Large language models (LMs) are capable of in-context learning from a few demonstrations (example-label pairs) to solve new tasks during inference. Despite the intuitive importance of high-quality demonstrations, previous work has observed that, in some settings, ICL performance is minimally affected by irrelevant labels (Min et al., 2022). We hypothesize that LMs perform ICL with irrelevant labels via two sequential processes: an inference function that solves the task, followed by a verbalization function that maps the inferred answer to the label space. Importantly, we hypothesize that the inference function is invariant to remappings of the label space (e.g., "true"/"false" to "cat"/"dog"), enabling LMs to share the same inference function across settings with different label words. We empirically validate this hypothesis with controlled layer-wise interchange intervention experiments. Our findings confirm the hypotheses on multiple datasets and tasks (natural language inference, sentiment analysis, and topic classification) and further suggest that the two functions can be localized in specific layers across various open-sourced models, including GEMMA-7B, MISTRAL-7B-V0.3, GEMMA-2-27B, and LLAMA-3.1-70B.

Gaussian processes (GPs) are Bayesian nonparametric models for function approximation with principled predictive uncertainty estimates. Deep Gaussian processes (DGPs) are multilayer generalizations of GPs that can represent complex marginal densities as well as complex mappings. As exact inference is either computationally prohibitive or analytically intractable in GPs and extensions thereof, some existing methods resort to variational inference (VI) techniques for tractable approximations. However, the expressivity of conventional approximate GP models critically relies on independent inducing variables that might not be informative enough for some problems. In this work we introduce amortized variational inference for DGPs, which learns an inference function that maps each observation to variational parameters. The resulting method enjoys a more expressive prior conditioned on fewer input dependent inducing variables and a flexible amortized marginal posterior that is able to model more complicated functions. We show with theoretical reasoning and experimental results that our method performs similarly or better than previous approaches at less computational cost.

Variational inference is a class of methods to approximate the posterior distribution of a probabilistic model. The classic factorized (or mean-field) variational inference (F-VI) fits a separate parametric distribution for each latent variable. The more modern amortized variational inference (A-VI) instead learns a common \textit{inference function}, which maps each observation to its corresponding latent variable's approximate posterior. Typically, A-VI is used as a cog in the training of variational autoencoders, however it stands to reason that A-VI could also be used as a general alternative to F-VI. In this paper we study when and why A-VI can be used for approximate Bayesian inference. We establish that A-VI cannot achieve a better solution than F-VI, leading to the so-called \textit{amortization gap}, no matter how expressive the inference function is. We then address a central theoretical question: When can A-VI attain F-VI's optimal solution? We derive conditions on the model which are necessary, sufficient, and verifiable under which the amortization gap can be closed. We show that simple hierarchical models, which encompass many models in machine learning and Bayesian statistics, verify these conditions. We demonstrate, on a broader class of models, how to expand the domain of AVI's inference function to improve its solution, and we provide examples, e.g. hidden Markov models, where the amortization gap cannot be closed. Finally, when A-VI can match F-VI's solution, we empirically find that the required complexity of the inference function does not grow with the data size and that A-VI often converges faster.

Reliance on scanned documents and fax communication for healthcare referrals leads to high administrative costs and errors that may affect patient care. In this work we propose a hybrid model leveraging LayoutLMv3 along with domain-specific rules to identify key patient, physician, and exam-related entities in faxed referral documents. We explore some of the challenges in applying a document understanding model to referrals, which have formats varying by medical practice, and evaluate model performance using MUC-5 metrics to obtain appropriate metrics for the practical use case. Our analysis shows the addition of domain-specific rules to the transformer model yields greatly increased precision and F1 scores, suggesting a hybrid model trained on a curated dataset can increase efficiency in referral management.

