With the proliferation of large pre-trained language models (PLMs), fine-tuning all model parameters becomes increasingly inefficient, particularly when dealing with numerous downstream tasks that entail substantial training and storage costs. Several approaches aimed at achieving parameter-efficient fine-tuning (PEFT) have been proposed. Among them, Low-Rank Adaptation (LoRA) stands out as an archetypal method, incorporating trainable rank decomposition matrices into each target module. Nevertheless, LoRA does not consider the varying importance of each layer. To address these challenges, we introduce PRILoRA, which linearly allocates a different rank for each layer, in an increasing manner, and performs pruning throughout the training process, considering both the temporary magnitude of weights and the accumulated statistics of the input to any given layer. We validate the effectiveness of PRILoRA through extensive experiments on eight GLUE benchmarks, setting a new state of the art.
We present a new layer in which dynamic (i.e., input-dependent) Infinite Impulse Response (IIR) filters of order two are used to process the input sequence prior to applying conventional attention. The input is split into chunks, and the coefficients of these filters are determined based on previous chunks to maintain causality. Despite their relatively low order, the causal adaptive filters are shown to focus attention on the relevant sequence elements. The new layer is grounded in control theory, and is shown to generalize diagonal state-space layers. The layer performs on-par with state-of-the-art networks, with a fraction of their parameters and with time complexity that is sub-quadratic with input size. The obtained layer is favorable to layers such as Heyna, GPT2, and Mega, both with respect to the number of parameters and the obtained level of performance on multiple long-range sequence problems.
We seek to semantically describe a set of images, capturing both the attributes of single images and the variations within the set. Our procedure is analogous to Principle Component Analysis, in which the role of projection vectors is replaced with generated phrases. First, a centroid phrase that has the largest average semantic similarity to the images in the set is generated, where both the computation of the similarity and the generation are based on pretrained vision-language models. Then, the phrase that generates the highest variation among the similarity scores is generated, using the same models. The next phrase maximizes the variance subject to being orthogonal, in the latent space, to the highest-variance phrase, and the process continues. Our experiments show that our method is able to convincingly capture the essence of image sets and describe the individual elements in a semantically meaningful way within the context of the entire set. Our code is available at: https://github.com/OdedH/textual-pca.
Language modeling with BERT consists of two phases of (i) unsupervised pre-training on unlabeled text, and (ii) fine-tuning for a specific supervised task. We present a method that leverages the second phase to its fullest, by applying an extensive number of parallel classifier heads, which are enforced to be orthogonal, while adaptively eliminating the weaker heads during training. We conduct an extensive inter- and intra-dataset evaluation, showing that our method improves the generalization ability of BERT, sometimes leading to a +9% gain in accuracy. These results highlight the importance of a proper fine-tuning procedure, especially for relatively smaller-sized datasets. Our code is attached as supplementary.
The recently introduced hateful meme challenge demonstrates the difficulty of determining whether a meme is hateful or not. Specifically, both unimodal language models and multimodal vision-language models cannot reach the human level of performance. Motivated by the need to model the contrast between the image content and the overlayed text, we suggest applying an off-the-shelf image captioning tool in order to capture the first. We demonstrate that the incorporation of such automatic captions during fine-tuning improves the results for various unimodal and multimodal models. Moreover, in the unimodal case, continuing the pre-training of language models on augmented and original caption pairs, is highly beneficial to the classification accuracy.
When training neural models, it is common to combine multiple loss terms. The balancing of these terms requires considerable human effort and is computationally demanding. Moreover, the optimal trade-off between the loss terms can change as training progresses, e.g., for adversarial terms. In this work, we generalize the Adam optimization algorithm to handle multiple loss terms. The guiding principle is that for every layer, the gradient magnitude of the terms should be balanced. To this end, the Multi-Term Adam (MTAdam) computes the derivative of each loss term separately, infers the first and second moments per parameter and loss term, and calculates a first moment for the magnitude per layer of the gradients arising from each loss. This magnitude is used to continuously balance the gradients across all layers, in a manner that both varies from one layer to the next and dynamically changes over time. Our results show that training with the new method leads to fast recovery from suboptimal initial loss weighting and to training outcomes that match or improve conventional training with the prescribed hyperparameters of each method.
Unsupervised word translation from non-parallel inter-lingual corpora has attracted much research interest. Very recently, neural network methods trained with adversarial loss functions achieved high accuracy on this task. Despite the impressive success of the recent techniques, they suffer from the typical drawbacks of generative adversarial models: sensitivity to hyper-parameters, long training time and lack of interpretability. In this paper, we make the observation that two sufficiently similar distributions can be aligned correctly with iterative matching methods. We present a novel method that first aligns the second moment of the word distributions of the two languages and then iteratively refines the alignment. Extensive experiments on word translation of European and Non-European languages show that our method achieves better performance than recent state-of-the-art deep adversarial approaches and is competitive with the supervised baseline. It is also efficient, easy to parallelize on CPU and interpretable.
We study the topmost weight matrix of neural network language models. We show that this matrix constitutes a valid word embedding. When training language models, we recommend tying the input embedding and this output embedding. We analyze the resulting update rules and show that the tied embedding evolves in a more similar way to the output embedding than to the input embedding in the untied model. We also offer a new method of regularizing the output embedding. Our methods lead to a significant reduction in perplexity, as we are able to show on a variety of neural network language models. Finally, we show that weight tying can reduce the size of neural translation models to less than half of their original size without harming their performance.