Ella Koresh

Ronit D. Gross, Yarden Tzach, Tal Halevi et al.

A prominent achievement of natural language processing (NLP) is its ability to understand and generate meaningful human language. This capability relies on complex feedforward transformer block architectures pre-trained on large language models (LLMs). However, LLM pre-training is currently feasible only for a few dominant companies due to the immense computational resources required, limiting broader research participation. This creates a critical need for more accessible alternatives. In this study, we explore whether tiny language models (TLMs) exhibit the same key qualitative features of LLMs. We demonstrate that TLMs exhibit a clear performance gap between pre-trained and non-pre-trained models across classification tasks, indicating the effectiveness of pre-training, even at a tiny scale. The performance gap increases with the size of the pre-training dataset and with greater overlap between tokens in the pre-training and classification datasets. Furthermore, the classification accuracy achieved by a pre-trained deep TLM architecture can be replicated through a soft committee of multiple, independently pre-trained shallow architectures, enabling low-latency TLMs without affecting classification accuracy. Our results are based on pre-training BERT-6 and variants of BERT-1 on subsets of the Wikipedia dataset and evaluating their performance on FewRel, AGNews, and DBPedia classification tasks. Future research on TLM is expected to further illuminate the mechanisms underlying NLP, especially given that its biologically inspired models suggest that TLMs may be sufficient for children or adolescents to develop language. The data and code that support the findings of this study are openly available on https://github.com/Rg32601/Tiny-Language-Models .

Shalom Rosner, Ronit D. Gross, Ella Koresh et al.

Spontaneous symmetry breaking in statistical mechanics primarily occurs during phase transitions at the thermodynamic limit where the Hamiltonian preserves inversion symmetry, yet the low-temperature free energy exhibits reduced symmetry. Herein, we demonstrate the emergence of spontaneous symmetry breaking in natural language processing (NLP) models during both pre-training and fine-tuning, even under deterministic dynamics and within a finite training architecture. This phenomenon occurs at the level of individual attention heads and is scaled-down to its small subset of nodes and also valid at a single-nodal level, where nodes acquire the capacity to learn a limited set of tokens after pre-training or labels after fine-tuning for a specific classification task. As the number of nodes increases, a crossover in learning ability occurs, governed by the tradeoff between a decrease following random-guess among increased possible outputs, and enhancement following nodal cooperation, which exceeds the sum of individual nodal capabilities. In contrast to spin-glass systems, where a microscopic state of frozen spins cannot be directly linked to the free-energy minimization goal, each nodal function in this framework contributes explicitly to the global network task and can be upper-bounded using convex hull analysis. Results are demonstrated using BERT-6 architecture pre-trained on Wikipedia dataset and fine-tuned on the FewRel classification task.

Ella Koresh, Ronit D. Gross, Yuval Meir et al.

Convolutional neural networks (CNNs) evaluate short-range correlations in input images which progress along the layers, whereas vision transformer (ViT) architectures evaluate long-range correlations, using repeated transformer encoders composed of fully connected layers. Both are designed to solve complex classification tasks but from different perspectives. This study demonstrates that CNNs and ViT architectures stem from a unified underlying learning mechanism, which quantitatively measures the single-nodal performance (SNP) of each node in feedforward (FF) and multi-head attention (MHA) sub-blocks. Each node identifies small clusters of possible output labels, with additional noise represented as labels outside these clusters. These features are progressively sharpened along the transformer encoders, enhancing the signal-to-noise ratio. This unified underlying learning mechanism leads to two main findings. First, it enables an efficient applied nodal diagonal connection (ANDC) pruning technique without affecting the accuracy. Second, based on the SNP, spontaneous symmetry breaking occurs among the MHA heads, such that each head focuses its attention on a subset of labels through cooperation among its SNPs. Consequently, each head becomes an expert in recognizing its designated labels, representing a quantitative MHA modus vivendi mechanism. This statistical mechanics inspired viewpoint enables to reveal macroscopic behavior of the entire network from the microscopic performance of each node. These results are based on a compact convolutional transformer architecture trained on the CIFAR-100 and Flowers-102 datasets and call for their extension to other architectures and applications, such as natural language processing.

Ronit D. Gross, Tal Halevi, Ella Koresh et al.

The emergence of spontaneous symmetry breaking among a few heads of multi-head attention (MHA) across transformer blocks in classification tasks was recently demonstrated through the quantification of single-nodal performance (SNP). This finding indicates that each head focuses its attention on a subset of labels through cooperation among its SNPs. This underlying learning mechanism is generalized to large-scale MHA (LS-MHA) using a single matrix value representing single-head performance (SHP), analogous to single-filter performance in convolutional neural networks (CNNs). The results indicate that each SHP matrix comprises multiple unit clusters such that each label being explicitly recognized by a few heads with negligible noise. This leads to an increased signal-to-noise ratio (SNR) along the transformer blocks, thereby improving classification accuracy. These features give rise to several distinct vision transformer (ViT) architectures that achieve the same accuracy but differ in their LS-MHA structures. As a result, their soft committee yields superior accuracy, an outcome not typically observed in CNNs which rely on hundreds of filters. In addition, a significant reduction in latency is achieved without affecting the accuracy by replacing the initial transformer blocks with convolutional layers. This substitution accelerates early-stage learning, which is then improved by subsequent transformer layers. The extension of this learning mechanism to natural language processing tasks, based on quantitative differences between CNNs and ViT architectures, has the potential to yield new insights in deep learning. The findings are demonstrated using compact convolutional transformer architectures trained on the CIFAR-100 dataset.

Yarden Tzach, Yuval Meir, Ronit D. Gross et al.

Pruning the parameters and structure of neural networks reduces the computational complexity, energy consumption, and latency during inference. Recently, a novel underlying mechanism for successful deep learning (DL) was presented based on a method that quantitatively measures the single filter performance in each layer of a DL architecture, and a new comprehensive mechanism of how deep learning works was presented. This statistical mechanics inspired viewpoint enables to reveal the macroscopic behavior of the entire network from the microscopic performance of each filter and their cooperative behavior. Herein, we demonstrate how this understanding paves the path to high quenched dilution of the convolutional layers of deep architectures without affecting their overall accuracy using applied filter cluster connections (AFCC). AFCC is exemplified on VGG-11 and EfficientNet-B0 architectures trained on CIFAR-100, and its high pruning outperforms other techniques using the same pruning magnitude. Additionally, this technique is broadened to single nodal performance and highly pruning of fully connected layers, suggesting a possible implementation to considerably reduce the complexity of over-parameterized AI tasks.

Yarden Tzach, Ronit D. Gross, Ella Koresh et al.

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.