This paper introduces Neurocache, an approach to extend the effective context size of large language models (LLMs) using an external vector cache to store its past states.Like recent vector retrieval approaches, Neurocache uses an efficient k-nearest-neighbor (kNN) algorithm to retrieve relevant past states and incorporate them into the attention process.Neurocache improves upon previous methods by (1) storing compressed states, which reduces cache size; (2) performing a single retrieval operation per token which increases inference speed; and(3) extending the retrieval window to neighboring states, which improves both language modeling and downstream task accuracy.Our experiments show the effectiveness of Neurocache both for models trained from scratch and for pre-trained models such as Llama2-7B and Mistral-7B when enhanced with the cache mechanism.We also compare Neurocache with text retrieval methods and show improvements in single-document questionanswering and few-shot learning tasks.We made the source code available under: https: //github.com/alisafaya/neurocache

Semih Yagcioglu, Osman Batur İnce, Aykut Erdem, Erkut Erdem, Desmond Elliott, Deniz Yuret. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Drawings are powerful means of pictorial abstraction and communication. Understanding diverse forms of drawings, including digital arts, cartoons, and comics, has been a major problem of interest for the computer vision and computer graphics communities. Although there are large amounts of digitized drawings from comic books and cartoons, they contain vast stylistic variations, which necessitate expensive manual labeling for training domain-specific recognizers. In this work, we show how self-supervised learning, based on a teacher-student network with a modified student network update design, can be used to build face and body detectors. Our setup allows exploiting large amounts of unlabeled data from the target domain when labels are provided for only a small subset of it. We further demonstrate that style transfer can be incorporated into our learning pipeline to bootstrap detectors using a vast amount of out-of-domain labeled images from natural images (i.e., images from the real world). Our combined architecture yields detectors with state-of-the-art (SOTA) and near-SOTA performance using minimal annotation effort. Our code can be accessed from https://github.com/barisbatuhan/DASS_Detector.

Researchers have recently begun exploring the use of StyleGAN-based models for real image editing. One particularly interesting application is using natural language descriptions to guide the editing process. Existing approaches for editing images using language either resort to instance-level latent code optimization or map predefined text prompts to some editing directions in the latent space. However, these approaches have inherent limitations. The former is not very efficient, while the latter often struggles to effectively handle multi-attribute changes. To address these weaknesses, we present CLIPInverter, a new text-driven image editing approach that is able to efficiently and reliably perform multi-attribute changes. The core of our method is the use of novel, lightweight text-conditioned adapter layers integrated into pretrained GAN-inversion networks. We demonstrate that by conditioning the initial inversion step on the Contrastive Language-Image Pre-training (CLIP) embedding of the target description, we are able to obtain more successful edit directions. Additionally, we use a CLIP-guided refinement step to make corrections in the resulting residual latent codes, which further improves the alignment with the text prompt. Our method outperforms competing approaches in terms of manipulation accuracy and photo-realism on various domains including human faces, cats, and birds, as shown by our qualitative and quantitative results.

Self-supervised learning allows AI systems to learn effective representations from large amounts of data using tasks that do not require costly labeling. Mode collapse, i.e., the model producing identical representations for all inputs, is a central problem to many self-supervised learning approaches, making self-supervised tasks, such as matching distorted variants of the inputs, ineffective. In this article, we argue that a straightforward application of information maximization among alternative latent representations of the same input naturally solves the collapse problem and achieves competitive empirical results. We propose a self-supervised learning method, CorInfoMax, that uses a second-order statistics-based mutual information measure that reflects the level of correlation among its arguments. Maximizing this correlative information measure between alternative representations of the same input serves two purposes: (1) it avoids the collapse problem by generating feature vectors with non-degenerate covariances; (2) it establishes relevance among alternative representations by increasing the linear dependence among them. An approximation of the proposed information maximization objective simplifies to a Euclidean distance-based objective function regularized by the log-determinant of the feature covariance matrix. The regularization term acts as a natural barrier against feature space degeneracy. Consequently, beyond avoiding complete output collapse to a single point, the proposed approach also prevents dimensional collapse by encouraging the spread of information across the whole feature space. Numerical experiments demonstrate that CorInfoMax achieves better or competitive performance results relative to the state-of-the-art SSL approaches.

