Enterprise Machine Learning’s Biggest Challenge

By Srini Sekaran • October 3, 2019

Worldwide spending on artificial intelligence systems has surged to $35.8 billion in 2019, according to IDC. And for good reason: machine learning (ML) initiatives are no longer sidelined as peripheral innovation projects—they are at the epicenter of modern business, and critical for sustainability.

It is imperative that companies accelerate their adoption of machine learning; enterprise integration has profound implications that can change the way we ideate, build, sell, and support. It has the potential to dramatically increase the efficacy of a business and if companies are to maintain competitive advantage, the need for adoption and integration of machine learning within core workflows is a necessity.

While the need for adoption is here, reality is further away. A survey from O’Reilly indicated that merely 15% of 11,000 respondents work for companies that have extensive experience using ML in production. This survey is indicative of not a lack of effort or a sclerotic organization, but rather points to a series of barriers that dissuade enterprise adoption. According to a Dimensional Research report, 8 out of 10 organizations engaged in AI and machine learning reported that their projects have stalled. The biggest barrier: a dearth of accessible, high-fidelity training data.

96% of these organizations have run into problems with the quality of training data, a foundational aspect of training machine learning models. Along with the procurement of high-quality training data, aligning business value with ML initiatives is another roadblock. The two cardinal challenges with machine learning in enterprises are:

  1. Procuring an adequate volume of high-quality training data
  2. Finding alignment with business value and reducing risk

The Training Data Gap: A Need for Relevancy

Models that underpin ML initiatives can be imprecise, even catastrophic, without an adequate training mechanism. Without sufficient, relevant data, a detrimental ripple effect may occur, with predictions pointing in an inaccurate direction. The industry agrees. According to an executive specializing in labeling data for machine learning initiatives, “the single largest obstacle to implementing machine learning models into production is the volume and quality of the training data.”

A comprehensive data management solution, optimal for AI and ML initiatives, has attributes that allow for:

  • Automated, or policy-driven data ingest of new data sources
  • A streamlined way to manage multiple, disparate data sources
  • The ability to test and iterate models before deployment

Mass data fragmentation contributes to this dearth of high-quality training data. In the majority of organizations, data is massively siloed, copies are redundantly manifested, and data is scattered across different environments. If the internal data of an organization is to be leveraged as a competitive edge, it must be readily accessible.

Cohesity directly addresses mass data fragmentation by consolidating silos and making it simple to protect, manage, and leverage data. With one platform that spans on-premise and public cloud environments, and protects a true gamut of workloads, Cohesity offers a streamlined way to manage the numerous, disparate data sources in your organization. Cohesity addresses the training data gap by making it simple to clone data, or provision virtual copies. And since your sources are protected with policy-driven backups, the need to perform additional data ingest for machine learning training data is eliminated—you can do more with fewer workflows, and fewer copies of data.

Clones of databases are identical to physical database copies with the ability to perform all operations, but with an important caveat: clones created on Cohesity are zero-cost and have no storage overhead. With greater access to malleable data, data scientists and engineers are liberated to experiment, and have enough data to thwart overfitting.

In training models, quantity of data alone is not a panacea. Quality of training data, in terms of relevance and being in sync with source data, is critical. Whether using cross-validation, early stopping, regularization, or tweaking parameters, the ability to iterate training models with relevant—not stale—data can have a profound impact on model efficacy. As time is the cornerstone input for forecasting applications, the importance of having training data updated with source data cannot be overstated. With the ability to refresh clones on demand, Cohesity makes it simple to provide highly relevant training data by syncing training data with source data.

The premise of a data warehouse treads a line between promise and ambiguity. If done well, a data warehouse can open the doors to a ceaseless training data pipeline. A data warehouse should be viewed, McKinsey notes, “as a service model for delivering business value within the enterprise, not a technology outcome.” For machine learning development, our intent is clear: deliver business value by streamlining access to high-quality data.

Machine Learning: A Human Process

The steps in training—and leveraging—a model that underpins machine learning initiatives are:

  1. Defining the problem and gathering the data necessary for the model
  2. Defining model performance success and corollary metrics
  3. Preparing data through sanitization
  4. Selecting a model
  5. Training the model with the prepared dataset
  6. Evaluating the performance of the model based on established guidelines
  7. Refining parameters and develop a model that does better than the baseline
  8. Using the model to predict outcomes for new input data

Evidently, machine learning is still a human-centric process.

From defining the problem to tackle, to defining the success metrics that drive the iterable training process, different human stakeholders influence and orchestrate model creation and training. As a result, all stakeholders need to be empowered to drive the creation of an efficient model.

Cohesity satisfies different needs for a variety of stakeholders while keeping the intent clear: democratizing access to high-quality training data.

  • Satisfying financial goals. Create virtual copies of sources that are fully readable and writable, portable, and readily consumable for model iteration—while using 95% less storage.
  • Addressing security, head-on. Identify sensitive data and automatically apply data masking to sanitize personally identifiable information (PII). With vulnerability assessment, you can be confident in provisioning virtual copies throughout your organization.
  • Democratizing accelerated access to data. Rapidly provision virtual data copies to development, training, testing, and reporting environments.

Streamlining Access to High-Quality Data

The biggest challenge in enterprise adoption of machine learning is a lack of high-quality data and satisfying the needs of the cardinal members of the machine learning process—people. Cohesity makes machine learning simple by addressing the most time-consuming phase of training models: the procurement of high-quality data. To learn more, visit https://www.cohesity.com/solution/devtest/.

Tags: