Inside | Real news, curated by real humans
Inside AI

Inside AI (Nov 24th, 2019)

Happy Sunday, and welcome to the weekend edition of InsideAI!  I'm Rob May, and I try to write this weekend commentary based on the trends I'm seeing from the thousands of AI companies that pitch me each year as an early stage investor.  

Let's kick off today's issue with the most popular AI articles of the week:

Cerebras unveiled today what it claims is the industry’s fastest AI computer. The CS-1 can replace hundreds of racks of GPU-based computers using hundreds of kilowatts of power, according to the company, but measures only 26 inches tall and consumes about 20 kilowatts itself. Andrew Feldman, Cerebras CEO, says that the system dramatically cuts down on training time and, in the case of deep neural networks, can achieve single image classification in microseconds. The software side of the system allows users to write machine learning models using standard frameworks like Pytorch and Tensorflow. Already, Argonne National Labs, which plans to house a future exascale supercomputer, said it has deployed a CS-1 to accelerate neural networks in cancer studies and study traumatic brain injuries, among other uses. - TECHCRUNCH

IDC and Forrester recently made their annual predictions for the AI industry next year and beyond. Based on their surveys and other data, the firms believe that:

  • 25 percent of Fortune 500 companies will add AI building blocks - such as text analytics and machine learning - to their Robotic Process Automation efforts, forming hundreds of new intelligent process automation (IPA) use cases.
  • About half of AI platform providers, global systems integrators, and managed service providers will emphasize IPA in their portfolios.
  • By 2022, 75 percent of enterprises will embed intelligent automation into their technology and process development, using AI software to uncover operational and experiential insights. - FORBES

Some experts are asking for a more formalized, documented machine learning process to curtail the loss of model information when a data scientist leaves a company. Oftentimes, a data scientist may forget to pass along information about how a machine learning model was developed. When that employee leaves, the information can be lost permanently. As IEEE Spectrum notes, problems can arise if the model was found to be biased, since the company may not know what data was used to train the model. Companies like Microsoft already have an idea of how to document the machine-learning process and encourage co-development of ML models as much as possible. - IEEE SPECTRUM

People who work in highly paid white-collar occupations may be most exposed to AI in the future, according to a new study released Wednesday by the Brookings Institution. The research is based on Stanford Ph.D. student Michael Webb's method that examines AI patents and compares them with a Department of Labor database that describes occupational duties. The results predicted that workers in high-tech digital services, such as software publishing and computer system design, actually face high exposure to AI, which means that the technology will alter how they work (which may or may not mean that their jobs will be replaced). According to the study, more educated people and groups such as white and Asian men are most likely to face this "high exposure" to AI, as well as high tech areas like California or Seattle. - BAY AREA NEWS GROUP

I don't normally write about research papers in this newsletter, but this week Jeff Dean from Google released an awesome paper on The Deep Learning Revolution And Its Implications For Computer Architecture and Chip Design.  If you are even moderately technical, you will be able to understand this paper, so I encourage you to read it.  But here is my favorite section:

Deep learning models have three properties that make them different than many other kinds of more general purpose computations. First, they are very tolerant of reduced-precision computations. Second,

the computations performed by most models are simply different compositions of a relatively small handful of operations like matrix multiplies, vector operations, application of convolutional kernels, and other dense linear algebra calculations [Vanhoucke ​et al.​ 2011]. Third, many of the mechanisms developed over the past 40 years to enable general-purpose programs to run with high performance on modern CPUs, such as branch predictors, speculative execution, hyperthreaded-execution processing cores, and deep cache memory hierarchies and TLB subsystems are unnecessary for machine learning computations. So, the opportunity exists to build computational hardware that is specialized for dense, low-precision linear algebra, and not much else, but is still programmable at the level of specifying programs as different compositions of mostly linear algebra-style operations.

Dean explains very clearly why AI hardware is needed, and is going to be a big industry.  More importantly, it is going to lead to a big paradigm shift in how we think about compute, and the the things we want to compute.  I've written about AI hardware off and on over the years but, this paper is excellent, and got me really excited so I wanted to share it.  

Happy Sunday, and thanks for reading!

@robmay

Subscribe to Inside AI