Self-attention vector output similarities reveal how machines pay attention

The self-attention mechanism has significantly advanced the field of natural language processing, facilitating the development of advanced language-learning machines. Although its utility is widely acknowledged, the precise mechanisms of self-attention underlying its advanced learning and the quantitative characterization of this learning process remains an open research question. This study introduces a new approach for quantifying information processing within the self-attention mechanism. The analysis conducted on the BERT-12 architecture reveals that, in the final layers, the attention map focuses on sentence separator tokens, suggesting a practical approach to text segmentation based on semantic features. Based on the vector space emerging from the self-attention heads, a context similarity matrix, measuring the scalar product between two token vectors was derived, revealing distinct similarities between different token vector pairs within each head and layer. The findings demonstrated that different attention heads within an attention block focused on different linguistic characteristics, such as identifying token repetitions in a given text or recognizing a token of common appearance in the text and its surrounding context. This specialization is also reflected in the distribution of distances between token vectors with high similarity as the architecture progresses. The initial attention layers exhibit substantially long-range similarities; however, as the layers progress, a more short-range similarity develops, culminating in a preference for attention heads to create strong similarities within the same sentence. Finally, the behavior of individual heads was analyzed by examining the uniqueness of their most common tokens in their high similarity elements. Each head tends to focus on a unique token from the text and builds similarity pairs centered around it.

Translation Entropy: A Statistical Framework for Evaluating Translation Systems

The translation of written language has been known since the 3rd century BC; however, its necessity has become increasingly common in the information age. Today, many translators exist, based on encoder-decoder deep architectures, nevertheless, no quantitative objective methods are available to assess their performance, likely because the entropy of even a single language remains unknown. This study presents a quantitative method for estimating translation entropy, with the following key finding. Given a translator, several sentences that differ by only one selected token of a given pivot sentence yield identical translations. Analyzing the statistics of this phenomenon across an ensemble of such sentences, consisting each of a pivot selected token, yields the probabilities of replacing this specific token with others while preserving the translation. These probabilities constitute the entropy of the selected token, and the average across all selected pivot tokens provides an estimate of the translator's overall translation entropy, which is enhanced along the decoder blocks. This entropic measure allows for the quantitative ranking of several publicly available translators and reveals whether mutual translation entropy is symmetric. Extending the proposed method to include the replacement of two tokens in a given pivot sentence demonstrates a multiplicative effect, where translation degeneracy is proportional to the product of the degeneracies of the two tokens. These findings establish translation entropy as a measurable property and objective benchmarking of artificial translators. Results are based on MarianMT, T5-Base and NLLB-200 translators.

Learning Mechanism Underlying NLP Pre-Training and Fine-Tuning

Natural language processing (NLP) enables the understanding and generation of meaningful human language, typically using a pre-trained complex architecture on a large dataset to learn the language and next fine-tune its weights to implement a specific task. Twofold goals are examined; to understand the mechanism underlying successful pre-training and to determine the interplay between the pre-training accuracy and the fine-tuning of classification tasks. The following main results were obtained; the accuracy per token (APT) increased with its appearance frequency in the dataset, and its average over all tokens served as an order parameter to quantify pre-training success, which increased along the transformer blocks. Pre-training broke the symmetry among tokens and grouped them into finite, small, strong match token clusters, as inferred from the presented token confusion matrix. This feature was sharpened along the transformer blocks toward the output layer, enhancing its performance considerably compared with that of the embedding layer. Consequently, higher-order language structures were generated by pre-training, even though the learning cost function was directed solely at identifying a single token. These pre-training findings were reflected by the improved fine-tuning accuracy along the transformer blocks. Additionally, the output label prediction confidence was found to be independent of the average input APT, as the input meaning was preserved since the tokens are replaced primarily by strong match tokens. Finally, although pre-training is commonly absent in image classification tasks, its underlying mechanism is similar to that used in fine-tuning NLP classification tasks, hinting at its universality. The results were based on the BERT-6 architecture pre-trained on the Wikipedia dataset and fine-tuned on the FewRel and DBpedia classification tasks.

