The Resilient ML Stack is Modular

So how can an organization stay agile within an ever-shifting ML landscape? Part of the answer lies with establishing a modular ML architecture. Comet will be joined on July 6th in a live webinar by the AI Infrastructure Alliance and Superb AI to discuss how to accelerate AI value with modular MLOps. Register now.



Sponsored Post

 
By Yasmeen Kashef

As a broad technology, AI has leaped and hobbled through the adoption cycle. With the market expected to break $500B by 20231, AI adoption accelerates even through the pandemic. Everyone agrees that AI is everywhere and it’s not going away2. The fear of being left behind or missing the opportunity to accelerate revenue growth is a mantra that haunts any go-to-market strategy. 

At the same time, the hobble is evidenced by the lack of models actually delivering business value. In a recent survey by Comet, 68% of respondents admit to scrapping 40-80% of their experiments altogether. We should be better off by now, right? Understandably, machine learning is complex. MLOps best practices continues to evolve as everyone learns together.

When you look at the evolving marketplace, the tooling landscape for ML is still a mess3. When Chip Huyen reviewed machine learning tools4 just over a year ago, the explosion of tools in the market was already accelerating.

 

Graph showing years on x-axis from 2008 to 2020 and number of ML tools on the y-axis. The graph shows a hockey-stick growth of less than 50 ML tools in 2008 to more than 200 ML tools in 2020.
https://huyenchip.com/2020/12/30/mlops-v2.html

 

If you look at the landscape today5, the vast number of options can be overwhelming. It’s also not obvious which tools are widely adopted and which will become consolidated. 

 

Collection of data and machine learning companies organized in various sections. Approximately more than 1,000 logos are represented.
https://mattturck.com/data2021/

 

Enterprises that are navigating this environment today are likely contending with the decisions on how to replace or augment their current unique in-house solutions. Considering that AI is often born out of R&D, the research is not designed to be production-level from the outset. This incurs its own technical debt. So when it’s time to divert engineering resources from maintaining internal tools to purchasing a replacement, the choice of which tool to invest into is not an easy one to make. 

To complicate it further, the Comet survey also found that budgeting for tools is inadequate with 88% of respondents having an annual budget of less than $75,000 for tools and infrastructure. Fortunately, most organizations are increasing their ML budgets for 2022. 

So how can an organization stay agile within an ever-shifting ML landscape? Part of the answer lies with establishing a modular ML architecture5. This enables companies to customize their workflows and more easily replace components of their in-house systems with market tools that have become more established. It also enables data scientists to work with tools that cater to their needs, while collaborating cross-functionally with other teams. 

We’re looking forward to watching this space evolve and mature. Why don’t you join us? Comet will be joined by the AI Infrastructure Alliance and Superb AI to discuss how to accelerate AI value with modular MLOps. We’ll be discussing the characteristics of the modern MLOps stack and challenges that companies face today with building an integrated MLOps strategy.

Register here to join us!

 
Banner image displaying details of Comet’s webinar on July 6th with speakers Gideon Mendels from Comet, James Le from Superb AI and Daniel Jeffries from AIIA.

 

Sources

 

  1. IDC. Feb 2022. IDC Forecasts Companies to Increase Spend on AI Solutions by 19.6% in 2022. https://www.idc.com/getdoc.jsp?containerId=prUS48881422
  2. Mike Loukides. Mar 2022. AI Adoption in the Enterprise 2022. https://www.oreilly.com/radar/ai-adoption-in-the-enterprise-2022/
  3. Mihail Eric. Mar 2022. MLOps Is a Mess But That’s to be Expected. https://www.mihaileric.com/posts/mlops-is-a-mess/
  4. Chip Huyen. Dec 2020. Machine Learning Tools Landscape v2 (+84 new tools). https://huyenchip.com/2020/12/30/mlops-v2.html
  5. Matt Turck. Sept 2021. Red Hot: The 2021 Machine Learning, AI and Data (MAD) Landscape. https://mattturck.com/data2021/
  6. Casber Wang. Mar 2022. The Future of AI Infrastructure is Becoming Modular: Why Best-of-Breed MLOps Solutions are Taking Off & Top Players to Watch. https://medium.com/sapphire-ventures-perspectives/the-future-of-ai-infrastructure-is-becoming-modular-why-best-of-breed-mlops-solutions-are-taking-fd85c6ca8bcf

 

Additional reading

 

  1. Gartner. Oct 2021. Gartner Identifies the Top Strategic Technology Trends for 2022. https://www.gartner.com/en/newsroom/press-releases/2021-10-18-gartner-identifies-the-top-strategic-technology-trends-for-2022
  2. Lj Miranda. May 2021. Navigating the MLOps tooling landscape (Part 1: The Lifecycle). https://ljvmiranda921.github.io/notebook/2021/05/10/navigating-the-mlops-landscape/
  3. Innoq. MLOps Infrastructure Stack. https://ml-ops.org/content/state-of-mlops
  4. Rohit Tandon. Jun 2021. ML-Oops to MLOps. https://www2.deloitte.com/us/en/blog/deloitte-on-cloud-blog/2021/ml-oops-to-mlops.html
  5. MLOps.community. Nov 2020. The Current MLOps Landscape // Nathan Benaich & Timothy Chen // MLOps Meetup  #43. https://www.youtube.com/watch?v=i6HZ2vjFLIs
  6. Francois Chollet. Jun 2022. Discussion thread on Tensorflow and Jax. https://twitter.com/fchollet/status/1539411350681636864