Princeton University is a prestigious Ivy League research institution renowned for its pioneering work in physics, computer science, and public policy. It cultivates fundamental discovery with societal and global impact. NECLA researchers worked with Princeton University to develop enhanced negative sample generation techniques for language-and-vision models. We collaborate to create more informative contrastive learning signals, leading to better alignment between visual inputs and language representations in multimodal AI systems.

Posts

Eric Blow Presents at the IEEE Photonics Conference Singapore on November 10th & 13th

/in /by

Eric Blow of NEC Labs will address how machine-learning methods applied to distributed acoustic-sensing data can monitor facility perimeters and detect intrusion via walk, dig, or drive events over buried optical fibre—for example achieving ~90% classification accuracy. Later in the week he will explore neuromorphic photonic RF sensing combining silicon photonics with FPGA-based recurrent neural networks, and his intern Yuxin Wang will present a finalist paper on scalable photonic neurons for automatic modulation classification.

Read more

Quantitative Bounds for Length Generalization in Transformers

/in /by

We provide quantitative bounds on the length of sequences required to be observed during training for a transformer to length generalize, e.g., to continue to perform well on sequences unseen during training. Our results improve on Huang et al. [8], who show that there is a finite training length beyond which length generalization is guaranteed, but for which they do not provide quantitative bounds.

Read more

DISC: Dynamic Decomposition Improves LLM Inference Scaling (SSI-FM)

/in /by

Inference scaling methods often rely on decomposing problems into steps, followed by sampling and selecting the best next steps. However, these steps and their sizes are typically fixed or depend on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically breaks down solution and reasoning traces into manageable steps during inference. By allocating compute more effectively, particularly by subdividing challenging steps and sampling them more frequently, dynamic decomposition significantly enhances inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

Read more

SFS: Smarter Code Space Search improves LLM Inference Scaling

/in /by

We frame code generation as a black-box optimization problem within the code space and demonstrate how optimization-inspired techniques can enhance inference scaling. Based on this perspective, we propose SCATTERED FOREST SEARCH (SFS), a novel approach that improves solution diversity and better exploits feedback during evolutionary search. Our theoretical analysis illustrates how these methods help avoid local optima during optimization, leading to more efficient exploration. Extensive experiments on HumanEval, MBPP, APPS, CodeContests, and Leetcode reveal significant performance gains. For instance, our method achieves a pass@1 rate of 67.1% on HumanEval+ and 87.2% on HumanEval with GPT-3.5, marking improvements of 8.6% and 4.3% over the state-of-the-art, while also halving the iterations needed to find the correct solution. Furthermore, our approach scales more efficiently than existing search techniques, including tree search, line search, and repeated sampling.

Read more

DISC: Dynamic Decomposition Improves LLM Inference Scaling (DL4C)

/in /by

Inference scaling methods often rely on decomposing problems into steps, followed by sampling and selecting the best next steps. However, these steps and their sizes are typically fixed or depend on domain knowledge. We propose dynamic decomposition, a method that adaptively and automatically breaks down solution and reasoning traces into manageable steps during inference. By allocating compute more effectively—particularly by subdividing challenging steps and sampling them more frequently—dynamic decomposition significantly enhances inference efficiency. Experiments on benchmarks such as APPS, MATH, and LiveCodeBench demonstrate that dynamic decomposition outperforms static approaches, including token-level, sentence-level, and single-step decompositions. These findings highlight the potential of dynamic decomposition to improve a wide range of inference scaling techniques.

Read more

Top 10 Most Legendary College Pranks of All-Time for April Fools’ Day

/in /by

At NEC Labs America, we celebrate innovation in all forms—even the brilliantly engineered college prank. From MIT’s police car on the Great Dome to Caltech hacking the Rose Bowl, these legendary stunts showcase next-level planning, stealth, and technical genius. Our Top 10 list honors the creativity behind pranks that made history (and headlines). This April Fools’ Day, we salute the hackers, makers, and mischief-makers who prove that brilliance can be hilarious.

Read more

Low-Latency Passive Thermal Stabilization of a Silicon Micro-Ring Resonator with Self-Heating

/in /by

Analog photonic information processing can be implemented with low chip area using wavelength-division multiplexed systems, which typically manipulate light using micro-ring resonators. Micro-rings are uniquely susceptible to thermal crosstalk, with negative system performance consequences if not addressed. Existing thermal sensitivity mitigation methods face drawbacks including high complexity, high latency, high digital and analog hardware requirements, and CMOS incompatibility. Here, we demonstrate a passive thermal desensitization mechanism for silicon micro-ring resonators exploiting self-heating resulting from optical absorption. We achieve a 49% reduction in thermal crosstalk sensitivity and 1 ?s adaptation latency using a system with no specialized micro-ring engineering, no additional control hardware, and no additional calibration. Our theoretical model indicates the potential for significant further desensitization gains with optimized microring designs. Self-heating desensitization can be combined with active thermal stabilization to achieve both responsiveness and accuracy or applied independently to thermally desensitize large photonic systems for signal processing or neural network inference.

Read more

Multi-terminal Germanium Photodetector in a Commercial Silicon Photonics Platform

/in /by

We report responsivity measurements of a multiterminal photodetection device in a commercial silicon photonics platform. The ratio of measured responsivities is found to track the relative terminal lengths. This can serve as a highly compact optoelectronic tap/diplexer. More importantly, complex biasing conditions of similar devices are promising for onchip reprogrammable opto-electronic responses in conventional silicon photonic platforms, with applications in reprogrammable photonics and neuromorphic photonics.

Read more

Radio-Frequency Linear Analysis and Optimization of Silicon Photonic Neural Networks

/in /by

Broadband analog signal processors utilizing silicon photonics have demonstrated a significant impact in numerous application spaces, offering unprecedented bandwidths, dynamic range, and tunability. In the past decade, microwave photonic techniques have been applied to neuromorphic processing, resulting in the development of novel photonic neural network architectures. Neuromorphic photonic systems can enable machine learning capabilities at extreme bandwidths and speeds. Herein, low-quality factor microring resonators are implemented to demonstrate broadband optical weighting. In addition, silicon photonic neural network architectures are critically evaluated, simulated, and optimized from a radio-frequency performance perspective. This analysis highlights the linear front-end of the photonic neural network, the effects of linear and nonlinear loss within silicon waveguides, and the impact of electrical preamplification.

Read more

Link Loss Analysis of Integrated Linear Weight Bank within Silicon Photonic Neural Network

/in /by

Over the last decade, silicon photonic neural networks have demonstrated the possibility of photonic-enabled machine learning at the edge. These systems enable low-latency ultra-wideband classifications, channel estimations, and many other signal characterization tasks within wireless environments. While these proof-of-concept experiments have yielded promising results, poor device and architectural designs have resulted in sub-optimal bandwidth and noise performance. As a result, the application space of this technology has been limited to GHz bandwidths and high signal-to-ratio input signals. By applying a microwave photonic perspective to these systems, the authors demonstrate high-bandwidth operation while optimizing for RF performance metrics: instantaneous bandwidth, link loss, noise figure, and dynamic range. The authors explore the extended capabilities due to these improved metrics and potential architectures to continue further optimization. The authors introduce novel architectures and RF analysis for RF-optimized neuromorphic photonic hardware.

Read more