ML Trends

Newsletter #7 - May 27, 2019

The battle for AI supremacy

The Battle for AI Supremacy

Google or Facebook? TensorFlow or PyTorch? Centralized or Distributed Machine Learning? 

The stage is set for a battle to win the biggest share of the AI pie, with a projected market size of around $170 billion by 2025 at stake.

In this edition of ML Trends, we showcase the biggest names in the AI/ML, in terms of companies, frameworks and AI-specific IOT devices.

The Companies

Which companies are part of the AI competition? If you guessed GAFAM, FAMGA, AMGAF or any other combination of the acronym, you guessed right. Google, Apple, Facebook, Amazon and Microsoft are some of the biggest names in the battle for consolidating the AI market. 

Since 2006, the combined spend by these companies on AI startups is approximately $6.4 Billion.

AI solutions have become deeply integrated within the frameworks of major Cloud-service providers. In the first week of May, Google launched the Google AI Platform, an end-to-end platform to build, run and manage machine learning projects. Microsoft unveiled a new service within Azure for AI & Blockchain projects. In March, AWS launched the AWS Deep Learning Containers service, which allows users to deploy Docker images preinstalled with popular deep learning frameworks.

In gadgets, Apple's upcoming SoC, the A13, which Apple calls the "Next-generation Neural Engine" will allow next-gen Apple mobile devices to use the advanced neural hardware to apply advanced computer vision, image processing and speech recognition algorithms. Google is banking on its proprietary TPU technology, which offers GPU power at scale with a low cost. Amazon is still developing it's Inferentia chip, which is a response to Google's TPU. Microsoft was rumored to be partnering with the Huawei to design it's AI chip. The biggest loser in this race? Nvidia.

The current trend seems to be a consolidation of AI by creating complete AI-based ecosystems to edge out market competitors. It'll be very interesting to see how this battle pans out. 

For more information, visit this link.

The Frameworks

The sudden rise of tools and techniques in AI and ML of late has been striking. There are a plethora of tools available, such as Chainer, TensorFlow, CNTK, Gluon, Torch, PyTorch just to name a few. However, there are two libraries that are starting to emerge as the preferred choice by the AI developer community.

TensorFlow is the best in class, but PyTorch is a new entrant in the field that is slowly eating away TensorFlow's market share. So, PyTorch vs TensorFlow, which one is better? How do the two deep learning libraries compare to one another?

As it stands, TensorFlow is certainly favored and used more than PyTorch in production, primarily due to a more mature ecosystem, a larger community support and a wide array of other tools that support it.

PyTorch is relatively new and has a smaller community than TensorFlow. But PyTorch is faster and more efficient. Because of its efficiency, speed & ease of use and clean, dynamic API, PyTorch is a great alternative to TensorFlow. Companies such as Facebook, Twitter, Uber, Instagram and many others are using PyTorch as their framework of choice to train deep learning models.

Instagram uses PyTorch for tasks such as automatic image caption generation, abusive text detection, fake news detection, automatic subtitle generation for videos, generation of audio captions for images to support the blind users.  

Uber has built a probabilistic language on top of PyTorch called Pyro, which it uses extensively in research pertaining to Reinforcement Learning and Genetic Architecture Search. 

Similarly there are various other tools that are built on top of PyTorch. Some of the more popular ones are: 

  1. AllenNLP: De Facto framework for building modern advanced NLP pipelines, along with spaCy; developed by the Allen.AI research lab.
  2. OpenMined: The only platform to do privacy preserving federated deep learning built by the OpenMined community.
  3. Flair: Another advanced deep learning library built with PyTorch at it's core by Zolando research.

TensorFlow is used by a wide range of organizations such as Tesla, DeepMind, Google Brain etc.  

  1. Tesla's self-driving systems incorporate a whole lot of Computer Vision models which are built on top of TensorFlow. 
  2. DeepMind uses TensorFlow and libraries built on top of TensorFlow, such as Sonnet, to power it's AI systems like AlphaGo, AlphaZero & AlphaStar.

According to a survey conducted by Developer Economics, 43% of AI developers use TensorFlow or PyTorch. Of them, 86% of them use TensorFlow.

Even though TensorFlow is the preferred AI framework, PyTorch is definitely gaining steam. TensorFlow is also going through a radical change, with TensorFlow 2.0's deep integration of Keras. Backward compatibility and developer disagreement might create problems for the TensorFlow community, at least in the short term. Watch this space! 

Single-board Computers?

Single-board computers used to be a fun way to learn programming. But, in the past year, companies have invested in developing Single-board computers specifically for AI.

Nvidia unveiled the Jetson Nano Dev Kit in March 2019. The Dev Kit contains a small, powerful computer that run s on 5 watts and can run multiple neural-networks in parallel, just for $99.

In the same month, Google announced the release of Google Coral, powered by the Google Edge TPU, which priced at $150, can run Inception v4 30 times faster than a CPU.

Microsoft revamped Kinect and announced the Azure Kinect Developer Kit, an all-in-one perception system for computer vision and speech solutions back in February. The Azure Kinect Developer Kit is probably the most powerful dev kit on the market, and in turn, the most expensive one with a price tag of $399.

Intel has partnered with UP Board in developing the UP Squared AI Vision X Developer Kit.

The combination of distributed learning and IOT-based inference potentially creates a huge market for consumer-based applications, it's no wonder that software and hardware companies are investing in this nascent but fast-developing market.

We hope you enjoyed this edition of our newsletter. If you've got a couple of minutes to spare, do take this survey and help us get better.

If you found this interesting, you can help us grow by asking your friends, colleagues and enemies to subscribe to our newsletter.
Enjoy the rest of your weekend.

The ML Trends team
Copyright © 2019 mlTrends, All rights reserved.