We introduce funion, a system providing end-to-end sender-receiver unlinkability for neural network inference. By leveraging the Pigeonhole storage protocol and BACAP (blinding-and-capability) scheme from the Echomix anonymity system, funion inherits the provable security guarantees of modern mixnets. Users can anonymously store input tensors in pseudorandom storage locations, commission compute services to process them via the neural network, and retrieve results with no traceable connection between input and output parties. This store-compute-store paradigm masks both network traffic patterns and computational workload characteristics, while quantizing execution timing into public latency buckets. Our security analysis demonstrates that funion inherits the strong metadata privacy guarantees of Echomix under largely the same trust assumptions, while introducing acceptable overhead for production-scale workloads. Our work paves the way towards an accessible platform where users can submit fully anonymized inference queries to cloud services.

Mixed precision accumulation for neural network inference guided by componentwise forward error analysis El-Mehdi El arar 1, Silviu-Ioan Filip 1, Theo Mary 2, and Elisa Riccietti 3 1 Inria, IRISA, Universit e de Rennes, 263 Av. G en eral Leclerc, F-35000, Rennes, France 2 Sorbonne Universit e, CNRS, LIP6, 4 Place Jussieu, F-75005, Paris, France 3 ENS de Lyon, CNRS, Inria, Universit e Claude Bernard Lyon 1 LIP, UMR 5668, 69342, Lyon cedex 07, France Abstract This work proposes a mathematically founded mixed precision accumulation strategy for the inference of neural networks. Our strategy is based on a new componentwise forward error analysis that explains the propagation of errors in the forward pass of neural networks. Specifically, our analysis shows that the error in each component of the output of a layer is proportional to the condition number of the inner product between the weights and the input, multiplied by the condition number of the activation function. These condition numbers can vary widely from one component to the other, thus creating a significant opportunity to introduce mixed precision: each component should be accumulated in a precision inversely proportional to the product of these condition numbers. We propose a practical algorithm that exploits this observation: it first computes all components in low precision, uses this output to estimate the condition numbers, and recomputes in higher precision only the components associated with large condition numbers. We test our algorithm on various networks and datasets and confirm experimentally that it can significantly improve the cost-accuracy tradeoff compared with uniform precision accumulation baselines. Keywords: Neural network, inference, error analysis, mixed precision, multiply-accumulate 1 Introduction Modern applications in artificial intelligence require increasingly complex models and thus increasing memory, time, and energy costs for storing and deploying large-scale deep learning models with parameter counts ranging in the millions and billions. This is a limiting factor both in the context of training and of inference. While the growing training costs can be tackled by the power of modern computing resources, notably GPU accelerators, the deployment of large-scale models leads to serious limitations in inference contexts with limited resources, such as embedded systems or applications that require real-time processing.

Multiparty computation approaches to secure neural network inference commonly rely on garbled circuits for securely executing nonlinear activation functions. However, garbled circuits require excessive communication between server and client, impose significant storage overheads, and incur large runtime penalties. To reduce these costs, we propose an alternative to garbled circuits: Tabula, an algorithm based on secure lookup tables. Our approach precomputes lookup tables during an offline phase that contains the result of all possible nonlinear function calls. Because these tables incur exponential storage costs in the number of operands and the precision of the input values, we use quantization to reduce these storage costs to make this approach practical. This enables an online phase where securely computing the result of a nonlinear function requires just a single round of communication, with communication cost equal to twice the number of bits of the input to the nonlinear function. In practice our approach costs 2 bytes of communication per nonlinear function call in the online phase. Compared to garbled circuits with 8-bit quantized inputs, when computing individual nonlinear functions during the online phase, experiments show Tabula with 8-bit activations uses between $280$-$560 \times$ less communication, is over $100\times$ faster, and uses a comparable (within a factor of 2) amount of storage; compared against other state-of-the-art protocols Tabula achieves greater than $40\times$ communication reduction. This leads to significant performance gains over garbled circuits with quantized inputs during the online phase of secure inference of neural networks: Tabula reduces end-to-end inference communication by up to $9 \times$ and achieves an end-to-end inference speedup of up to $50 \times$, while imposing comparable storage and offline preprocessing costs.

Abstract--With more scientific fields relying on neural networks (NNs) to process data incoming at extreme throughputs and latencies, it is crucial to develop NNs with all their parameters stored on-chip. In many of these applications, there is not enough time to go off-chip and retrieve weights. Even more so, off-chip memory such as DRAM does not have the bandwidth required to process these NNs as fast as the data is being produced (e.g., every 25 ns). As such, these extreme latency and bandwidth requirements have architectural implications for the hardware intended to run these NNs: 1) all NN parameters must fit on-chip, and 2) codesigning custom/reconfigurable logic is often required to meet these latency and bandwidth constraints. In our work, we show that many scientific NN applications must run fully on chip, in the extreme case requiring a custom chip to meet such stringent constraints.

