As Large Language Models (LLMs) scale, the question is not just how large they become, but how much of their capacity is effectively utilized.Existing scaling laws relate model size to loss, yet overlook how components exploit their latent space.In this work, we focus on Feed-Forward Networks (FFNs) and recast width selection as a spectral utilization optimization problem.Using a lightweight diagnostic suite: Hard Rank (participation ratio), Soft Rank (Shannon Rank), Spectral Concentration, and the composite Spectral Utilization Index (SUI), we quantify how many latent directions are meaningfully activated across LLaMA, GPT-2, and nGPT families.Our key finding is an Asymmetric Spectral Scaling Law: soft rank follows an almost perfect power law with FFN width, while hard rank grows only sublinearly, with high variance.This asymmetry suggests that widening FFNs mostly adds low-energy tail directions, while dominant-mode subspaces saturate early.Moreover, at larger widths, variance further collapses into a narrow subspace, leaving much of the latent space under-utilized.These results recast FFN width selection as a principled trade-off between tail capacity and dominant-mode capacity, offering concrete guidance for inference-efficient LLM design.

Large language model (LLM) based services are primarily structured as client-server interactions, with clients sending queries directly to cloud providers that host LLMs. This approach currently compromises data privacy as all queries must be processed in the cloud and in the clear. Fully Homomorphic Encryption (FHE) is a solution to this data privacy issue by enabling computations directly upon encrypted queries. However, running encrypted transformer inference is challenging as programmers must map standard kernels to the constrained instruction set provided by FHE. In this work, we explore implementations of linear algebra kernels needed for transformer inference in FHE and understand how network optimization can help mitigate FHE costs while remaining performant.We leverage the Orion PyTorch to FHE framework to benchmark several linear algebra kernels in order to profile two linear transformation methods, packed row and BSGS, and find that BSGS outperforms packed row methods by up to 13.7× at transformer-level scales. We also incorporate network-level pruning strategies that reduce FHE runtimes of feed forward layers by up to 11.46×. Furthermore, we extend Orion to include ciphertext-ciphertext matrix-matrix products, a key component in the self-attention blocks. And finally, we perform a roofline analysis of FHE primitives and encrypted linear transformations and find that (SIMD encoded) implementations are memory-bound with primitives having roughly 0.1 integer operations per byte of DRAM traffic. These findings illustrate the need for exploring alternative encoding schemes and models of computation within CKKS to unlock scalable private transformer inference. We conduct all experiments using the Orion framework which can be found at: https://github.com/baahl-nyu/orion.

In two-party machine learning prediction services, the client’s goal is to query a remote server’s trained machine learning model to perform neural network inference in some application domain. However, sensitive information can be obtained during this process by either the client or the server, leading to potential collection, unauthorized secondary use, and inappropriate access to personal information. These security concerns have given rise to Private Inference (PI), in which both the client’s personal data and the server’s trained model are kept confidential. State-of-the-art PI protocols consist of a pre-processing or offline phase and an online phase that combine several cryptographic primitives: Homomorphic Encryption (HE), Secret Sharing (SS), Garbled Circuits (GC), and Oblivious Transfer (OT). Despite the need and recent performance improvements, PI remains largely arcane today and is too slow for practical use.

The recent rise of privacy concerns has led researchers to devise methods for private neural inference—where inferences are made directly on encrypted data, never seeing inputs. The primary challenge facing private inference is that computing on encrypted data levies an impractically-high latency penalty, stemming mostly from non-linear operators like ReLU. Enabling practical and private inference requires new optimization methods that minimize network ReLU counts while preserving accuracy. This paper proposes DeepReDuce: a set of optimizations for the judicious removal of ReLUs to reduce private inference latency. The key insight is that not all ReLUs contribute equally to accuracy. We leverage this insight to drop, or remove, ReLUs from classic networks to significantly reduce inference latency and maintain high accuracy. Given a network architecture, DeepReDuce outputs a Pareto frontier of networks that tradeoff the number of ReLUs and accuracy. Compared to the state-of-the-art for private inference DeepReDuce improves accuracy and reduces ReLU count by up to 3.5% (iso-ReLU count) and 3.5x (iso-accuracy), respectively.

