Ian Colbert

Giuseppe Franco, Ian Colbert, Pablo Monteagudo-Lago et al.

Quantizing large language models (LLMs) to low-precision floating-point representations is central to efficient deployment, yet applying a single bit-width uniformly across all layers is sub-optimal in terms of both performance and accuracy. This work introduces dMX, a differentiable mixed-precision quantization framework for learnable floating-point bit-width assignment. We study its application for the microscaling floating-point (MXFP) family of data types defined by the Open Compute Project (OCP) standard. The per-layer bit-width assignment is formulated as a continuous optimization problem in which each layer's floating-point format format is parameterized by a scalar parameter, folding the multi-variate design space into a single learnable offset. During training this offset takes continuous values, avoiding sudden oscillations between discrete quantization formats. A temperature-based annealing schedule progressively discretizes the learned offsets, ensuring that the final configuration maps to hardware-compatible MXFP formats without abrupt transitions between training and inference behavior. A target-aware regularization term steers the average bit-width toward a user-specified budget, serving as a coarse-grained proxy for inference cost and balancing model quality against deployment efficiency. We performed experiments on different families of LLM, such as Llama, Qwen3, and SmolLM2, evaluating perplexity on WikiText-2 and accuracy on four zero-shot reasoning benchmarks. Across these settings, dMX consistently yields Pareto-dominating models and improves over Kullback-Leibler (KL) divergence-based layer-selection heuristics, efficiently navigating trade-offs between model quality and average bit-width.

Lara D'Agata, Carlos Agulló-Domingo, Óscar Vera-López et al.

Fully homomorphic encryption (FHE) has recently attracted significant attention as both a cryptographic primitive and a systems challenge. Given the latest advances in accelerated computing, FHE presents a promising opportunity for progress, with applications ranging from machine learning to information security. We target the most computationally intensive operation in deep neural networks from a hardware perspective, matrix multiplication (matmul), and adapt it for execution on AMD GPUs. We propose a new optimized method that improves the runtime and complexity of ciphertext matmul by using FIDESlib, a recent open-source FHE library designed specifically for GPUs. By exploiting sparsity in both operands, our sparse matmul implementation outperforms its CPU counterpart by up to $3.0\times$ and reduces the time complexity from cubic to semi-linear, demonstrating an improvement over existing FHE matmul implementations.

Ian Colbert, Giuseppe Franco, Fabian Grob et al.

When quantizing weights and activations to increasingly narrower representations, the cost of additions begins to dominate that of multiplications in multiply-accumulate (MAC) units. Recent studies show that reducing addition costs via low-precision accumulation improves throughput, power, and area across inference platforms, albeit with an increased risk of overflow. Accumulator-aware quantization research has so far only considered the quantization-aware training (QAT) paradigm, in which models are fine-tuned or trained from scratch with quantization in the loop. As models and datasets continue to grow in size, QAT techniques become increasingly more expensive, which has motivated the recent surge in post-training quantization (PTQ) research. To bridge this gap, we introduce AXE, the first accumulator-aware quantization framework explicitly designed to endow overflow avoidance guarantees to PTQ algorithms. We present theoretical motivation for AXE and demonstrate its flexibility by implementing it on top of two existing algorithms: GPFQ and OPTQ. We design AXE to support multi-stage accumulation, opening the door to full datapath optimization for the first time. We evaluate AXE using recent language generation models; when quantizing Llama3 8B for a 16-bit multi-stage accumulation datapath, AXE maintains up to 98% of the FP16 perplexity, surpassing naive bit width manipulation by up to 15%.

Ian Colbert, Alessandro Pappalardo, Jakoba Petri-Koenig

We present accumulator-aware quantization (A2Q), a novel weight quantization method designed to train quantized neural networks (QNNs) to avoid overflow when using low-precision accumulators during inference. A2Q introduces a unique formulation inspired by weight normalization that constrains the L1-norm of model weights according to accumulator bit width bounds that we derive. Thus, in training QNNs for low-precision accumulation, A2Q also inherently promotes unstructured weight sparsity to guarantee overflow avoidance. We apply our method to deep learning-based computer vision tasks to show that A2Q can train QNNs for low-precision accumulators while maintaining model accuracy competitive with a floating-point baseline. In our evaluations, we consider the impact of A2Q on both general-purpose platforms and programmable hardware. However, we primarily target model deployment on FPGAs because they can be programmed to fully exploit custom accumulator bit widths. Our experimentation shows accumulator bit width significantly impacts the resource efficiency of FPGA-based accelerators. On average across our benchmarks, A2Q offers up to a 2.3x reduction in resource utilization over 32-bit accumulator counterparts with 99.2% of the floating-point model accuracy.