The mechanism underlying successful deep learning

Deep architectures consist of tens or hundreds of convolutional layers (CLs) that terminate with a few fully connected (FC) layers and an output layer representing the possible labels of a complex classification task. According to the existing deep learning (DL) rationale, the first CL reveals localized features from the raw data, whereas the subsequent layers progressively extract higher-level features required for refined classification. This article presents an efficient three-phase procedure for quantifying the mechanism underlying successful DL. First, a deep architecture is trained to maximize the success rate (SR). Next, the weights of the first several CLs are fixed and only the concatenated new FC layer connected to the output is trained, resulting in SRs that progress with the layers. Finally, the trained FC weights are silenced, except for those emerging from a single filter, enabling the quantification of the functionality of this filter using a correlation matrix between input labels and averaged output fields, hence a well-defined set of quantifiable features is obtained. Each filter essentially selects a single output label independent of the input label, which seems to prevent high SRs; however, it counterintuitively identifies a small subset of possible output labels. This feature is an essential part of the underlying DL mechanism and is progressively sharpened with layers, resulting in enhanced signal-to-noise ratios and SRs. Quantitatively, this mechanism is exemplified by the VGG-16, VGG-6, and AVGG-16. The proposed mechanism underlying DL provides an accurate tool for identifying each filter's quality and is expected to direct additional procedures to improve the SR, computational complexity, and latency of DL.

Enhancing the success rates by performing pooling decisions adjacent to the output layer

Learning classification tasks of (2^nx2^n) inputs typically consist of \le n (2x2) max-pooling (MP) operators along the entire feedforward deep architecture. Here we show, using the CIFAR-10 database, that pooling decisions adjacent to the last convolutional layer significantly enhance accuracy success rates (SRs). In particular, average SRs of the advanced VGG with m layers (A-VGGm) architectures are 0.936, 0.940, 0.954, 0.955, and 0.955 for m=6, 8, 14, 13, and 16, respectively. The results indicate A-VGG8s' SR is superior to VGG16s', and that the SRs of A-VGG13 and A-VGG16 are equal, and comparable to that of Wide-ResNet16. In addition, replacing the three fully connected (FC) layers with one FC layer, A-VGG6 and A-VGG14, or with several linear activation FC layers, yielded similar SRs. These significantly enhanced SRs stem from training the most influential input-output routes, in comparison to the inferior routes selected following multiple MP decisions along the deep architecture. In addition, SRs are sensitive to the order of the non-commutative MP and average pooling operators adjacent to the output layer, varying the number and location of training routes. The results call for the reexamination of previously proposed deep architectures and their SRs by utilizing the proposed pooling strategy adjacent to the output layer.

Efficient shallow learning as an alternative to deep learning

The realization of complex classification tasks requires training of deep learning (DL) architectures consisting of tens or even hundreds of convolutional and fully connected hidden layers, which is far from the reality of the human brain. According to the DL rationale, the first convolutional layer reveals localized patterns in the input and large-scale patterns in the following layers, until it reliably characterizes a class of inputs. Here, we demonstrate that with a fixed ratio between the depths of the first and second convolutional layers, the error rates of the generalized shallow LeNet architecture, consisting of only five layers, decay as a power law with the number of filters in the first convolutional layer. The extrapolation of this power law indicates that the generalized LeNet can achieve small error rates that were previously obtained for the CIFAR-10 database using DL architectures. A power law with a similar exponent also characterizes the generalized VGG-16 architecture. However, this results in a significantly increased number of operations required to achieve a given error rate with respect to LeNet. This power law phenomenon governs various generalized LeNet and VGG-16 architectures, hinting at its universal behavior and suggesting a quantitative hierarchical time-space complexity among machine learning architectures. Additionally, the conservation law along the convolutional layers, which is the square-root of their size times their depth, is found to asymptotically minimize error rates. The efficient shallow learning that is demonstrated in this study calls for further quantitative examination using various databases and architectures and its accelerated implementation using future dedicated hardware developments.