[Adobe Stock]
Lightscline Co-Founders Ankur Verma, Ayush Goyal, and Soundar Kumara, Ph.D., are co-authors of the paper in Nature Scientific Reports.
Lightscline co-founder and CEO Ankur Verma, Ph.D., claims this approach can “Reduce your AI infrastructure and human capital costs by 100×.” While real-world results will vary, Verma clarifies that the 100× figure is more of an “order-of-magnitude” estimate in savings across edge computing, bandwidth, cloud resources, and processing time. According to Verma, these savings can be realized in four main areas: first, source of data collection—edge compute power; second, transmission bandwidth; third, cloud computing resources (FLOPS); and four, overall cloud processing time. The paper also demonstrates significant computational savings: Lightscline’s Shift-Invariant and spectrally stable Undersampled Network (SIUN) can cut FLOPS by a factor of 435 compared to a conventional CNN. This 435× reduction translates directly into lower compute costs.
A “selective attention” approach
Undersampled sine waves (A, B, C, D) illustrating how a periodic signal can appear random when sampled sparsely. Inspired by Figure 1 in [Verma et al., Sci Rep 14, 32041 (2024), https://doi.org/10.1038/s41598-024-83706-8
This “selective attention” approach is already being put to the test in real-world environments. In pilot tests with a Fortune 150 company, Lightscline demonstrated strong accuracy while training on only a fraction of raw sensor readings. According to Verma, their data was similar to the bearing-fault signals from the CWRU dataset used in the paper, so the estimated performance benefits were on the order of 100–400×. The company is exploring wearable sensor applications with NASA HumanWorks, aiming for enabling AI inference capabilities on low SWaP-C (size, weight, power, and costs) devices.
When less is more
In deep learning, models focusing on complex problems with high-dimensional inputs, such as image recognition and natural language processing, often benefit from larger datasets. This allows them to learn complex patterns, avoid overfitting, and generalize better to unseen data. Reinforcement learning (RL) shares some similarities in that more experience can lead to improved performance.
Ankur Verma, Ph.D.
“Lightscline’s neural-network approach discards most of the raw sensor data while preserving critical information. The company reports 96% accuracy using only 30% of one benchmark dataset, and a 435× reduction in FLOPS compared to standard CNNs. Real-world trials have shown 100-400x FLOPS reductions on just 10% of raw data, demonstrating a potential to reduce AI infrastructure costs by an order of magnitude.”
But not all data challenges are created equal. When it comes to sensor data in industrial or edge computing scenarios, collecting every bit of information at all times may be unnecessary—and potentially wasteful. That’s where Lightscline’s approach diverges. “We’re not arguing that you don’t need more data,” Verma said. “We’re saying that for sensor data, you don’t need to collect more than a certain fraction—you don’t need to collect at a certain rate,” he said. “What we’re saying is that augmenting a small dataset is more beneficial than operating on 100% of a dataset.”
Consider the example of underwater mapping for oil and gas exploration. “Every hour of raw data might require 40 hours of processing,” Verma says. Lightscline’s method frees data scientists from exhaustive manual feature extraction. Verma adds that with Lightscline’s end-to-end technique, the total time can be reduced from 40+ hours of manual effort to about 1 hour, enabling data teams to focus on scaling the number of models rather than labor-intensive data pre-processing. “We’re moving into this post-Nyquist era—how do we find the important information for different tasks? That’s problem-dependent sampling.”
“The fundamental question is: How can we find the important information for different tasks,” Verma said. “How can we generalize that? This architecture generalizes the sampling rate according to the information in your signal.”
Hardware-agnostic strategy
The versatility of Lightscline’s approach extends beyond software to encompass a wide range of hardware. They have deployed their models on everything from a $5 Raspberry Pi Pico microcontroller to an NVIDIA Jetson Nano with its 128-core GPU. The company also supports Arduino and the Intel/Ryzen processor.
from lightscline.lightscline import LightsclineEdge # Load data into Lightscline ls = LightsclineEdge(data=data, fs=SAMPLING_FREQUENCY) # Reduce the amount of data by 90% of the original ls.reduce_and_preprocess_data(per_reduction=90) # Train the model ls.train_model(verbose=True, n_iters=1000) # Checking the results ls.test_model()
At the right is Lightscline’s core Python implementation, requiring just four lines of code: loading data, reducing it by 90%, training the model, and testing the results.
Lightscline’s work adds to a broader movement questioning whether Nyquist–Shannon is still the final word on sampling. The company demonstrates that it is possible to move beyond Nyquist–Shannon constraints and achieve high accuracy with less data, a key point in sensor data analysis for applications ranging from industrial machinery and wearable devices to underwater exploration and satellite operations.
Even Meta has investigated similar undersampling approaches with its Byte Latent Transformer (BLT), a framework that processes information at the raw byte level and allocates compute only where data complexity requires it. “Even in language models, there’s this general notion of downsampling, adaptive sampling (like this [Meta] paper is showing), and sparsifying the network using low-rank matrices. In the language models,” Verma said. Meta’s and Lightscline’s work underscores a larger trend in compressing or undersampling data without undermining performance. “Nyquist defines how fast you sample,” Verma notes, “but we’re redefining what we save and discard. In many ways, less really can be more.”
Tell Us What You Think!