Unconditional flow-matching trains diffusion models to transport samples from a source distribution to a target distribution by enforcing that the flows between sample pairs are unique. However, in conditional settings (e.g., class-conditioned models), this uniqueness is no longer guaranteed--flows from different conditions may overlap, leading to more ambiguous generations. We introduce Contrastive Flow Matching, an extension to the flow matching objective that explicitly enforces uniqueness across all conditional flows, enhancing condition separation. Our approach adds a contrastive objective that maximizes dissimilarities between predicted flows from arbitrary sample pairs. We validate Contrastive Flow Matching by conducting extensive experiments across varying model architectures on both class-conditioned (ImageNet-1k) and text-to-image (CC3M) benchmarks. Notably, we find that training models with Contrastive Flow Matching (1) improves training speed by a factor of up to 9x, (2) requires up to 5x fewer de-noising steps and (3) lowers FID by up to 8.9 compared to training the same models with flow matching. We release our code at: https://github.com/gstoica27/DeltaFM.git.

Recent model merging methods demonstrate that the parameters of fully-finetuned models specializing in distinct tasks can be combined into one model capable of solving all tasks without retraining. Yet, this success does not transfer well when merging LoRA finetuned models. We study this phenomenon and observe that the weights of LoRA finetuned models showcase a lower degree of alignment compared to their fully-finetuned counterparts. We hypothesize that improving this alignment is key to obtaining better LoRA model merges, and propose KnOTS to address this problem. KnOTS uses the SVD to jointly transform the weights of different LoRA models into an aligned space, where existing merging methods can be applied. In addition, we introduce a new benchmark that explicitly evaluates whether merged models are general models. Notably, KnOTS consistently improves LoRA merging by up to 4.3% across several vision and language benchmarks, including our new setting. We release our code at: https://github.com/gstoica27/KnOTS.

Typical deep visual recognition models are capable of performing the one task they were trained on. In this paper, we tackle the extremely difficult problem of combining distinct models with different initializations, each solving a separate task, into one multi-task model without any additional training. Prior work in model merging permutes one model to the space of the other then averages them together. While this works for models trained on the same task, we find that this fails to account for the differences in models trained on disjoint tasks. Thus, we introduce "ZipIt!", a general method for merging two arbitrary models of the same architecture that incorporates two simple strategies. First, in order to account for features that aren't shared between models, we expand the model merging problem to allow for merging features within each model by defining a general "zip" operation. Second, we add support for partially zipping the models up until a specified layer, naturally creating a multi-head model. We find that these two changes combined account for 20-60% improvement over prior work, making it more feasible to merge models trained on disjoint tasks without retraining.

Embryo quality assessment is vital in assisted reproductive technology, aiding specialists in quickly selecting the most viable embryos for implantation. Leveraging the power of Deep Learning methods enables quick and accurate prediction, reducing the workload of medical specialists and increasing their productivity. Deep Learning algorithms learn recurring patterns and feature representations for a given training dataset, being able to generalize and perform predictions on previously unseen testing data. However, Deep Learning models are prone to be biased in the event of imbalanced training data, which is a common problem for many datasets. We have used batch-level data balancing to ameliorate the effect of data imbalance and employed data augmentation techniques, transfer learning, test time augmentation and model ensembles in order to compare the performance of multiple Deep Learning training configurations on the Embryo Fertility Classification task.

This paper presents a complete pipeline for automatic diagnosis of various liver masses, consisting of components based on deep learning models for both the automatic segmentation of liver and its lesions, as well as the classification of masses into 5 classes, namely simple cysts, hemangiomas, hepatocellular carcinoma, calcifications and metastases. It simultaneously receives 3 Computed Tomography (CT) scans as input for a single patient, representing different acquisition phases, all necessary for an accurate diagnostic. All visible liver masses are automatically segmented into a single class, then extracted and preprocessed as 2D images in order to be sent as input to a classifier that determines the type of lesion. The study is performed on CT scans obtained in collaboration with a regional clinic from 184 symptomatic patients. A thorough experimental analysis comparatively evaluates several deep learning models, leveraging the latest advances in the computer vision field for tackling segmentation and classification tasks. The best models record a mean Dice score of 0.8653 for the segmentation component and an accuracy of 0.9272 for the classification component with an F1 score of 0.9135. The results demonstrate the feasibility of constructing fully automated systems to support the diagnosis of liver disease.

