Reah Miyara predicts that AI fairness and bias problems will worsen before companies improve their governance and practices.
2
Enterprises will stop deploying models without understanding their behavior in production, which will push ML observability into regular use.
3
The ML infrastructure ecosystem will become more crowded and complex, while unstructured data, citizen data scientists, and model robustness gain attention.
Summary
Reah Miyara gives several predictions for MLOps in 2022 in conversation with Demetrios Brinkmann. He expects AI fairness and bias issues to get worse before they improve, partly because regulation and governance are difficult to define. Enterprises will also stop shipping AI blindly as production conditions change, making ML observability useful for monitoring behavior, finding root causes, and improving models. Miyara expects citizen data scientists to grow through low-code and no-code tools, although he thinks these users will initially consume model services rather than choose algorithms themselves. He predicts a more complex ML infrastructure ecosystem with more investment, new companies, and eventual consolidation. Unstructured data and embeddings will become more useful as teams work with text, images, video, and sensor data. Finally, model robustness will receive more attention because models can fail under environmental changes and adversarial manipulation. Brinkmann adds that regulation will generate plenty of discussion without much action.
AI fairness problems will worsen before companies improve them
Reah Miyara predicts that AI fairness and bias issues will get worse before they get better. He points to facial recognition, image classification, search rankings, and natural-language bots as examples of systems that can harm people. Models are built by humans and trained on historical data, which can leave blind spots for underprivileged and underrepresented groups. Miyara says companies should adopt more inclusive governance because it improves systems, rather than treating fairness as an ethics-washing exercise. He draws on his work helping develop responsible AI principles at Google.
Production models need observability instead of blind deployment
Miyara expects enterprises to stop shipping AI without understanding how models behave across the ML lifecycle. A model that worked on training data can behave differently during a global pandemic, a viral TikTok trend, or another change in human behavior and the surrounding environment. He says ordinary alerts are insufficient. ML observability should help teams monitor behavior proactively, identify the feature-value combinations behind failures, perform root-cause analysis, and decide whether to improve, retrain, or redeploy a model. He connects this prediction to his experience moving from controlled research settings into real production environments.
Citizen data scientists will use ML through simpler tools
Miyara predicts that citizen data scientists will become more visible as low-code and no-code tools let domain specialists apply ML without deep engineering or infrastructure skills. His examples include a risk manager working on outbreak prevention and a structural engineer predicting metal erosion. He expects the early version to involve services that let people upload a blueprint, select relevant features, and interpret a probability or performance chart. He would prefer these users to choose the model type and algorithm themselves, but he thinks that will come later. Boot camps, online courses, and fewer degree requirements are already expanding this group.
The ML infrastructure ecosystem will become more crowded
Miyara predicts that the ML infrastructure ecosystem will become much more complex. He compares it with software development, where tools such as GitHub and Datadog became established across parts of the stack over time. ML infrastructure is still in an early adoption phase, while companies and venture capital firms continue investing in startups and internal teams. Miyara expects the sector to become inflated and eventually consolidate, although Demetrios Brinkmann thinks the clear winners may not emerge until much later. They discuss acquisitions including Algorithmia and Determined, along with the steady arrival of new tools.
Unstructured data and embeddings will become more useful
Miyara expects more organizations to work with unstructured data such as text, mobile activity, social posts, sensor data, audio, video, and images. These data types are expensive and difficult to interpret, but they contain information about customer behavior, sentiment, and purchasing patterns. He identifies embeddings as a useful technique, describing them as dense vector representations of images, text, and video. He expects wider access to ML to increase demand for ways to represent and use this data. Demetrios Brinkmann connects the prediction to data-centric ML, where better labeling methods could reduce the amount of labeled data needed for some results.
Robustness will include resistance to environmental and adversarial change
Miyara defines model robustness as resistance to performance regressions under different conditions. A computer vision system might work in sunny weather and fail in rain or snow, so teams need to understand how changes in the environment affect its output. He also discusses adversarial manipulation, where someone changes an image through pixelation, dithering, or blurring to confuse a classifier. One example shown in the conversation is an image of a gun that a classifier identified as a camera. Miyara expects observability to surface these vulnerabilities faster and provide recommendations that improve model stability.
Regulation will be discussed more than it is implemented
Demetrios Brinkmann adds his own prediction at the end: there will be more discussion of laws, regulation, and governance for machine learning, but little will get done. He connects this to proposed European Union requirements around robust data sets and robust models, where the meaning of robustness is open to interpretation. Brinkmann also raises the ethics of data labeling services and the conditions under which labels are produced. Miyara agrees that laws and governance around fairness and robustness are difficult to define, although he expects model robustness to receive more attention during 2022.
