We are Training Much Faster Now

Faster, better training is here outlined and described. 

We’re Training AI Twice as Fast This Year as Last New MLPerf rankings show training times plunging     By SAMUEL K. MOORE    in IEEE Spectrum

According to the best measures we’ve got, a set of benchmarks called MLPerf, machine-learning systems can be trained nearly twice as quickly as they could last year. It’s a figure that outstrips Moore’s Law, but also one we’ve come to expect. Most of the gain is thanks to software and systems innovations, but this year also gave the first peek at what some new processors, notably from Graphcore and Intel subsidiary Habana Labs, can do.

The once-crippling time it took to train a neural network to do its task is the problem that launched startups like Cerebras and SambaNova and drove companies like Google to develop machine-learning accelerator chips in house. But the new MLPerf data shows that training time for standard neural networks has gotten a lot less taxing in a short period of time. And that speedup has come from much more than just the advance of Moore’s Law.

This capability has only incentivized machine-learning experts to dream big. So the size of new neural networks continues to outpace computing power.

Called by some “the Olympics of machine learning,” MLPerf consists of eight benchmark tests: image recognition, medical-imaging segmentation, two versions of object detection, speech recognition, natural-language processing, recommendation, and a form of gameplay called reinforcement learning. (One of the object-detection benchmarks was updated for this round to a neural net that is closer to the state of the art.) Computers and software from 21 companies and institutions compete on any or all of the tests. This time around, officially called MLPerf Training 2.0, they collectively submitted 250 results.

Very few commercial and cloud systems were tested on all eight, but Nvidia director of product development for accelerated computing Shar Narasimhan gave an interesting example of why systems should be able to handle such breadth: Imagine a person with a smartphone snapping a photo of a flower and asking the phone: “What kind of flower is this?” It seems like a single request, but answering it would likely involve 10 different machine-learning models, several of which are represented in MLPerf.

To give a taste of the data, for each benchmark we’ve listed the fastest results for commercially available computers and cloud offerings (Microsoft Azure and Google Cloud) by how many machine-learning accelerators (usually GPUs) were involved. Keep in mind that some of these will be a category of one. For instance, there really aren’t that many places that can devote thousands of GPUs to a task. Likewise, there are some benchmarks where systems beat their nearest competitor by a matter of seconds or where five or more entries landed within a few minutes of each other. So if you’re curious about the nuances of AI performance, check out the complete list.  .... '   (much more at link) 

Vertical Thinking Broke Bottleneck in Powering High-Performance Computers

Vertical Thinking Broke Bottleneck in Powering High-Performance Computers

Princeton University, May 2, 2022  In Princeton Engineering

A team of researchers from Princeton University, Dartmouth College, Intel, and Google have developed a new method of power delivery that meets the needs of modern high-performance computers without sacrificing speed or efficiency. The method works with both small systems and large datacenters. The goal of the research was to deliver power to smaller areas to enable microprocessors to sit closer together, increase efficiency to reduce costs and prevent overheating, and quickly switch power among components to meet the demands of microprocessors. The researchers accomplished these goals by using capacitors rather than magnetics to process power, and by building the systems vertically instead of horizontally. Princeton's Minjie Chen said the result of their efforts is "a fully functioning system 10 times smaller than the best off-the-shelf." ... ' 

Computing Speed Limit?

Interesting look, does it clearly include quantum? 

Speed Limit of Computers Detected

Graz University of Technology (Austria), Christoph Pelzl, March 25, 2022

At Germany's Ludwig Maximilian University of Munich and the Max Planck Institute of Quantum Optics and Austria's Vienna (TU Wien) and Graz Universities of Technology, researchers have determined that a microchip's maximum signal transmission speed is about 1 petahertz (1 million gigahertz), roughly 100,000 times faster than current transistors. The researchers focused on increasing the switching signals of transistors to increase the speed of data transmission. They applied an ultra-short laser pulse with frequency in the extreme UV range to a lithium fluoride sample and found that by putting the electrons in the lithium fluoride into a more energetic state, the material became an electrical conductor for a brief period. The excited electrons were steered in a desired direction with a second laser pulse that lasted slightly longer. TU Wien's Joachim Burgförder said, "At about 1 petahertz there is an upper limit for controlled optoelectronic processes."

Fast Doglike Robotics

Faster tracking surveillance? Note its training abilities. 

 Mini Cheetah Runs

Massachusetts Institute of Technology (MIT) researchers have designed a new version of the Mini Cheetah robot that can achieve high running speeds more efficiently than its predecessors.

The robot learns in real time through an experiential model, and is capable of absorbing 100 days' worth of experience on diverse terrains in three hours by training its neural network in a simulator.

"The intuition behind why the robot's running skills work well in the real world is: of all the environments it sees in this simulator, some will teach the robot skills that are useful in the real world," said MIT's Gabriel Margolis and Ge Yang.

"When operating in the real world, our controller identifies and executes the relevant skills in real time."

From ZDNet   View Full Article  

How Fast Can Quantum Computers Process Information?

Fundamental question.   Always thinking back to the complexity question being posed.  

How Fast Can Quantum Computers Process Information?