We introduce Morse, a recurrent encoder-decoder model that produces morphological analyses of each word in a sentence. The encoder turns the relevant information about the word and its context into a fixed size vector representation and the decoder generates the sequence of characters for the lemma followed by a sequence of individual morphological features. We show that generating morphological features individually rather than as a combined tag allows the model to handle rare or unseen tags and to outperform whole-tag models. In addition, generating morphological features as a sequence rather than, for example, an unordered set allows our model to produce an arbitrary number of features that represent multiple inflectional groups in morphologically complex languages. We obtain state-of-the-art results in nine languages of different morphological complexity under low-resource, high-resource, and transfer learning settings. We also introduce TrMor2018, a new high-accuracy Turkish morphology data set. Our Morse implementation and the TrMor2018 data set are available online to support future research. 1 See https://github.com/ai-ku/Morse.jl for a Morse implementation in Julia/Knet (Yuret, 2016 ) and https://github.com/ai-ku/TrMor2018 for the new Turkish data set.

Summary Reduction and broadcast operations are commonly used in machine learning algorithms for different purposes. They widely appear in the calculation of the gradient values of a loss function, which are one of the core structures of neural networks. Both operations are implemented naively in many libraries usually for scalar reduction or broadcast; however, to our knowledge, there are no optimized multidimensional implementations available. This fact limits the performance of machine learning models requiring these operations to be performed on tensors. In this work, we address the problem and propose two new strategies that extend the existing implementations to perform on tensors. We introduce formal definitions of both operations using tensor notations, investigate their mathematical properties, and exploit these properties to provide an efficient solution for each. We implement our parallel strategies and test them on a CUDA enabled Tesla K40 m GPU accelerator. Our performant implementations achieve up to 75% of the peak device memory bandwidth on different tensor sizes and dimensions. Significant speedups against the implementations available in the Knet Deep Learning framework are also achieved for both operations.

We describe and evaluate a character-level tagger for language-independent Named Entity Recognition (NER). Instead of words, a sentence is represented as a sequence of characters. The model consists of stacked bidirectional LSTMs which inputs characters and outputs tag probabilities for each character. These probabilities are then converted to consistent word level named entity tags using a Viterbi decoder. We are able to achieve close to state-of-the-art NER performance in seven languages with the same basic model using only labeled NER data and no hand-engineered features or other external resources like syntactic taggers or Gazetteers.

Learning syntactic categories is a fundamental task in language acquisition. Previous studies show that co-occurrence patterns of preceding and following words are essential to group words into categories. However, the neighboring words, or frames, are rarely repeated exactly in the data. This creates data sparsity and hampers learning for frame based models. In this work, we propose a paradigmatic representation of word context which uses probable substitutes instead of frames. Our experiments on child-directed speech show that models based on probable substitutes learn more accurate categories with fewer examples compared to models based on frames.

The encoder-decoder framework for neural machine translation (NMT) has been shown effective in large data scenarios, but is much less effective for low-resource languages.We present a transfer learning method that significantly improves BLEU scores across a range of low-resource languages.Our key idea is to first train a high-resource language pair (the parent model), then transfer some of the learned parameters to the low-resource pair (the child model) to initialize and constrain training.Using our transfer learning method we improve baseline NMT models by an average of 5.6 BLEU on four low-resource language pairs.Ensembling and unknown word replacement add another 2 BLEU which brings the NMT performance on low-resource machine translation close to a strong syntax based machine translation (SBMT) system, exceeding its performance on one language pair.Additionally, using the transfer learning model for re-scoring, we can improve the SBMT system by an average of 1.3 BLEU, improving the state-of-the-art on low-resource machine translation.

We propose a framework for devising empirically testable algorithms for bridging the communication gap between humans and robots.We instantiate our framework in the context of a problem setting in which humans give instructions to robots using unrestricted natural language commands, with instruction sequences being subservient to building complex goal configurations in a blocks world.We show how one can collect meaningful training data and we propose three neural architectures for interpreting contextually grounded natural language commands.The proposed architectures allow us to correctly understand/ground the blocks that the robot should move when instructed by a human who uses unrestricted language.The architectures have more difficulty in correctly understanding/grounding the spatial relations required to place blocks correctly, especially when the blocks are not easily identifiable.

