8 MLops predictions for enterprise machine learning in 2023

The MLops landscape is thriving, in a global market that was estimated to $612 million in 2021 and is expected to exceed $6 billion in 2028. However, it is also highly fragmented, with hundreds of MLops vendors competing for end-user artificial intelligence (AI) operational ecosystems.

MLops emerged less than a decade ago as a set of best practices to address one of the main obstacles preventing the enterprise from putting AI into action: the transition from development and training to production environments. This is essential because almost one in two AI pilots never get into production.

What trends will emerge in the MLops landscape in 2023? Several AI and ML experts shared their predictions with VentureBeat:

1. MLops goes beyond hype

“MLops will not just be a subject of hype, but rather a source of empowerment for data scientists to take machine learning models into production. The primary goal is to streamline the development process of machine learning solutions.

“As organizations push to promote best practices for producing AI, applying MLops to bridge the gap between machine learning and data engineering will seamlessly unify these functions. It will be vital in the evolving challenges of scaling AI systems. The companies that embrace it next year and accelerate this transition will reap the rewards.”

—Steve HarrisCEO of Mindtech

2. Data scientists prefer out-of-the-box industry-specific and domain-specific ML models

“In 2023, we will see a greater number of pre-built machine learning [ML] models that become available to data scientists. They encapsulate area expertise within an initial ML model, which then accelerates time-to-value and time-to-market for data scientists and their organizations. For example, these ready-to-use ML models help eliminate or reduce the amount of time data scientists have to spend retraining and refining models. Look at the work the Hugging face The AI ​​community is already driving a marketplace for ready-to-use ML models.

“What I expect next year and beyond is an increase in industry-specific and domain-specific pre-built ML models, enabling data scientists to work on more targeted problems using a well-defined set of underlying data and without having to spend time becoming a subject matter expert on an area that does not belong to the core of their organization.”

— Torsten Grabsdirector of product management, Snowflake

3. AI and ML workloads running in Kubernetes will overtake non-Kubernetes deployments

“AI and ML workloads are picking up steam, but the dominant projects are not currently on Kubernetes. We expect this to shift in 2023.

“There has been a tremendous amount of focus over the past year on customizing Kubernetes with new projects that make it more attractive to developers. These efforts have also focused on adapting Kubernetes offerings to run the compute-intensive needs of AI and ML on GPUs to maintain quality of service while hosted on Kubernetes.”

— Patrick McFadinVP of Developer Relations, DataStax

4. Operational efficiency will be a line item for 2023 ML budgets

“Investments focused on operational efficiency have been taking place for several years, but this will be a focus in 2023, especially as macroeconomic factors unfold and a limited talent pool remains. Those who advance their organization with machine learning (ML) and advanced technologies find the most success in designing workflows with the human-in-the-loop aspect. This approach provides much-needed guardrails if the technology crashes or needs additional oversight, while allowing both parties to work efficiently side-by-side.

“Expect some initial backlash and hesitation in educating the masses about the ML quality assurance process, largely due to a lack of understanding of how the learning systems work and the resulting rigor. One aspect that still raises doubt, but is an essential distinguishing feature between ML and the static, traditional technology we’ve come to know, is ML’s ability to learn and adapt over time. If we can better educate leaders on how to leverage the full value of ML – and the guiding hand to achieve operational efficiencies – we will see a lot of progress in the coming years.”

— Tony LeeCTO at Hyperscience

5. Prioritization of ML projects will focus on revenue and business value

“Looking at ongoing ML projects, teams will need to be much more efficient given the recent layoffs and look to automation to move projects forward. Other teams will need to develop more structure and set deadlines to ensure projects are completed effectively. Different business units will need to communicate more, improve collaboration and share knowledge so that these now smaller teams can act as one cohesive unit.

“In addition, teams will also need to prioritize what type of projects to work on in order to make the most impact in a short period of time. I see machine learning projects that can be grouped into two types: marketable features that management believes will increase sales and beat the competition, and revenue optimization projects that directly impact sales. Projects with salable features are likely to be delayed as they are difficult to execute quickly and instead the now smaller ML teams will focus more on revenue optimization as it can increase real revenue. Performance is critical for all business units right now and ML is not immune to it.”

— Gideon MendelsCEO and co-founder of MLops platform, Comet

6. Enterprise ML teams will become more data-centric than model-centric

“Enterprise ML teams are becoming more data-centric than model-centric. If the input data is wrong and if the labels are wrong, then the model itself will be wrong, leading to a higher rate of false positive or false negative predictions. What it means is that there is a lot more focus on making sure that clean and well labeled data is used for training.

“For example, if Spanish words are accidentally used to train a model expecting English words, you could be in for a surprise. This makes MLops even more important. Data quality and ML observability are emerging as key trends as teams try to manage data before training and verify model effectiveness after production.

— Ashish Kakranclient, Thomvest Ventures

7. Edge ML will grow as MLops teams expand to focus on end-to-end processes

“As the cloud continues to provide unparalleled resources and flexibility, more and more enterprises are realizing the real value of running ML at the edge — close to the source of the data where decision-making takes place. This happens for a variety of reasons, such as the need to reduce latency for autonomous devices, to reduce cloud ingestion and storage costs, or because of a lack of connectivity in remote locations where highly secure systems cannot be connected to the open internet.

“Because edge ML implementation is more than just putting some code into a device, edge ML will experience tremendous growth as MLops teams expand to focus on the entire end-to-end process.”

— Vid Jainfounder and CEO of Wallaroo AI

8. Feature engineering is automated and simplified

“Feature engineering, the process of understanding, categorizing and preparing input data in a way usable by machine learning models, is a particularly intriguing area.

“While data warehouses and streaming capabilities have simplified data ingestion and AutoML platforms have democratized model development, the required feature engineering in the middle of this process is still a largely manual challenge. It requires domain knowledge to extract context and meaning, data science to transform the data, and data engineering to deploy the ‘features’ in production models. We expect to see significant progress in automating and simplifying this process.”

— Rudina Seserifounder and managing partner of Glasswing Ventures