Ian Colbert, Mehdi Saeedi

Recent advancements in deep reinforcement learning have brought forth an impressive display of highly skilled artificial agents capable of complex intelligent behavior. In video games, these artificial agents are increasingly deployed as non-playable characters (NPCs) designed to enhance the experience of human players. However, while it has been shown that the convincing human-like behavior of NPCs leads to increased engagement in video games, the believability of an artificial agent's behavior is most often measured solely by its proficiency at a given task. Recent work has hinted that proficiency alone is not sufficient to discern human-like behavior. Motivated by this, we build a non-parametric two-sample hypothesis test designed to compare the behaviors of artificial agents to those of human players. We show that the resulting $p$-value not only aligns with anonymous human judgment of human-like behavior, but also that it can be used as a measure of similarity.

Ian Colbert, Alessandro Pappalardo, Jakoba Petri-Koenig

We introduce a quantization-aware training algorithm that guarantees avoiding numerical overflow when reducing the precision of accumulators during inference. We leverage weight normalization as a means of constraining parameters during training using accumulator bit width bounds that we derive. We evaluate our algorithm across multiple quantized models that we train for different tasks, showing that our approach can reduce the precision of accumulators while maintaining model accuracy with respect to a floating-point baseline. We then show that this reduction translates to increased design efficiency for custom FPGA-based accelerators. Finally, we show that our algorithm not only constrains weights to fit into an accumulator of user-defined bit width, but also increases the sparsity and compressibility of the resulting weights. Across all of our benchmark models trained with 8-bit weights and activations, we observe that constraining the hidden layers of quantized neural networks to fit into 16-bit accumulators yields an average 98.2% sparsity with an estimated compression rate of 46.5x all while maintaining 99.2% of the floating-point performance.

Alexander Cann, Ian Colbert, Ihab Amer

The widespread adoption of deep neural networks in computer vision applications has brought forth a significant interest in adversarial robustness. Existing research has shown that maliciously perturbed inputs specifically tailored for a given model (i.e., adversarial examples) can be successfully transferred to another independently trained model to induce prediction errors. Moreover, this property of adversarial examples has been attributed to features derived from predictive patterns in the data distribution. Thus, we are motivated to investigate the following question: Can adversarial defenses, like adversarial examples, be successfully transferred to other independently trained models? To this end, we propose a deep learning-based pre-processing mechanism, which we refer to as a robust transferable feature extractor (RTFE). After examining theoretical motivation and implications, we experimentally show that our method can provide adversarial robustness to multiple independently pre-trained classifiers that are otherwise ineffective against an adaptive white box adversary. Furthermore, we show that RTFEs can even provide one-shot adversarial robustness to models independently trained on different datasets.

Sai Sanjeet, Ian Colbert, Pablo Monteagudo-Lago et al.

Recent post-training quantization (PTQ) methods have adopted block rotations to diffuse outliers prior to rounding. While this reduces the overhead of full-vector rotations, the effect of block structure on outlier suppression remains poorly understood. To fill this gap, we present the first systematic, non-asymptotic analysis of outlier suppression for block Hadamard rotations. Our analysis reveals that outlier suppression is fundamentally limited by the geometry of the input vector. In particular, post-rotation outliers are deterministically minimized when the pre-rotation $\ell_1$ norm mass is evenly distributed across blocks. Guided by these insights, we introduce MixQuant, a block rotation-aware PTQ framework that redistributes activation mass via permutations prior to rotation. We propose a greedy mass diffusion algorithm to calibrate permutations by equalizing the expected blockwise $\ell_1$ norms. To avoid adding inference overhead, we identify permutation-equivariant regions in transformer architectures to merge the resulting permutations into model weights before deployment. Experiments show that MixQuant consistently improves accuracy across all block sizes, recovering up to 90% of the full-vector rotation perplexity when quantizing Llama3 1B to INT4 with block size 16, compared to 46% without permutations.

Daniel Campa, Mehdi Saeedi, Ian Colbert et al.