We introduce FDA5 for efficient parameterization, optimization, and implementation of feature decay algorithms (FDA), a class of instance selection algorithms that use feature decay. FDA increase the diversity of the selected training set by devaluing features (i.e., n-grams) that have already been included. FDA5 decides which instances to select based on three functions used for initializing and decaying feature values and scaling sentence scores controlled with five parameters. We present optimization techniques that allow FDA5 to adapt these functions to in-domain and out-of-domain translation tasks for different language pairs. In a transductive learning setting, selection of training instances relevant to the test set can improve the final translation quality. In machine translation experiments performed on the 2 million sentence English-German section of the Europarl corpus, we show that a subset of the training set selected by FDA5 can gain up to 3.22 BLEU points compared to a randomly selected subset of the same size, can gain up to 0.41 BLEU points compared to using all of the available training data using only 15% of it, and can reach within 0.5 BLEU points to the full training set result by using only 2.7% of the full training data. FDA5 peaks at around 8M words or 15% of the full training set. In an active learning setting, FDA5 minimizes the human effort by identifying the most informative sentences for translation and FDA gains up to 0.45 BLEU points using 3/5 of the available training data compared to using all of it and 1.12 BLEU points compared to random training set. In translation tasks involving English and Turkish, a morphologically rich language, FDA5 can gain up to 11.52 BLEU points compared to a randomly selected subset of the same size, can achieve the same BLEU score using as little as 4% of the data compared to random instance selection, and can exceed the full dataset result by 0.78 BLEU points. FDA5 is able to reduce the time to build a statistical machine translation system to about half with 1M words using only 3% of the space for the phrase table and 8% of the overall space when compared with a baseline system using all of the training data available yet still obtain only 0.58 BLEU points difference with the baseline system in out-of-domain translation.

Most traditional distributional similarity models fail to capture syntagmatic patterns that group together multiple word features within the same joint context.In this work we introduce a novel generic distributional similarity scheme under which the power of probabilistic models can be leveraged to effectively model joint contexts.Based on this scheme, we implement a concrete model which utilizes probabilistic n-gram language models.Our evaluations suggest that this model is particularly wellsuited for measuring similarity for verbs, which are known to exhibit richer syntagmatic patterns, while maintaining comparable or better performance with respect to competitive baselines for nouns.Following this, we propose our scheme as a framework for future semantic similarity models leveraging the substantial body of work that exists in probabilistic language modeling.

Lexical substitutes have found use in areas such as paraphrasing, text simplification, machine translation, word sense disambiguation, and part of speech induction. However the computational complexity of accurately identifying the most likely substitutes for a word has made large scale experiments difficult. In this letter we introduce a new search algorithm, FASTSUBS, that is guaranteed to find the K most likely lexical substitutes for a given word in a sentence based on an n-gram language model. The computation is sublinear in both K and the vocabulary size V. An implementation of the algorithm and a dataset with the top 100 substitutes of each token in the WSJ section of the Penn Treebank are available at http://goo.gl/jzKH0.

We introduce a generative probabilistic model, the noisy channel model, for unsupervised word sense disambiguation. In our model, each context C is modeled as a distinct channel through which the speaker intends to transmit a particular meaning S using a possibly ambiguous word W. To reconstruct the intended meaning the hearer uses the distribution of possible meanings in the given context P(S|C) and possible words that can express each meaning P(W|S). We assume P(W|S) is independent of the context and estimate it using WordNet sense frequencies. The main problem of unsupervised WSD is estimating context-dependent P(S|C) without access to any sense-tagged text. We show one way to solve this problem using a statistical language model based on large amounts of untagged text. Our model uses coarse-grained semantic classes for S internally and we explore the effect of using different levels of granularity on WSD performance. The system outputs fine-grained senses for evaluation, and its performance on noun disambiguation is better than most previously reported unsupervised systems and close to the best supervised systems.