This past year, the first robots capable of reproduction were created, ransomware attacks on critical utilities across California and Florida were thwarted, and a test to predict Alzheimer’s Disease prior to symptom onset was discovered.
What do these three feats have in common?
They were all made possible by machine learning models. Machine learning (ML) and artificial intelligence (AI) are at the heart of today’s major technological and scientific breakthroughs.
But AI and ML are not solely used for major discovery – they have become a staple of data-driven technology and ubiquitous in our daily lives. ML models show us which products we’ll enjoy from our favorite designer, protect our bank accounts from fraud, and determine which protein combinations will best treat diseases.
These models are also critical to the inner workings of modern companies. In fact, McKinsey reports that more than 56% of enterprises have adopted AI for one or more important use-cases, with the top three being service operation optimization, AI-based enhancement of products, and customer service automation.
Though the majority of companies understand the importance of AI and ML adoption, they often lack the right infrastructure, tools, and talent to get the most out of their models.
First and foremost, data is the fuel on which ML models feed. To derive value from a model, the data on which it is trained and run must be clean and properly formatted. Companies such as Scale, Labelbox, and Vody ensure quality model inputs through proper data labeling, management, and training.
Once a company has completed data prep, it must successfully deploy the model and support it in production. There are several companies in the newly minted “MLOps” space that facilitate faster model deployment, more efficient model scale, and reliable and unbiased model insights.
Bringing models into production is one of the most difficult steps of the ML life cycle – only 10-20% of models succeed. This step can tie up data science talent for months in an attempt to refine model performance. Machine learning acceleration platforms, such as OctoML and Deci use ML themselves to optimize model performance for hardware targets and reduce deployment time.
Once models are deployed, they must be run efficiently and scale effectively. AI optimization and orchestration company Run:ai virtualizes hardware resources to optimally allocate them across ML initiatives, saving customers time and money.
Even when delivered at speed, insights and recommendations are worthless if companies are unable to trust their accuracy. Zillow and Instacart experienced firsthand the pain and financial impact caused by machine learning model drift – when a model’s predictive power degrades due to changes in the environment. The former lost more than $500 million when its house buying model was not adjusted to reflect changing market conditions, while the latter faced significant struggles during the onset of the pandemic when its inventory prediction models dropped from 93% to 61% accuracy.
Beyond the problem of drift, models can also be inherently flawed due to biases in the data used to train them. Goldman Sachs came under scrutiny after models used to determine credit limits for its Apple Card offering were found to be gender biased.
To protect against model drift and bias, companies such as Fiddler, Arthur , and WhyLabs offer solutions for model monitoring and explainability. These solutions offer insight into the “black box” of machine learning by allowing a company to track model performance over time, understand the drivers of model recommendations, and address anomalies.
Join us to learn more about the Next Generation of Machine Learning at this years’ The Montgomery Summit 2022 presented by March Capital on May 24th and 25th, where we will hear from several of the companies mentioned above that are enabling enterprises to uplevel their ML efforts. To register for the 2022 Summit, please click the following link: https://www.montgomerysummit.com/montysummit_registration/. If you have any questions, please contact firstname.lastname@example.org.