Navigation path traces play a crucial role in video game design, serving as a vital resource for both enhancing player engagement and fine-tuning non-playable character behavior. Generating such paths with human-like realism can enrich the overall gaming experience, and evaluating path traces can provide game designers insights into player interactions. Despite the impressive recent advancements in deep learning-based generative modeling, the video game industry hesitates to adopt such models for path generation, often citing their complex training requirements and interpretability challenges. To address these problems, we propose a novel path generation and evaluation approach that is grounded in principled nonparametric statistics and provides precise control while offering interpretable insights. Our path generation method fuses two statistical techniques: (1) nonparametric model-free transformations that capture statistical characteristics of path traces through time; and (2) copula models that capture statistical dependencies in space. For path evaluation, we adapt a nonparametric three-sample hypothesis test designed to determine if the generated paths are overfit (mimicking the original data too closely) or underfit (diverging too far from it). We demonstrate the precision and reliability of our proposed methods with empirical analysis on two existing gaming benchmarks to showcase controlled generation of diverse navigation paths. Notably, our novel path generator can be fine-tuned with user controllable parameters to create navigation paths that exhibit varying levels of human-likeness in contrast to those produced by neural network-based agents. The code is available at https://github.com/daniel-campa/mf-copula.

Haoyu Zhang, Shihao Zhang, Ian Colbert et al.

Post-training quantization (PTQ) has become a crucial tool for reducing the memory and compute costs of modern deep neural networks, including large language models (LLMs). Among PTQ algorithms, the OPTQ framework-also known as GPTQ-has emerged as a leading method due to its computational efficiency and strong empirical performance. Despite its widespread adoption, however, OPTQ lacks rigorous quantitative theoretical guarantees. This paper presents the first quantitative error bounds for both deterministic and stochastic variants of OPTQ, as well as for Qronos, a recent related state-of-the-art PTQ algorithm. We analyze how OPTQ's iterative procedure induces quantization error and derive non-asymptotic 2-norm error bounds that depend explicitly on the calibration data and a regularization parameter that OPTQ uses. Our analysis provides theoretical justification for several practical design choices, including the widely used heuristic of ordering features by decreasing norm, as well as guidance for selecting the regularization parameter. For the stochastic variant, we establish stronger infinity-norm error bounds, which enable control over the required quantization alphabet and are particularly useful for downstream layers and nonlinearities. Finally, we extend our analysis to Qronos, providing new theoretical bounds, for both its deterministic and stochastic variants, that help explain its empirical advantages.

Giuseppe Franco, Pablo Monteagudo-Lago, Ian Colbert et al.

The size of a model has been a strong predictor of its quality, as well as its cost. As such, the trade-off between model cost and quality has been well-studied. Post-training optimizations like quantization and pruning have typically focused on reducing the overall volume of pre-trained models to reduce inference costs while maintaining model quality. However, recent advancements have introduced optimization techniques that, interestingly, expand models post-training, increasing model size to improve quality when reducing volume. For instance, to enable 4-bit weight and activation quantization, incoherence processing often necessitates inserting online Hadamard rotations in the compute graph, and preserving highly sensitive weights often calls for additional higher precision computations. However, if application requirements cannot be met, the prevailing solution is to relax quantization constraints. In contrast, we demonstrate post-training model expansion is a viable strategy to improve model quality within a quantization co-design space, and provide theoretical justification. We show it is possible to progressively and selectively expand the size of a pre-trained large language model (LLM) to improve model quality without end-to-end retraining. In particular, when quantizing the weights and activations to 4 bits for Llama3 1B, we reduce the gap to full-precision perplexity by an average of 9% relative to both QuaRot and SpinQuant with only 5% more parameters, which is still a 3.8% reduction in volume relative to a BF16 reference model.

Aidan Ferguson, Perry Gibson, Lara D'Agata et al.

The deployment of deep neural networks (DNNs) in privacy-sensitive environments is constrained by computational overheads in fully homomorphic encryption (FHE). This paper explores unstructured sparsity in FHE matrix multiplication schemes as a means of reducing this burden while maintaining model accuracy requirements. We demonstrate that sparsity can be exploited in arbitrary matrix multiplication, providing runtime benefits compared to a baseline naive algorithm at all sparsity levels. This is a notable departure from the plaintext domain, where there is a trade-off between sparsity and the overhead of the sparse multiplication algorithm. In addition, we propose three sparse multiplication schemes in FHE based on common plaintext sparse encodings. We demonstrate the performance gain is scheme-invariant; however, some sparse schemes vastly reduce the memory storage requirements of the encrypted matrix at high sparsity values. Our proposed sparse schemes yield an average performance gain of 2.5x at 50% unstructured sparsity, with our multi-threading scheme providing a 32.5x performance increase over the equivalent single-threaded sparse computation when utilizing 64 cores.

