

Recently, skeleton-based human action recognition has been receiving significant attention from various research communities due to the availability of depth sensors and real-time depth-based 3D skeleton estimation algorithms. In this work, we use rolling maps for recognizing human actions from 3D skeletal data. The rolling map is a well-defined mathematical concept that has not been explored much by the vision community. First, we represent each skeleton using the relative 3D rotations between various body parts. Since 3D rotations are members of the special orthogonal group SO3, our skeletal representation becomes a point in the Lie group SO3× ... × SO3, which is also a Riemannian manifold. Then, using this representation, we model human actions as curves in this Lie group. Since classification of curves in this non-Euclidean space is a difficult task, we unwrap the action curves onto the Lie algebra so3× ... × so3(which is a vector space) by combining the logarithm map with rolling maps, and perform classification in the Lie algebra. Experimental results on three action datasets show that the proposed approach performs equally well or better when compared to state-of-the-art.

We propose a novel end-to-end trainable deep network architecture for image denoising based on a Gaussian Conditional Random Field (GCRF) model. In contrast to the existing discriminative denoising methods that train a separate model for each individual noise level, the proposed deep network explicitly models the input noise variance and hence is capable of handling a range of noise levels. Our deep network, which we refer to as deep GCRF network, consists of two sub-networks: (i) a parameter generation network that generates the pairwise potential parameters based on the noisy input image, and (ii) an inference network whose layers perform the computations involved in an iterative GCRF inference procedure. We train two deep GCRF networks (each network operates over a range of noise levels: one for low input noise levels and one for high input noise levels) discriminatively by maximizing the peak signal-to-noise ratio measure. Experiments on Berkeley segmentation and PASCALVOC datasets show that the proposed approach produces results on par with the state of-the-art without training a separate network for each individual noise level.

In contrast to the existing approaches that use discrete Conditional Random Field (CRF) models, we propose to use a Gaussian CRF model for the task of semantic segmentation. We propose a novel deep network, which we refer to as Gaussian Mean Field (GMF) network, whose layers perform mean field inference over a Gaussian CRF. The proposed GMF network has the desired property that each of its layers produces an output that is closer to the maximum a posteriori solution of the Gaussian CRF compared to its input. By combining the proposed GMF network with deep Convolutional Neural Networks (CNNs), we propose a new end-to-end trainable Gaussian conditional random field network. The proposed Gaussian CRF network is composed of three sub-networks: (i) a CNN-based unary network for generating unary potentials, (ii) a CNN-based pairwise network for generating pairwise potentials, and (iii) a GMF network for performing Gaussian CRF inference. When trained end-to-end in a discriminative fashion, and evaluated on the challenging PASCALVOC 2012 segmentation dataset, the proposed Gaussian CRF network outperforms various recent semantic segmentation approaches that combine CNNs with discrete CRF models.

Recently, cross-modal synthesis of subject-specific scans has been receiving significant attention from the medical imaging community. Though various synthesis approaches have been introduced in the recent past, most of them are either tailored to a specific application or proposed for the supervised setting, i.e., they assume the availability of training data from the same set of subjects in both source and target modalities. But, collecting multiple scans from each subject is undesirable. Hence, to address this issue, we propose a general unsupervised cross-modal medical image synthesis approach that works without paired training data. Given a source modality image of a subject, we first generate multiple target modality candidate values for each voxel independently using cross-modal nearest neighbor search. Then, we select the best candidate values jointly for all the voxels by simultaneously maximizing a global mutual information cost function and a local spatial consistency cost function. Finally, we use coupled sparse representation for further refinement of synthesized images. Our experiments on generating T1-MRI brain scans from T2-MRI and vice versa demonstrate that the synthesis capability of the proposed unsupervised approach is comparable to various state-of-the-art supervised approaches in the literature.

Recently introduced cost-effective depth sensors coupled with the real-time skeleton estimation algorithm of Shotton et al. [16] have generated a renewed interest in skeleton-based human action recognition. Most of the existing skeleton-based approaches use either the joint locations or the joint angles to represent a human skeleton. In this paper, we propose a new skeletal representation that explicitly models the 3D geometric relationships between various body parts using rotations and translations in 3D space. Since 3D rigid body motions are members of the special Euclidean group SE(3), the proposed skeletal representation lies in the Lie group SE(3)×.. .×SE(3), which is a curved manifold. Using the proposed representation, human actions can be modeled as curves in this Lie group. Since classification of curves in this Lie group is not an easy task, we map the action curves from the Lie group to its Lie algebra, which is a vector space. We then perform classification using a combination of dynamic time warping, Fourier temporal pyramid representation and linear SVM. Experimental results on three action datasets show that the proposed representation performs better than many existing skeletal representations. The proposed approach also outperforms various state-of-the-art skeleton-based human action recognition approaches.

In computer vision applications, features often lie on Riemannian manifolds with known geometry. Popular learning algorithms such as discriminant analysis, partial least squares, support vector machines, etc., are not directly applicable to such features due to the non-Euclidean nature of the underlying spaces. Hence, classification is often performed in an extrinsic manner by mapping the manifolds to Euclidean spaces using kernels. However, for kernel based approaches, poor choice of kernel often results in reduced performance. In this paper, we address the issue of kernel selection for the classification of features that lie on Riemannian manifolds using the kernel learning approach. We propose two criteria for jointly learning the kernel and the classifier using a single optimization problem. Specifically, for the SVM classifier, we formulate the problem of learning a good kernel-classifier combination as a convex optimization problem and solve it efficiently following the multiple kernel learning approach. Experimental results on image set-based classification and activity recognition clearly demonstrate the superiority of the proposed approach over existing methods for classification of manifold features.