Senseye is at the cutting-edge of neuroscience and computer vision. Their technology turns high definition, high frame-rate video footage into actionable intelligence about internal cognitive and behavioral processes accessible by businesses, academics, and government organizations. See how they achieved 120x faster machine learning with Saturn Cloud:
How to make a safer world
Leveraging insights about mental fatigue, drug and alcohol impairment, cognitive load, and advanced biometrics enables organizations to increase workplace safety, employee wellness, make training programs more efficient, tailor products to individuals, and much more. Senseye’s clients use these insights to enhance and grow their business, and to interact with their partners, employees, and consumers on a profound level.
The All Too Real Challenge of Computer Vision Data
Senseye pushes machines to their limits with the volume of HD video footage they use to train their machine learning models. The computing power needed to process large volumes of video data with custom technology is a limiting factor for the extraction of information from raw frames that are used to drive Senseye’s data products. Ultimate product effectiveness and success is directly tied to building and testing new ideas quickly at large scales.
When Speed is Everything
Recently, Senseye partnered with Saturn Cloud to make a breakthrough performance improvement in their machine learning work. Saturn Cloud is a data science and machine learning platform, which offers the highest-speed computing tools on the market.
Haven’t heard of Saturn? Here’s all you need to know: Random Forest on GPUs: 2000x Faster than Apache Spark
To test Saturn Cloud’s capabilities, Senseye chose one of their most compute-intensive processing pipelines. It was comprised of 900 ultra-high-definition videos, every 10 minutes in duration and recorded at 120Hz, which amounted to nearly 7.5TB on disk.
The video files are transformed into chunks of three-dimensional arrays and fed through several custom PyTorch models with more traditional analytic computer vision techniques to compute features, which are eventually used in downstream machine learning algorithms.
“With the scale we work at, it can take 60 days to extract features for machine learning. Lags like these affect our ability to iterate solutions, improve software, and increase performance. We are forced to carefully choose which solutions we want to test at scale with incomplete information, or run more solution candidates on smaller subsamples of data, which negatively affects the robustness and performance of our algorithms” — Andrew Sommerlot, Director of ML & AI
This process Senseye previously used equates to roughly 90 minutes of processing time per video per GPU (Titan RTX). Using the legacy on-premise GPU machine, the total processing time is 60 days.
“Fast iterations at full scale are ideal when creating an AI-driven product that needs high performance and confidence when deployed in real-world applications”
Let’s talk about “Highly Scalable Python Analytics”
Saturn Cloud allows teams to scale to an endless number of GPUs on AWS. What does this mean? It means the limiting factor of on-premise computing for Senseye is totally eliminated, opening a path to lightning-fast machine learning.
Senseye leveraged Saturn to scale up to 160 T4 GPUs in the cloud, moving away from the limits of the previously-utilized on-premise machines. This improvement required only adding 10 lines of code — a trivial change for the performance improvement.
Now, each video takes only 40 seconds to process, enabling Senseye to iterate every day as opposed to once every two months.
“Oh My God! The new workflow is 120x faster…”
This reduces the total runtime of 60 days down to just 11 hours in total. What does this entail? Well for starters, Senseye can build more accuracy into its predictions, and ultimately more value to its bottom line: fewer workplace accidents by its end-users means safer work environments and less insurance claims for its proprietors.
“Taking runtime down from 60 days to 11 hours is such an incredible improvement. We are able to fit in many more iterations on our models. This has a significant positive impact on the effectiveness of our product, which takes many iterations to perform at the standard necessary for our customers.” — Seth Weisberg, Principal Machine Learning Scientist (Senseye)
Let’s lay out the numbers side by side.