Naveen Verma

Yung-Chin Chen, Chung Peng Lee, Ze-Wei Liou et al.

Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges. While prior hypotheses characterize these as high-level scalar biases, we argue that they are merely the scalar intermediates of rigid, structural vector biases in the spike-carrying tokens. We show that these tokens converge to constant vectors after normalization that drive the attention sink and value-state drain mechanisms. We geometrically substantiate this by analyzing the coordination of projection weights: $W_K$ contrastively amplifies the vector, $W_Q$ aligns semantic tokens toward it, and $W_V$ projects it into the spectral null-space. Furthermore, we reveal that the model actively preserves these structural biases against Rotary Positional Embedding (RoPE) perturbations by localizing them in "zones of rotational stability" utilizing low-frequency bands and coherent channel pairs. Leveraging this, we propose INSERTQUANT, a post-training quantization (PTQ) framework that clamps spikes and restores their function via pre-computed template vectors. This renders activations strictly spike-free, enabling robust low-bit quantization with high fidelity. INSERTQUANT achieves parity with state-of-the-art per-tensor quantization methods on LLMs and uniquely generalizes beyond text to other modalities such as ViTs.

Zhiwu Zheng, Prakhar Kumar, Yenan Chen et al.

Soft robots have drawn great interest due to their ability to take on a rich range of shapes and motions, compared to traditional rigid robots. However, the motions, and underlying statics and dynamics, pose significant challenges to forming well-generalized and robust models necessary for robot design and control. In this work, we demonstrate a five-actuator soft robot capable of complex motions and develop a scalable simulation framework that reliably predicts robot motions. The simulation framework is validated by comparing its predictions to experimental results, based on a robot constructed from piezoelectric layers bonded to a steel-foil substrate. The simulation framework exploits the physics engine PyBullet, and employs discrete rigid-link elements connected by motors to model the actuators. We perform static and AC analyses to validate a single-unit actuator cantilever setup and observe close agreement between simulation and experiments for both the cases. The analyses are extended to the five-actuator robot, where simulations accurately predict the static and AC robot motions, including shapes for applied DC voltage inputs, nearly-static "inchworm" motion, and jumping (in vertical as well as vertical and horizontal directions). These motions exhibit complex non-linear behavior, with forward robot motion reaching ~1 cm/s. Our open-source code can be found at: https://github.com/zhiwuz/sfers.

Zhiwu Zheng, Prakhar Kumar, Yenan Chen et al.

Electrically-driven soft robots based on piezoelectric actuators may enable compact form factors and maneuverability in complex environments. In most prior work, piezoelectric actuators are used to control a single degree of freedom. In this work, the coordinated activation of five independent piezoelectric actuators, attached to a common metal foil, is used to implement inchworm-inspired crawling motion in a robot that is less than 0.5 mm thick. The motion is based on the control of its friction to the ground through the robot's shape, in which one end of the robot (depending on its shape) is anchored to the ground by static friction, while the rest of its body expands or contracts. A complete analytical model of the robot shape, which includes gravity, is developed to quantify the robot shape, friction, and displacement. After validation of the model by experiments, the robot's five actuators are collectively sequenced for inchworm-like forward and backward motion.

Hongyang Jia, Yinqi Tang, Hossein Valavi et al.

This paper presents a programmable in-memory-computing processor, demonstrated in a 65nm CMOS technology. For data-centric workloads, such as deep neural networks, data movement often dominates when implemented with today's computing architectures. This has motivated spatial architectures, where the arrangement of data-storage and compute hardware is distributed and explicitly aligned to the computation dataflow, most notably for matrix-vector multiplication. In-memory computing is a spatial architecture where processing elements correspond to dense bit cells, providing local storage and compute, typically employing analog operation. Though this raises the potential for high energy efficiency and throughput, analog operation has significantly limited robustness, scale, and programmability. This paper describes a 590kb in-memory-computing accelerator integrated in a programmable processor architecture, by exploiting recent approaches to charge-domain in-memory computing. The architecture takes the approach of tight coupling with an embedded CPU, through accelerator interfaces enabling integration in the standard processor memory space. Additionally, a near-memory-computing datapath both enables diverse computations locally, to address operations required across applications, and enables bit-precision scalability for matrix/input-vector elements, through a bit-parallel/bit-serial (BP/BS) scheme. Chip measurements show an energy efficiency of 152/297 1b-TOPS/W and throughput of 4.7/1.9 1b-TOPS (scaling linearly with the matrix/input-vector element precisions) at VDD of 1.2/0.85V. Neural network demonstrations with 1-b/4-b weights and activations for CIFAR-10 classification consume 5.3/105.2 $μ$J/image at 176/23 fps, with accuracy at the level of digital/software implementation (89.3/92.4 $\%$ accuracy).