Tian Liu, Alex Cann, Ian Colbert et al.

While the rapid advancements in the reinforcement learning (RL) research community have been remarkable, the adoption in commercial video games remains slow. In this paper, we outline common challenges the Game AI community faces when using RL-driven NPCs in practice, and highlight the intersection of RL with traditional behavior trees (BTs) as a crucial juncture to be explored further. Although the BT+RL intersection has been suggested in several research papers, its adoption is rare. We demonstrate the viability of this approach using AMD Schola -- a plugin for training RL agents in Unreal Engine -- by creating multi-task NPCs in a complex 3D environment inspired by the commercial video game ``The Last of Us". We provide detailed methodologies for jointly training RL models with BTs while showcasing various skills.

Ian Colbert, Alessandro Pappalardo, Jakoba Petri-Koenig et al.

Quantization techniques commonly reduce the inference costs of neural networks by restricting the precision of weights and activations. Recent studies show that also reducing the precision of the accumulator can further improve hardware efficiency at the risk of numerical overflow, which introduces arithmetic errors that can degrade model accuracy. To avoid numerical overflow while maintaining accuracy, recent work proposed accumulator-aware quantization (A2Q), a quantization-aware training method that constrains model weights during training to safely use a target accumulator bit width during inference. Although this shows promise, we demonstrate that A2Q relies on an overly restrictive constraint and a sub-optimal weight initialization strategy that each introduce superfluous quantization error. To address these shortcomings, we introduce: (1) an improved bound that alleviates accumulator constraints without compromising overflow avoidance; and (2) a new strategy for initializing quantized weights from pre-trained floating-point checkpoints. We combine these contributions with weight normalization to introduce A2Q+. We support our analysis with experiments that show A2Q+ significantly improves the trade-off between accumulator bit width and model accuracy and characterize new trade-offs that arise as a consequence of accumulator constraints.

Ian Colbert, Jake Daly, Norm Rubin

GPU compilers are complex software programs with many optimizations specific to target hardware. These optimizations are often controlled by heuristics hand-designed by compiler experts using time- and resource-intensive processes. In this paper, we developed a GPU compiler autotuning framework that uses off-policy deep reinforcement learning to generate heuristics that improve the frame rates of graphics applications. Furthermore, we demonstrate the resilience of these learned heuristics to frequent compiler updates by analyzing their stability across a year of code check-ins without retraining. We show that our machine learning-based compiler autotuning framework matches or surpasses the frame rates for 98% of graphics benchmarks with an average uplift of 1.6% up to 15.8%.

Xinyu Zhang, Ian Colbert, Ken Kreutz-Delgado et al.

Quantization and pruning are core techniques used to reduce the inference costs of deep neural networks. State-of-the-art quantization techniques are currently applied to both the weights and activations; however, pruning is most often applied to only the weights of the network. In this work, we jointly apply novel uniform quantization and unstructured pruning methods to both the weights and activations of deep neural networks during training. Using our methods, we empirically evaluate the currently accepted prune-then-quantize paradigm across a wide range of computer vision tasks and observe a non-commutative nature when applied to both the weights and activations of deep neural networks. Informed by these observations, we articulate the non-commutativity hypothesis: for a given deep neural network being trained for a specific task, there exists an exact training schedule in which quantization and pruning can be introduced to optimize network performance. We identify that this optimal ordering not only exists, but also varies across discriminative and generative tasks. Using the optimal training schedule within our training framework, we demonstrate increased performance per memory footprint over existing solutions.