This position paper presents a theoretical framework for enhancing explainable artificial intelligence (xAI) through emergent communication (EmCom), focusing on creating a causal understanding of AI model outputs. We explore the novel integration of EmCom into AI systems, offering a paradigm shift from conventional associative relationships between inputs and outputs to a more nuanced, causal interpretation. The framework aims to revolutionize how AI processes are understood, making them more transparent and interpretable. While the initial application of this model is demonstrated on synthetic data, the implications of this research extend beyond these simple applications. This general approach has the potential to redefine interactions with AI across multiple domains, fostering trust and informed decision-making in healthcare and in various sectors where AI's decision-making processes are critical.

The rise of neural network-based machine learning ushered in high-level libraries, including TensorFlow and PyTorch, to support their functionality. Computational fluid dynamics (CFD) researchers have benefited from this trend and produced powerful neural networks that promise shorter simulation times. For example, multilayer perceptrons (MLPs) and Long Short Term Memory (LSTM) recurrent-based (RNN) architectures can represent sub-grid physical effects, like turbulence. Implementing neural networks in CFD solvers is challenging because the programming languages used for machine learning and CFD are mostly non-overlapping, We present the roseNNa library, which bridges the gap between neural network inference and CFD. RoseNNa is a non-invasive, lightweight (1000 lines), and performant tool for neural network inference, with focus on the smaller networks used to augment PDE solvers, like those of CFD, which are typically written in C/C++ or Fortran. RoseNNa accomplishes this by automatically converting trained models from typical neural network training packages into a high-performance Fortran library with C and Fortran APIs. This reduces the effort needed to access trained neural networks and maintains performance in the PDE solvers that CFD researchers build and rely upon. Results show that RoseNNa reliably outperforms PyTorch (Python) and libtorch (C++) on MLPs and LSTM RNNs with less than 100 hidden layers and 100 neurons per layer, even after removing the overhead cost of API calls. Speedups range from a factor of about 10 and 2 faster than these established libraries for the smaller and larger ends of the neural network size ranges tested.

MLaaS (ML-as-a-Service) offerings by cloud computing platforms are becoming increasingly popular these days. Pre-trained machine learning models are deployed on the cloud to support prediction based applications and services. For achieving higher throughput, incoming requests are served by running multiple replicas of the model on different machines concurrently. Incidence of straggler nodes in distributed inference is a significant concern since it can increase inference latency, violate SLOs of the service. In this paper, we propose a novel coded inference model to deal with stragglers in distributed image classification. We propose modified single shot object detection models, Collage-CNN models, to provide necessary resilience efficiently. A Collage-CNN model takes collage images formed by combining multiple images as its input and performs multi-image classification in one shot. We generate custom training collages using images from standard image classification datasets and train the model to achieve high classification accuracy. Deploying the Collage-CNN models in the cloud, we demonstrate that the 99th percentile latency can be reduced by 1.45X to 2.46X compared to replication based approaches and without compromising prediction accuracy.

NVIDIA TensorRT is a high-performance deep learning inference library for production environments. Power efficiency and speed of response are two key metrics for deployed deep learning applications, because they directly affect the user experience and the cost of the service provided. Tensor RT automatically optimizes trained neural networks for run-time performance, delivering up to 16x higher energy efficiency (performance per watt) on a Tesla P100 GPU compared to common CPU-only deep learning inference systems (see Figure 1). Figure 2 shows the performance of NVIDIA Tesla P100 and K80 running inference using TensorRT with the relatively complex GoogLenet neural network architecture. In this post we will show you how you can use Tensor RT to get the best efficiency and performance out of your trained deep neural network on a GPU-based deployment platform.

Over the last few months we have seen NVIDIA's Pascal GPUs roll out among their consumer cards, and now the time has come for the Tesla line to get its own Pascal update. To that end, at today's GTC Beijing 2016 keynote, NVIDIA CEO Jen-Hsun Huang has announced the next generation of NVIDIA's neural network inferencing cards, the Tesla P40 and Tesla P4. These cards are the direct successor to the current Tesla M40 and M4 products, and with the addition of the Pascal architecture, NVIDIA is promising a major leap in inferencing performance. We've covered NVIDIA's presence in and plans for the deep learning market for some time now. Overall the deep learning market is a rapidly growing market, and one that has proven very successful for NVIDIA as the underlying neural networks map well to their GPU architectures.

Today at ICML 2016, NVIDIA announced its latest Deep Learning SDK updates, including DIGITS 4, cuDNN 5.1 (CUDA Deep Neural Network Library) and the new GPU Inference Engine. NVIDIA GPU Inference Engine (GIE) is a high-performance deep learning inference solution for production environments. Power efficiency and speed of response are two key metrics for deployed deep learning applications, because they directly affect the user experience and the cost of the service provided. GIE automatically optimizes trained neural networks for run-time performance, delivering up to 16x higher performance per watt on a Tesla M4 GPU compared to the CPU-only systems commonly used for inference today. Figure 1 shows GIE inference performance per watt of the relatively complex GoogLeNet running on a Tesla M4. GIE can deliver 20 Images/s/Watt on the simpler AlexNet benchmark.