This paper proposes improvements over earlier work by Nazareth and Blei (2022) for estimating the depth of Bayesian neural networks. Here, we propose a discrete truncated normal distribution over the network depth to independently learn its mean and variance. Posterior distributions are inferred by minimizing the variational free energy, which balances the model complexity and accuracy. Our method improves test accuracy on the spiral data set and reduces the variance in posterior depth estimates.

Due to high accuracy, BERT-like models have been widely adopted by discriminative text mining and web searching. However, large BERT-like models suffer from inefficient online inference, as they face the following two problems on GPUs. First, they rely on the large model depth to achieve high accuracy, which linearly increases the sequential computation on GPUs. Second, stochastic and dynamic online workloads cause extra costs. In this paper, we present Academus for low-latency online inference of BERT-like models. At the core of Academus is the novel student parallelism, which adopts boosting ensemble and stacking distillation to distill the original deep model into an equivalent group of parallel and shallow student models. This enables Academus to achieve the lower model depth (e.g., two layers) than baselines and consequently the lowest inference latency without affecting the accuracy.For occasional workload bursts, it can temporarily decrease the number of students with minimal accuracy loss to improve throughput. Additionally, it employs specialized system designs for student parallelism to better handle stochastic online workloads. We conduct comprehensive experiments to verify the effectiveness. The results show that Academus outperforms the baselines by 4.1X~1.6X in latency without compromising accuracy, and achieves up to 22.27X higher throughput for workload bursts.

High-quality sentence embeddings are fundamental in many natural language processing (NLP) tasks, such as semantic textual similarity (STS) and retrieval-augmented generation (RAG). Nevertheless, most existing methods leverage fixed-length embeddings from full-layer language models, which lack the scalability to accommodate the diverse available resources across various applications. Viewing this gap, we propose a novel sentence embedding model $\mathrm{Espresso}$ $\mathrm{Sentence}$ $\mathrm{Embeddings}$ (ESE) with two learning processes. First, the learn-to-express process encodes more salient representations to lower layers. Second, the learn-to-compress process compacts essential features into the initial dimensions using Principal Component Analysis (PCA). This way, ESE can scale model depth via the former process and embedding size via the latter. Extensive experiments on STS and RAG suggest that ESE can effectively produce high-quality embeddings with less model depth and embedding size, enhancing embedding inference efficiency.

Humans have a tendency to see 'human'-like qualities in objects around them. We name our cars, and talk to pets and even household appliances, as if they could understand us as other humans do. This behavior, called anthropomorphism, is also seeing traction in Machine Learning (ML), where human-like intelligence is claimed to be perceived in Large Language Models (LLMs). In this position paper, considering professional incentives, human biases, and general methodological setups, we discuss how the current search for Artificial General Intelligence (AGI) is a perfect storm for over-attributing human-like qualities to LLMs. In several experiments, we demonstrate that the discovery of human-interpretable patterns in latent spaces should not be a surprising outcome. Also in consideration of common AI portrayal in the media, we call for the academic community to exercise extra caution, and to be extra aware of principles of academic integrity, in interpreting and communicating about AI research outcomes.

Deep graph models (e.g., graph neural networks and graph transformers) have become important techniques for leveraging knowledge across various types of graphs. Yet, the scaling properties of deep graph models have not been systematically investigated, casting doubt on the feasibility of achieving large graph models through enlarging the model and dataset sizes. In this work, we delve into neural scaling laws on graphs from both model and data perspectives. We first verify the validity of such laws on graphs, establishing formulations to describe the scaling behaviors. For model scaling, we investigate the phenomenon of scaling law collapse and identify overfitting as the potential reason. Moreover, we reveal that the model depth of deep graph models can impact the model scaling behaviors, which differ from observations in other domains such as CV and NLP. For data scaling, we suggest that the number of graphs can not effectively metric the graph data volume in scaling law since the sizes of different graphs are highly irregular. Instead, we reform the data scaling law with the number of edges as the metric to address the irregular graph sizes. We further demonstrate the reformed law offers a unified view of the data scaling behaviors for various fundamental graph tasks including node classification, link prediction, and graph classification. This work provides valuable insights into neural scaling laws on graphs, which can serve as an essential step toward large graph models.

