MLOps Before It Became Cool

Nebulus Blob of Intelligence // Karl Fezer

In this podcast, we drew philosophical tangents around the meaning of intelligence with Karl Fezer, AI Evangelist at Lockheed Martin, from his paper "All about defining intelligence"

The hype cycle of Artificial Intelligence (AI) and its discussion amongst people, initiated the movement to neatly define AI and Intelligence by extension. Taking inspiration from the work of scientists and philosophers over the years on cognition, helped them ground metrics for defining intelligence, based on different attributes.

AI Winters
With AI it is always a question of when and how because the constant goal is mimicry of human behavior.

In Alan Turing's time, the AI ideology was grounded on Intelligence by assumption or proxy. Its stand that intelligence was skewed to the beliefs of the observer and what they perceived as intelligent. It's not ideal because humans are not great observers of reality.

The second winter of AI in the 1960s theorized a bunch of techniques based on biomimicry of the human brain/cognition. Their theoretical advancement and scientific fiction laid the groundwork for the current state of AI today, irrespective of their limited computing resources.

In the current winter, data is exploding and quickly getting beyond our scope. The models are becoming too big to understand, which is why we are beginning to need explainability.

Video || Spotify || Apple

Designing Machine Learning Systems by Chip Huyen.

Using Machine Learning

ML is not the best solution to all problems, so consider whether it is necessary or cost-effective before embarking on an ML project.

Examining the core approach of ML solutions, which is to learn complex patterns from existing data and use these patterns to make predictions on unseen data, aids in understanding the implications of ML adoption for the problems that ML can solve.

ML is increasingly being used in both enterprise and consumer applications. Since the mid-2010s, there has been an explosion of applications that use ML to provide consumers with superior or previously impossible services.

Despite the fact that the market for consumer ML applications is expanding, the majority of ML use cases remain in the enterprise world. Enterprise machine learning applications have very different requirements and considerations than consumer applications.
There are many exceptions, but in most cases, enterprise applications may have stricter accuracy requirements while being more tolerant of latency requirements.

Read Here

MLOps at Ultraleap // Sam Jenkins

At the last Bistrol meetup, we heard from Sam Jenkins, Senior MLOps Engineer at Ultraleap. Ultraleap is a company that leverages ML to make human interaction with the digital world seamless interactive, and hands-on, using interactive interfaces like VR, AR, interactive screens, etc.

Generic MLOps
Generics is Ultraleap's custom MLOps platform, essentially a containerized system orchestrated with Docker. It has been the backbone for seven years for the development process of Ultraleap's solution. It has over 400,000 jobs across 40,000 experiments with 160 jobs or ML/data workflow (pipe types).

Ultraleap's MLOps path takes a turn to the edge, at the model serving phase. This introduces a unique set of problems contrary to what is commonly encountered in traditional MLOps.

The prediction performance, monitoring, and triggering path are chopped off, and the feature store is managed at runtime when inputting the data into training pipelines.

Generic's Feature X Pitfalls X Future
It provides experiment lineage pipelines, hardware and environment selection, visualization capabilities, declarative model architecture, container-based process, and a git-based system with forkable end-to-end pipelines.

As feature-packed as Generic is, it isn't easy to use for infrastructure monitoring, data management, and environment provisioning.

In the future, Generic plans to maintain its valuable features, utilize open source to add features easily, and well-supported tooling, where appropriate, in alignment with the modern tech stack.

YouTube

VMO2 MLOps Cookbook // Berk Mollamustafaoğlu

The MLOps Platform at VMO2 allows data scientists and analysts to explore data, iterate on ML-based solutions, and productionise them to make a real impact.

At the heart of this tooling sits Vertex AI Pipelines, which is Google Cloud Platform’s service that uses Kubeflow Pipelines (KFP) under the hood for creating end-to-end machine learning pipelines.

In this article, they share how they solve the problem of managing multiple container environments to allow for leaner and faster pipelines, enabling them scale the number of productionised ML products further.

On the MLOps Platform, each pipeline is made up of a series of Vertex AI Pipelines components, and each component consists of
Python code containing the component logic and a container environment for this component code to run in.

This blog was written By Berk Mollamustafaoğlu, Senior MLOps Engineer at Virgin Media O2.

On behalf of the VMO2 MLOps team: Alex Jones, Ariyo Oluwasanmi, Gleb Lukicov, and Sidney Radcliffe.

Read Here