Ian Colbert, Ken Kreutz-Delgado, Srinjoy Das

A novel energy-efficient edge computing paradigm is proposed for real-time deep learning-based image upsampling applications. State-of-the-art deep learning solutions for image upsampling are currently trained using either resize or sub-pixel convolution to learn kernels that generate high fidelity images with minimal artifacts. However, performing inference with these learned convolution kernels requires memory-intensive feature map transformations that dominate time and energy costs in real-time applications. To alleviate this pressure on memory bandwidth, we confine the use of resize or sub-pixel convolution to training in the cloud by transforming learned convolution kernels to deconvolution kernels before deploying them for inference as a functionally equivalent deconvolution. These kernel transformations, intended as a one-time cost when shifting from training to inference, enable a systems designer to use each algorithm in their optimal context by preserving the image fidelity learned when training in the cloud while minimizing data transfer penalties during inference at the edge. We also explore existing variants of deconvolution inference algorithms and introduce a novel variant for consideration. We analyze and compare the inference properties of convolution-based upsampling algorithms using a quantitative model of incurred time and energy costs and show that using deconvolution for inference at the edge improves both system latency and energy efficiency when compared to their sub-pixel or resize convolution counterparts.

Siqiao Ruan, Ian Colbert, Ken Kreutz-Delgado et al.

The use of Deep Learning hardware algorithms for embedded applications is characterized by challenges such as constraints on device power consumption, availability of labeled data, and limited internet bandwidth for frequent training on cloud servers. To enable low power implementations, we consider efficient bitwidth reduction and pruning for the class of Deep Learning algorithms known as Discriminative Deep Belief Networks (DDBNs) for embedded-device classification tasks. We train DDBNs with both generative and discriminative objectives under an approximate computing framework and analyze their power-at-performance for supervised and semi-supervised applications. We also investigate the out-of-distribution performance of DDBNs when the inference data has the same class structure yet is statistically different from the training data owing to dynamic real-time operating environments. Based on our analysis, we provide novel insights and recommendations for choice of training objectives, bitwidth values, and accuracy sensitivity with respect to the amount of labeled data for implementing DDBN inference with minimum power consumption on embedded hardware platforms subject to accuracy tolerances.

Alexander Potapov, Ian Colbert, Ken Kreutz-Delgado et al.

Stochastic-sampling-based Generative Neural Networks, such as Restricted Boltzmann Machines and Generative Adversarial Networks, are now used for applications such as denoising, image occlusion removal, pattern completion, and motion synthesis. In scenarios which involve performing such inference tasks with these models, it is critical to determine metrics that allow for model selection and/or maintenance of requisite generative performance under pre-specified implementation constraints. In this paper, we propose a new metric for evaluating generative model performance based on $p$-values derived from the combined use of Maximum Mean Discrepancy (MMD) and permutation-based (PT-based) resampling, which we refer to as PT-MMD. We demonstrate the effectiveness of this metric for two cases: (1) Selection of bitwidth and activation function complexity to achieve minimum power-at-performance for Restricted Boltzmann Machines; (2) Quantitative comparison of images generated by two types of Generative Adversarial Networks (PGAN and WGAN) to facilitate model selection in order to maximize the fidelity of generated images. For these applications, our results are shown using Euclidean and Haar-based kernels for the PT-MMD two sample hypothesis test. This demonstrates the critical role of distance functions in comparing generated images against their corresponding ground truth counterparts as what would be perceived by human users.

Ian Colbert, Ken Kreutz-Delgado, Srinjoy Das

The power budget for embedded hardware implementations of Deep Learning algorithms can be extremely tight. To address implementation challenges in such domains, new design paradigms, like Approximate Computing, have drawn significant attention. Approximate Computing exploits the innate error-resilience of Deep Learning algorithms, a property that makes them amenable for deployment on low-power computing platforms. This paper describes an Approximate Computing design methodology, AX-DBN, for an architecture belonging to the class of stochastic Deep Learning algorithms known as Deep Belief Networks (DBNs). Specifically, we consider procedures for efficiently implementing the Discriminative Deep Belief Network (DDBN), a stochastic neural network which is used for classification tasks, extending Approximation Computing from the analysis of deterministic to stochastic neural networks. For the purpose of optimizing the DDBN for hardware implementations, we explore the use of: (a)Limited precision of neurons and functional approximations of activation functions; (b) Criticality analysis to identify nodes in the network which can operate at reduced precision while allowing the network to maintain target accuracy levels; and (c) A greedy search methodology with incremental retraining to determine the optimal reduction in precision for all neurons to maximize power savings. Using the AX-DBN methodology proposed in this paper, we present experimental results across several network architectures that show significant power savings under a user-specified accuracy loss constraint with respect to ideal full precision implementations.