The success of modern neural networks has prompted study of the connection between memorisation and generalisation: overparameterised models generalise well, despite being able to perfectly fit (memorise) completely random labels. To carefully study this issue, Feldman proposed a metric to quantify the degree of memorisation of individual training examples, and empirically computed the corresponding memorisation profile of a ResNet on image classification bench-marks. While an exciting first glimpse into what real-world models memorise, this leaves open a fundamental question: do larger neural models memorise more? We present a comprehensive empirical analysis of this question on image classification benchmarks. We find that training examples exhibit an unexpectedly diverse set of memorisation trajectories across model sizes: most samples experience decreased memorisation under larger models, while the rest exhibit cap-shaped or increasing memorisation. We show that various proxies for the Feldman memorization score fail to capture these fundamental trends. Lastly, we find that knowledge distillation, an effective and popular model compression technique, tends to inhibit memorisation, while also improving generalisation. Specifically, memorisation is mostly inhibited on examples with increasing memorisation trajectories, thus pointing at how distillation improves generalisation.

For humans, language production and comprehension is sensitive to the hierarchical structure of sentences. In natural language processing, past work has questioned how effectively neural sequence models like transformers capture this hierarchical structure when generalizing to structurally novel inputs. We show that transformer language models can learn to generalize hierarchically after training for extremely long periods -- far beyond the point when in-domain accuracy has saturated. We call this phenomenon \emph{structural grokking}. On multiple datasets, structural grokking exhibits inverted U-shaped scaling in model depth: intermediate-depth models generalize better than both very deep and very shallow transformers. When analyzing the relationship between model-internal properties and grokking, we find that optimal depth for grokking can be identified using the tree-structuredness metric of \citet{murty2023projections}. Overall, our work provides strong evidence that, with extended training, vanilla transformers discover and use hierarchical structure.

For years the model performance in machine learning obeyed a power-law relationship with the model size. For the consideration of parameter efficiency, recent studies focus on increasing model depth rather than width to achieve better performance. In this paper, we study how model width affects the Transformer model through a parameter-efficient multi-path structure. To better fuse features extracted from different paths, we add three additional operations to each sublayer: a normalization at the end of each path, a cheap operation to produce more features, and a learnable weighted mechanism to fuse all features flexibly. Extensive experiments on 12 WMT machine translation tasks show that, with the same number of parameters, the shallower multi-path model can achieve similar or even better performance than the deeper model. It reveals that we should pay more attention to the multi-path structure, and there should be a balance between the model depth and width to train a better large-scale Transformer.

In this paper, we propose a highly parameter-efficient approach to scaling pre-trained language models (PLMs) to a deeper model depth. Unlike prior work that shares all parameters or uses extra blocks, we design a more capable parameter-sharing architecture based on matrix product operator (MPO). MPO decomposition can reorganize and factorize the information of a parameter matrix into two parts: the major part that contains the major information (central tensor) and the supplementary part that only has a small proportion of parameters (auxiliary tensors). Based on such a decomposition, our architecture shares the central tensor across all layers for reducing the model size and meanwhile keeps layer-specific auxiliary tensors (also using adapters) for enhancing the adaptation flexibility. To improve the model training, we further propose a stable initialization algorithm tailored for the MPO-based architecture. Extensive experiments have demonstrated the effectiveness of our proposed model in reducing the model size and achieving highly competitive performance.

The trend in industrial automation is towards networking, intelligence and autonomy. Digital Twins, which serve as virtual representations, are becoming increasingly important in this context. The Digital Twin of a modular production system contains many different models that are mostly created for specific applications and fulfil different requirements. Especially simulation models, which are created in the development phase, can be used during the operational phase for applications such as prognosis or operation-parallel simulation. Due to the high heterogeneity of the model landscape in the context of a modular production system, the plant operator is faced with the challenge of adapting the models in order to ensure an application-oriented realism in the event of changes to the asset and its environment or the addition of applications. Therefore, this paper proposes a concept for the continuous model adaption in the Digital Twin of a modular production system during the operational phase. The benefits are then demonstrated by an application scenario and an agent-based realisation.