The simultaneous rise of machine learning as a service and concerns over user privacy have increasingly motivated the need for private inference (PI). While recent work demonstrates PI is possible using cryptographic primitives, the computational overheads render it impractical. State-of-art deep networks are inadequate in this context because the source of slowdown in PI stems from the ReLU operations whereas optimizations for plaintext inference focus on reducing FLOPs. In this paper we re-think ReLU computations and propose optimizations for PI tailored to properties of neural networks. Specifically, we reformulate ReLU as an approximate sign test and introduce a novel truncation method for the sign test that significantly reduces the cost per ReLU. These optimizations result in a specific type of stochastic ReLU. The key observation is that the stochastic fault behavior is well suited for the fault-tolerant properties of neural network inference. Thus, we provide significant savings without impacting accuracy. We collectively call the optimizations Circa and demonstrate improvements of up to 4.7$\times$ storage and 3$\times$ runtime over baseline implementations; we further show that Circa can be used on top of recent PI optimizations to obtain 1.8$\times$ additional speedup.

In recent years, researchers have focused on reducing the model size and number of computations (measured as “multiply-accumulate” or MAC operations) of DNNs. The energy consumption of a DNN depends on both the number of MAC operations and the energy efficiency of each MAC operation. The former can be estimated at design time; however, the latter depends on the intricate data reuse patterns and underlying hardware architecture. Hence, estimating it at design time is challenging. This article shows that the conventional approach to estimate the data reuse, viz. arithmetic intensity, does not always correctly estimate the degree of data reuse in DNNs since it gives equal importance to all the data types. We propose a novel model, termed “data type aware weighted arithmetic intensity” (DI), which accounts for the unequal importance of different data types in DNNs. We evaluate our model on 25 state-of-the-art DNNs on two GPUs. We show that our model accurately models data-reuse for all possible data reuse patterns for different types of convolution and different types of layers. We show that our model is a better indicator of the energy efficiency of DNNs. We also show its generality using the central limit theorem.

The capability of the self-attention mechanism to model the long-range dependencies has catapulted its deployment in vision models. Unlike convolution operators, self-attention offers infinite receptive field and enables compute- efficient modeling of global dependencies. However, the existing state-of-the-art attention mechanisms incur high compute and/or parameter overheads, and hence unfit for compact convolutional neural networks (CNNs). In this work, we propose a simple yet effective "Ultra-Lightweight Subspace Attention Mechanism" (ULSAM), which infers different attention maps for each feature map subspace. We argue that leaning separate attention maps for each feature subspace enables multi-scale and multi-frequency feature representation, which is more desirable for fine-grained image classification. Our method of subspace attention is orthogonal and complementary to the existing state-of-the- arts attention mechanisms used in vision models. ULSAM is end-to-end trainable and can be deployed as a plug-and- play module in the pre-existing compact CNNs. Notably, our work is the first attempt that uses a subspace attention mechanism to increase the efficiency of compact CNNs. To show the efficacy of ULSAM, we perform experiments with MobileNet-V1 and MobileNet-V2 as backbone architectures on ImageNet-1K and three fine-grained image classification datasets. We achieve ≈13% and ≈25% reduction in both the FLOPs and parameter counts of MobileNet-V2 with a 0.27% and more than 1% improvement in top-1 accuracy on the ImageNet-1K and fine-grained image classification datasets (respectively). Code and trained models are available at https://github.com/Nandan91/ULSAM.

The remarkable predictive performance of deep neural networks (DNNs) has led to their adoption in service domains of unprecedented scale and scope. However, the widespread adoption and growing commercialization of DNNs have underscored the importance of intellectual property (IP) protection. Devising techniques to ensure IP protection has become necessary due to the increasing trend of outsourcing the DNN computations on the untrusted accelerators in cloud-based services. The design methodologies and hyper-parameters of DNNs are crucial information, and leaking them may cause massive economic loss to the organization. Furthermore, the knowledge of DNN’s architecture can increase the success probability of an adversarial attack where an adversary perturbs the inputs and alters the prediction. In this work, we devise a two-stage attack methodology “DeepPeep,” which exploits the distinctive characteristics of design methodologies to reverse-engineer the architecture of building blocks in compact DNNs. We show the efficacy of “DeepPeep” on P100 and P4000 GPUs. Additionally, we propose intelligent design maneuvering strategies for thwarting IP theft through the DeepPeep attack and proposed “Secure MobileNet-V1.” Interestingly , compared to vanilla MobileNet-V1, secure MobileNet-V1 provides a significant reduction in inference latency (≈60%) and improvement in predictive performance (≈2%) with very low memory and computation overheads.