In this paper we introduce Efficient Dynamic Batch Adaptation (EDBA), which improves on a previous method that works by adjusting the composition and the size of the current batch. Our improvements allow for Dynamic Batch Adaptation to feasibly scale up for bigger models and datasets, drastically improving model convergence and generalization. We show how the method is still able to perform especially well in data-scarce scenarios, managing to obtain a test accuracy on 100 samples of CIFAR-10 of 90.68%, while the baseline only reaches 23.79%. On the full CIFAR-10 dataset, EDBA reaches convergence in ∼120 epochs while the baseline requires ∼300 epochs.

In this paper we introduce Backforward Propagation, a method of completely eliminating Internal Covariate Shift (ICS). Unlike previous methods, which only indirectly reduce the impact of ICS while introducing other biases, we are able to have a surgical view at the effects ICS has on training neural networks. Our experiments show that ICS has a weight regularizing effect on models, and completely removing it enables for faster convergence of the neural network.

TACRED is one of the largest and most widely used sentence-level relation extraction datasets. Proposed models that are evaluated using this dataset consistently set new state-of-the-art performance. However, they still exhibit large error rates despite leveraging external knowledge and unsupervised pretraining on large text corpora. A recent study suggested that this may be due to poor dataset quality. The study observed that over 50% of the most challenging sentences from the development and test sets are incorrectly labeled and account for an average drop of 8% f1-score in model performance. However, this study was limited to a small biased sample of 5k (out of a total of 106k) sentences, substantially restricting the generalizability and broader implications of its findings. In this paper, we address these shortcomings by: (i) performing a comprehensive study over the whole TACRED dataset, (ii) proposing an improved crowdsourcing strategy and deploying it to re-annotate the whole dataset, and (iii) performing a thorough analysis to understand how correcting the TACRED annotations affects previously published results. After verification, we observed that 23.9% of TACRED labels are incorrect. Moreover, evaluating several models on our revised dataset yields an average f1-score improvement of 14.3% and helps uncover significant relationships between the different models (rather than simply offsetting or scaling their scores by a constant factor). Finally, aside from our analysis we also release Re-TACRED, a new completely re-annotated version of the TACRED dataset that can be used to perform reliable evaluation of relation extraction models.

We consider the task of knowledge graph link prediction. Given a question consisting of a source entity and a relation (e.g., Shakespeare and BornIn), the objective is to predict the most likely answer entity (e.g., England). Recent approaches tackle this problem by learning entity and relation embeddings. However, they often constrain the relationship between these embeddings to be additive (i.e., the embeddings are concatenated and then processed by a sequence of linear functions and element-wise non-linearities). We show that this type of interaction significantly limits representational power. For example, such models cannot handle cases where a different projection of the source entity is used for each relation. We propose to use contextual parameter generation to address this limitation. More specifically, we treat relations as the context in which source entities are processed to produce predictions, by using relation embeddings to generate the parameters of a model operating over source entity embeddings. This allows models to represent more complex interactions between entities and relations. We apply our method on two existing link prediction methods, including the current state-of-the-art, resulting in significant performance gains and establishing a new state-of-the-art for this task. These gains are achieved while also reducing convergence time by up to 28 times.

Noninvasive molecular and functional imaging in vivo is promising for detecting and monitoring various physiological conditions in animals and ultimately humans. To this end, we present a novel noninvasive technology, spectroscopic photoacoustic tomography (SPAT), which offers both strong optical absorption contrast and high ultrasonic spatial resolution. Optical contrast allows spectroscopic separation of signal contributions from multiple optical absorbers (e.g., oxyhemoglobin, deoxyhemoglobin, and a molecular contrast agent), thus enabling simultaneous molecular and functional imaging. SPAT successfully imaged with high resolution the distribution of a molecular contrast agent targeting integrin overexpressed in human U87 glioblastomas in nude mouse brains. Simultaneously, SPAT also imaged the hemoglobin oxygen saturation and the total hemoglobin concentration of the vasculature, which revealed hypoxia in tumor neovasculature. Therefore, SPAT can potentially lead to better understanding of the interrelationships between hemodynamics and specific biomarkers associated with tumor progression.