Learning the complex structure of multivariate time series has been one of the major interests across many application domains, including economics, transportation, manufacturing [Fortuin et al., 2020, Wu et al., 2021, Li et al., 2019, Zhou et al., 2021]. While there has been much progress in the data-driven learning and processing complex time series, it still remains as a challenging topic, in particular, when the data is corrupted [Cao et al., 2018, Kreindler and Lumsden, 2006, Yoon et al., 2018, Du et al., 2022]. In this paper, we consider the forecasting task which aims to make prediction of future values using historical data that may contain missing values. In addition, for many industrial problems, the time series features can be in two categories: auxiliary features (X) that provide information about the state of a system and target variables (Y) that depends on the auxiliary features and may convey valuable information. For example, in the operation of a chemical reactor, the auxiliary features include temperature, pressure and concentration of chemicals observed through a sensor network, while the target variable may include the quality of the material and throughput. We are interested in the time series problem where the data set consists of X and Y. In general, X is more readily available, as it is obtained from a sensor network, while Y may be temporally sparse since it may be expensive or difficult to collect the data. This so-called soft sensor problem has been of interest in many industrial applications [Shardt et al., 2015, Yuan et al., 2021].

Monitoring machine learning (ML) predictions can help improve the quality of deployed models. Capturing the data from inferences made in production can enable you to monitor your deployed models and detect deviations in model quality. Early and proactive detection of these deviations enables you to take corrective actions, such as retraining models, auditing upstream systems, or fixing quality issues. AWS Lambda is a serverless compute service that can provide real-time ML inference at scale. In this post, we demonstrate a sample data capture feature that can be deployed to a Lambda ML inference workload.

Forecasting is an important part of running every business. You need to have an idea about what and how much to produce, especially if lead times are long, in order to have stock available for your customers. If you order too much, you'll have excess inventory which carries cost. If you order too little, you might miss out on lost sales. Products, user features, and accompanying data are quickly changing. This means data drift is a given.

When AI architects think about ML Serving, they focus primarily on speeding up the inference function in the Serving layer. When the solution is deployed, the cost of serving alarms those responsible for budgets, leading to abandoning of solutions. The default architecture that architects come up with is a synchronous one. An ML Service API, typical a REST API sits in front of the serving layer. It takes care of standard API functions like authentication and load balancing.

An important task in the Internet of Things (IoT) is field monitoring, where multiple IoT nodes take measurements and communicate them to the base station or the cloud for processing, inference, and analysis. This communication becomes costly when the measurements are high-dimensional (e.g., videos or time-series data). The IoT networks with limited bandwidth and low power devices may not be able to support such frequent transmissions with high data rates. To ensure communication efficiency, this article proposes to model the measurement compression at IoT nodes and the inference at the base station or cloud as a deep neural network (DNN). We propose a new framework where the data to be transmitted from nodes are the intermediate outputs of a layer of the DNN. We show how to learn the model parameters of the DNN and study the tradeoff between the communication rate and the inference accuracy. The experimental results show that we can save approximately 96% transmissions with only a degradation of 2.5% in inference accuracy. Our findings have the potentiality to enable many new IoT data analysis applications generating large amount of measurements. I. INTRODUCTION Distributed detection, monitoring, and classification are important tasks in wireless sensor networks or the Internet of Things. In these tasks, sensor nodes collect measurements and send them to a base station/gateway, and further to an application server in the IP network that can then make a decision or an inference based on the measurements from the nodes.

We consider efficiency in the implementation of deep neural networks. Hardware accelerators are gaining interest as machine learning becomes one of the drivers of high-performance computing. In these accelerators, the directed graph describing a neural network can be implemented as a directed graph describing a Boolean circuit. We make this observation precise, leading naturally to an understanding of practical neural networks as discrete functions, and show that so-called binarised neural networks are functionally complete. In general, our results suggest that it is valuable to consider Boolean circuits as neural networks, leading to the question of which circuit topologies are promising. We argue that continuity is central to generalisation in learning, explore the interaction between data coding, network topology, and node functionality for continuity, and pose some open questions for future research. As a first step to bridging the gap between continuous and Boolean views of neural network accelerators, we present some recent results from our work on LUTNet, a novel Field-Programmable Gate Array inference approach. Finally, we conclude with additional possible fruitful avenues for research bridging the continuous and discrete views of neural networks.