Given the massive amounts of hype and promise surrounding artificial intelligence (AI) and related technologies, it’s become increasingly difficult to make critical innovation and investment decisions in the space. In our recent blog, Navigating the AI Hype: An Outcome-Focused Framework, we presented a framework that reduces AI failures by starting with the end in mind. This outcome-focused framework considers the capabilities of today’s AI tools, emerging solutions, and challenges surrounding AI implementation. We outlined AI’s hype, failures, and the framework in our previous post. In this blog, we will share our #LuxTake on the challenges and emerging solutions in AI.
As mentioned in our previous post, at a high level, any AI project has four major steps; problem selection, data preparation, model selection and training, and deployment (see below). In this post, we will focus on the latter three steps.
Once a problem has been identified or selected, the next step is to prepare relevant data sets. Key challenges presented during this step surround data collection, data accessibility, and data cleanliness. One emerging solution for these challenges is to use machine learning-based data cleaning and wrangling tools to help automate the tedious and time-consuming process. These tools can automatically identify “dirty” data and remove or correct it. Investing in data cleanliness is often overlooked when building AI projects. Those interested should engage with vendors applying automated data cleaning and wrangling approaches.
MODEL SELECTION AND TRAINING
At the heart of any AI project are model selection and training. Model selection involves testing and selecting the best performing model, while training involves the process of adjusting parameters to fit a model to a dataset and optimizing its performance. Key challenges when it comes to selecting and training the right model are small datasets, privacy concerns, lack of data science talent, processing limits, and bias.
Emerging solutions like transfer learning are already making an impact on computer vision and natural language processing problems. Transfer learning allows for the knowledge learned in one task to be reused a starting point for a second task, thus requiring less data. For small datasets, this approach is a great solution as it is already seeing adoption by machine learning engineers. Another way to face the challenge of small datasets is with synthetic data. Synthetic data and simulations hold promise to overcome small dataset challenges in complex environments.
When it comes to combatting bias, dealing with privacy constrained applications, and the lack of data science talent, the tools are rapidly developing, but remain early stage and difficult to deploy at scale. For example, automated machine learning solutions, which seek to replicate the work for a data scientist, are emerging with claims to enable the democratization of AI. However, these tools tend to be overhyped. Those interested in addressing these challenges should monitor these solutions for improvements in performance and consider testing them in proof of concept projects when appropriate.
Deployment of AI systems involves exposing a trained machine learning model to new data for inferencing on an ongoing basis. This brings a whole new set of challenges. As deployment crosses the line from creating a prediction in a development environment to running models at scale it creates new issues such as the need to explain results, interpretability, edge computing requirements, and cybersecurity.
Unfortunately, in the near term, interpretability will have to be sacrificed for performance or vice versa. While several organizations are working on explainable AI tools that can aid in explaining complex models, it is unclear if the explanations produced will be useful or actionable. The challenge of cybersecurity, however, has solutions ready to be engaged with. Despite the early-stage nature of some adversarial robustness tools, those interested in deploying machine learning, especially in customer-facing or public applications should explore these toolsets in addition to understanding the overall cybersecurity landscape.
New AI processors and edge computing software will unlock a variety of new applications. Like AI processors for training, companies are designing dedicated edge processors for machine learning workloads which should be available in the market within the next few years. Everything from wearables to autonomous vehicles will require vast amounts of edge computing at extremely high processing efficiencies. Many new applications and opportunities will emerge from the arrival of edge computing and both software and hardware approaches should be explored.
Focusing on the outcomes and capabilities of today’s tools will help minimize the risk of AI failures. Rather than focusing on technologies first, start with the end in mind. Then, work backward to determine the level of AI needed to solve the problem and whether it’s feasible with today’s tools. Following this step, identify the key challenges, as outlined above, and determine whether emerging solutions will be capable of mitigating those challenges. Emerging tools can solve some challenges in implementing AI; however, technology readiness varies greatly.
- Navigating the AI Hype: An Outcome-Focused Framework
- Artificial Intelligence: A Framework to Identify Challenges and Guide Successful Outcomes (Members Only)
- Recent Webinar: Sensing for Insight: Getting the Most From Commercially Available Sensors
- A Taxonomy for Machine Learning Algorithms (Members Only)
- Five Challenges for Deep Learning (Members Only)