Vision Transformers (ViTs) have emerged as powerful backbones in computer vision, outperforming many traditional CNNs. However, their computational overhead, largely attributed to the self-attention mechanism, makes deployment on resource-constrained edge devices challenging. Multiple solutions rely on token pruning or token merging. In this paper, we introduce "Token Fusion" (ToFu), a method that amalgamates the benefits of both token pruning and token merging. Token pruning proves advantageous when the model exhibits sensitivity to input interpolations, while token merging is effective when the model manifests close to linear responses to inputs. We combine this to propose a new scheme called Token Fusion. Moreover, we tackle the limitations of average merging, which doesn't preserve the intrinsic feature norm, resulting in distributional shifts. To mitigate this, we introduce MLERP merging, a variant of the SLERP technique, tailored to merge multiple tokens while maintaining the norm distribution. ToFu is versatile, applicable to ViTs with or without additional training. Our empirical evaluations indicate that ToFu establishes new benchmarks in both classification and image generation tasks concerning computational efficiency and model accuracy.
Recent work has demonstrated a remarkable ability to customize text-to-image diffusion models to multiple, fine-grained concepts in a sequential (i.e., continual) manner while only providing a few example images for each concept. This setting is known as continual diffusion. Here, we ask the question: Can we scale these methods to longer concept sequences without forgetting? Although prior work mitigates the forgetting of previously learned concepts, we show that its capacity to learn new tasks reaches saturation over longer sequences. We address this challenge by introducing a novel method, STack-And-Mask INcremental Adapters (STAMINA), which is composed of low-ranked attention-masked adapters and customized MLP tokens. STAMINA is designed to enhance the robust fine-tuning properties of LoRA for sequential concept learning via learnable hard-attention masks parameterized with low rank MLPs, enabling precise, scalable learning via sparse adaptation. Notably, all introduced trainable parameters can be folded back into the model after training, inducing no additional inference parameter costs. We show that STAMINA outperforms the prior SOTA for the setting of text-to-image continual customization on a 50-concept benchmark composed of landmarks and human faces, with no stored replay data. Additionally, we extended our method to the setting of continual learning for image classification, demonstrating that our gains also translate to state-of-the-art performance in this standard benchmark.
In this paper, we establish a connection between the parameterization of flow-based and energy-based generative models, and present a new flow-based modeling approach called energy-based normalizing flow (EBFlow). We demonstrate that by optimizing EBFlow with score-matching objectives, the computation of Jacobian determinants for linear transformations can be entirely bypassed. This feature enables the use of arbitrary linear layers in the construction of flow-based models without increasing the computational time complexity of each training iteration from $\mathcal{O}(D^2L)$ to $\mathcal{O}(D^3L)$ for an $L$-layered model that accepts $D$-dimensional inputs. This makes the training of EBFlow more efficient than the commonly-adopted maximum likelihood training method. In addition to the reduction in runtime, we enhance the training stability and empirical performance of EBFlow through a number of techniques developed based on our analysis on the score-matching methods. The experimental results demonstrate that our approach achieves a significant speedup compared to maximum likelihood estimation, while outperforming prior efficient training techniques with a noticeable margin in terms of negative log-likelihood (NLL).
Recent works demonstrate a remarkable ability to customize text-to-image diffusion models while only providing a few example images. What happens if you try to customize such models using multiple, fine-grained concepts in a sequential (i.e., continual) manner? In our work, we show that recent state-of-the-art customization of text-to-image models suffer from catastrophic forgetting when new concepts arrive sequentially. Specifically, when adding a new concept, the ability to generate high quality images of past, similar concepts degrade. To circumvent this forgetting, we propose a new method, C-LoRA, composed of a continually self-regularized low-rank adaptation in cross attention layers of the popular Stable Diffusion model. Furthermore, we use customization prompts which do not include the word of the customized object (i.e., "person" for a human face dataset) and are initialized as completely random embeddings. Importantly, our method induces only marginal additional parameter costs and requires no storage of user data for replay. We show that C-LoRA not only outperforms several baselines for our proposed setting of text-to-image continual customization, which we refer to as Continual Diffusion, but that we achieve a new state-of-the-art in the well-established rehearsal-free continual learning setting for image classification. The high achieving performance of C-LoRA in two separate domains positions it as a compelling solution for a wide range of applications, and we believe it has significant potential for practical impact.
Singular value decomposition (SVD) is one of the most popular compression methods that approximate a target matrix with smaller matrices. However, standard SVD treats the parameters within the matrix with equal importance, which is a simple but unrealistic assumption. The parameters of a trained neural network model may affect task performance unevenly, which suggests non-equal importance among the parameters. Compared to SVD, the decomposition method aware of parameter importance is the more practical choice in real cases. Unlike standard SVD, weighted value decomposition is a non-convex optimization problem that lacks a closed-form solution. We systematically investigated multiple optimization strategies to tackle the problem and examined our method by compressing Transformer-based language models. Further, we designed a metric to predict when the SVD may introduce a significant performance drop, for which our method can be a rescue strategy. The extensive evaluations demonstrate that our method can perform better than current SOTA methods in compressing Transformer-based language models.