By The Jerusalem Post, December 29, 2021

Physicists at Germany's University of Bonn and the Technion-Israel Institute of Technology have investigated the determinants of quantum-computer information processing speed. The researchers theoretically deduced the minimum time for quantum gates to transform the wave function and the information contained.

Technion's Gal Ness said the team "used fast light pulses to create a so-called quantum superposition of two states of [a cesium] atom. Figuratively speaking, the atom behaves as if it had two different colors at the same time."

The atom clones were compared at intervals via quantum interference to ascertain when a significant change of the matter wave transpired. Technion Professor Yoav Sagi said the results indicated the minimum wave-change time shortens as energy uncertainty increases, and demonstrated a speed limit imposed by the atom's average energy.

Jerusalem Post:      

Blinkist for Speeding up Content Acquisition

This brought to mind some work I was involved with at the University of Pennsylvania's Language Laboratory.   We worked with more efficient ways to deliver books and reading material, and one of the experiments was to compress books and make them available for select classes.  Overall a quite simple technique. Then measured the effort/effort/use/value that was achieved.    In general this worked well with some kinds of class content, we touched on neural based techniques   The approach called Blinkist, mentioned below,  is apparently doing some something similar.  I have not tested this, nor has Engadget,  who originally posted the below.  Also I have not received any compensation for posting this.   But am intrigued by the application.  May take a further look.  

Read bestselling books in 15 minutes with Blinkist   in Engadget

Blinkist Premium offers thousands of condensed nonfiction books and podcasts that you can process in just 15 minutes, with 70 new titles added every month.

Every year, we tell ourselves that we need to read more. Perhaps we’ll crack open that book we were gifted months ago. Maybe pick up an Amazon bestseller will finally get us into the habit. And yet, the reading list keeps growing.

Between your professional and personal life, there’s little time for intellectually engaging pursuits. That’s where Blinkist comes in handy. This app contains condensed ideas from thousands of bestselling nonfiction books, so you can stay up to date with your daily reading while you go about your busy schedule. Right now, you can purchase a two-year Blinkist Premium subscription for just $99 — that’s a $285 discount.

Blinkist identifies the main ideas from popular podcasts and nonfiction books and condenses them into digestible, 15-minute text and audio files. You can read or listen to over 4,500 summarized bestsellers ranging in topics from personal development to psychology. This subscription gives you unlimited access to everything in the Blinkist library, including 70 new titles that are added every month.... "  ... ' 

Reinventing for Speed

McKinsey always does an interesting job looking at these problems.Increase the speed while minimizing the risk of error.  reviewing.

Ready, set, go: Reinventing the organization for speed in the post-COVID-19 era
June 26, 2020      By Aaron De Smet, Daniel Pacthod, Charlotte Relyea, and Bob Sternfels

When the coronavirus pandemic erupted, companies had to change. Many business-as-usual approaches to serving customers, working with suppliers, and collaborating with colleagues—or just getting anything done—would have failed. They had to increase the speed of decision making, while improving productivity, using technology and data in new ways, and accelerating the scope and scale of innovation. And it worked. Organizations in a wide range of sectors and geographies have accomplished difficult tasks and achieved positive results in record time:

Redeploying talent. A global telco redeployed 1,000 store employees to inside sales and retrained them in three weeks.

Launching new business models. A US-based retailer launched curbside delivery in two days versus the previously-planned 18 months.

Improving productivity. An industrial factory ran at 90-percent-plus capacity with 40 percent of the workforce.

Developing new products. An engineering company designed and manufactured ventilators within a week.

Shifting operations. Coordinating with local officials, a major shipbuilder switched from three shifts to two, with thousands of employees.   ... "

Google Seeks Patent for ML Navigation Solution

Machine Learning patent for a particular application by Google:

Google seeks patent for ML model speed prediction to improve navigation services
The prediction of the speed of the vehicle can be used to predict a travel time, or recommend a route to the user.

By: Sajan C Kumar in FinancialExpress

In a bid to further strengthen its navigation services, American tech major Google has moved the Indian patent office seeking a patent to its new machine learning (ML) model for prediction of the speed of vehicles on particular routes, which will provide users the accurate travel time. ... " 

Speeding Up Learning Inference by 2X

New methods, technical:

New Technique Speeds Up Deep-Learning Inference on TensorFlow by 2x
by  Anthony Alford  in InfoQ

Researchers at North Carolina State University recently presented a paper at the International Conference on Supercomputing (ICS) on their new technique, "deep reuse" (DR), that can speed up inference time for deep-learning neural networks running on TensorFlow by up to 2x, with almost no loss of accuracy.

Dr. Xipeng Shen, along with graduate student Lin Ning, authored the paper describing the technique, which requires no special hardware or changes to the deep-learning model. By taking advantage of similarities in the data values that are input into a neural network layer, DR eliminates redundant computation during inference, reducing the total time taken. Reducing computation also reduces power consumption, a key feature for mobile or embedded applications. In experiments running several common computer-vision deep-learning models on GPUs, including CifarNet, AlexNet, and VGG-19, DR achieved from 1.75X to 2.02X speedup, with an increase in error of 0.0005. In some cases, DR actually improved accuracy slightly. In similar experiments on a mobile phone, DR "achieves an average of 2.12x speedup for CifarNet and 2.55X for AlexNet."  .... "