Factorizing a large matrix into small matrices is a popular strategy for model compression. Singular value decomposition (SVD) plays a vital role in this compression strategy, approximating a learned matrix with fewer parameters. However, SVD minimizes the squared error toward reconstructing the original matrix without gauging the importance of the parameters, potentially giving a larger reconstruction error for those who affect the task accuracy more. In other words, the optimization objective of SVD is not aligned with the trained model's task accuracy. We analyze this previously unexplored problem, make observations, and address it by introducing Fisher information to weigh the importance of parameters affecting the model prediction. This idea leads to our method: Fisher-Weighted SVD (FWSVD). Although the factorized matrices from our approach do not result in smaller reconstruction errors, we find that our resulting task accuracy is much closer to the original model's performance. We perform analysis with the transformer-based language models, showing our weighted SVD largely alleviates the mismatched optimization objectives and can maintain model performance with a higher compression rate. Our method can directly compress a task-specific model while achieving better performance than other compact model strategies requiring expensive model pre-training. Moreover, the evaluation of compressing an already compact model shows our method can further reduce 9% to 30% parameters with an insignificant impact on task accuracy.
Continual learning describes a setting where machine learning models learn novel concepts from continuously shifting training data, while simultaneously avoiding degradation of knowledge on previously seen classes (a phenomenon known as the catastrophic forgetting problem) which may disappear from the training data for extended periods of time. Current approaches for continual learning of a single expanding task (aka class-incremental continual learning) require extensive rehearsal of previously seen data to avoid this degradation of knowledge. Unfortunately, rehearsal comes at a sharp cost to memory and computation, and it may also violate data-privacy. Instead, we explore combining knowledge distillation and parameter regularization in new ways to achieve strong continual learning performance without rehearsal. Specifically, we take a deep dive into common continual learning techniques: prediction distillation, feature distillation, L2 parameter regularization, and EWC parameter regularization. We first disprove the common assumption that parameter regularization techniques fail for rehearsal-free continual learning of a single, expanding task. Next, we explore how to leverage knowledge from a pre-trained model in rehearsal-free continual learning and find that vanilla L2 parameter regularization outperforms EWC parameter regularization and feature distillation. We then highlight the impact of the rehearsal-free continual learning settings with a classifier expansion benchmark, showing that a strategy based on our findings combined with a positive/negative label balancing heuristic can close the performance gap between the upper bound and the existing strategies by up to roughly 50%. Finally, we show that a simple method consisting of pre-training, L2 regularization, and prediction distillation can even outperform rehearsal-based methods on the common CIFAR-100 benchmark.
Knowledge distillation (KD) is a substantial strategy for transferring learned knowledge from one neural network model to another. A vast number of methods have been developed for this strategy. While most method designs a more efficient way to facilitate knowledge transfer, less attention has been put on comparing the effect of knowledge sources such as features, logits, and gradients. This work provides a new perspective to motivate a set of knowledge distillation strategies by approximating the classical KL-divergence criteria with different knowledge sources, making a systematic comparison possible in model compression and incremental learning. Our analysis indicates that logits are generally a more efficient knowledge source and suggests that having sufficient feature dimensions is crucial for the model design, providing a practical guideline for effective KD-based transfer learning.
Domain classification is the fundamental task in natural language understanding (NLU), which often requires fast accommodation to new emerging domains. This constraint makes it impossible to retrain all previous domains, even if they are accessible to the new model. Most existing continual learning approaches suffer from low accuracy and performance fluctuation, especially when the distributions of old and new data are significantly different. In fact, the key real-world problem is not the absence of old data, but the inefficiency to retrain the model with the whole old dataset. Is it potential to utilize some old data to yield high accuracy and maintain stable performance, while at the same time, without introducing extra hyperparameters? In this paper, we proposed a hyperparameter-free continual learning model for text data that can stably produce high performance under various environments. Specifically, we utilize Fisher information to select exemplars that can "record" key information of the original model. Also, a novel scheme called dynamical weight consolidation is proposed to enable hyperparameter-free learning during the retrain process. Extensive experiments demonstrate that baselines suffer from fluctuated performance and therefore useless in practice. On the contrary, our proposed model CCFI significantly and consistently outperforms the best state-of-the-art method by up to 20% in average accuracy, and each component of CCFI contributes effectively to overall performance.
It is well known that vision classification models suffer from poor calibration in the face of data distribution shifts. In this paper, we take a geometric approach to this problem. We propose Geometric Sensitivity Decomposition (GSD) which decomposes the norm of a sample feature embedding and the angular similarity to a target classifier into an instance-dependent and an instance-independent component. The instance-dependent component captures the sensitive information about changes in the input while the instance-independent component represents the insensitive information serving solely to minimize the loss on the training dataset. Inspired by the decomposition, we analytically derive a simple extension to current softmax-linear models, which learns to disentangle the two components during training. On several common vision models, the disentangled model outperforms other calibration methods on standard calibration metrics in the face of out-of-distribution (OOD) data and corruption with significantly less complexity. Specifically, we surpass the current state of the art by 30.8% relative improvement on corrupted CIFAR100 in Expected Calibration Error. Code available at https://github.com/GT-RIPL/Geometric-Sensitivity-Decomposition.git.