We recently had the opportunity to brief Alan Pelz-Sharpe, the founder of Deep Analysis, on our newest offering, Nuxeo Insight, a service for AI in content management]. Alan was kind enough to write a very positive “Vignette Analysis” on Nuxeo Insight and I thought I would share some quotes from this Deep Analysis article as well as provide you with an update on where we are with artificial intelligence and what we have planned on our roadmap for this year.
First, here is the top-line summary from Alan’s report:
Nuxeo Insight has gone further than most in bridging the gap between highly trained data scientists and non-technical business users, making it one of the most effective and useful AI systems for content analysis that we have seen.
It doesn’t get much better than that. Well, actually, it does. Alan goes on to say, “Nuxeo is one of the more advanced offerings we have seen from a content management vendor…” As artificial intelligence is a particular point of focus for Deep Analysis and Alan has seen so many different offerings in our space, we are particularly pleased with his positive response to our new service. Further, he has really captured the essence of our offering, as one of our key goals is to put the power of fully trained machine-learning models in the hands of our customers, thus “bridging the gap” between data scientists and business users.
Nuxeo’s AI Strategy
In simple terms, Nuxeo uses established third-party technologies to undertake commonplace activities such as optical character recognition (OCR), speech to text, and some basic classifications. It layers on its own ML models to undertake more granular analysis, with the end result being specific customer “bots” to meet specific customer needs, trained on the customer’s data.
This is one of the best summaries of Nuxeo’s AI strategy that I have seen – a combination of common, commodity services and Nuxeo Insight, which enables customers to train their own, business-specific models with their own data and content.
Nuxeo has introduced a standard, high-performance enrichment framework that allows us to quickly and easily integrate commodity, third-party AI services. In fact, today, Nuxeo offers out-of-the-box integration with nine different cloud-based, AI offerings including:
- Google Vision
- Google Document Understanding
- Amazon Rekognition Video
- Amazon Comprehend
- Amazon Textract
- Amazon Transcribe
- Amazon Translate
Nuxeo Insight is our own AI for content service that leverages proven components like Amazon SageMaker to train custom ML models for Nuxeo customers. Insight and the custom models allows our customers to go further, and to generate better, more meaningful results. Our initial focus is on data enrichment and entity extraction – using Nuxeo Insight to add layers of metadata to both content and digital assets. Some early business use cases from various customers include product identification, IP identification, asset and content classification, and records identification. Deeper, more analytical use cases include things like fraud detection and intelligent customer servicing. The key thing here is that these are use cases which simply can’t be supported by generic models or services.
And, finally, as Alan has pointed out, these services can be combined. So, for example, a customer could leverage Amazon Comprehend for OCR or sentiment analysis, and then combine this with a custom model to enable intelligent, guided customer service experience.
Human Levels of Understanding
Nuxeo claims that in some cases, training times on site can be as short as six hours to deliver 85-90% accuracy rates. Clearly, this will vary depending on the circumstances and data quality.
A very simple goal of artificial intelligence is to understand content and its associated data as well as a knowledgeable human, but at scale. And this is what we deliver with Nuxeo Insight. As Alan states above, we can achieve incredible accuracy rates in a very short period of time. Typically, above 85%-90% accuracy is considered human equivalent. But what you really have to consider with AI is the short training times and the ability of the technology to operate at extreme scale.
For example, one Nuxeo customer has recently completed a POC with Nuxeo Insight. They are using our Insight services to do product identification, talent identification (which models appear in photos and images) and even cross-product identification. With a relatively small training set of less than 2,000 images with corresponding metadata, we were able to train an ML model that is delivering 95%+ accuracy. Using Insight, we were able to train this model in just a few short hours. And, you can throw thousands of photos and images at this model and get near-instantaneous results. Now, think about how long it would take to train a human to accurately identify all of these different products and talent and how long it would take a human to go through thousands of images to pick out specific products, talent, etc.
This is the real promise of AI and it will have an enormous impact on our industry and our customers. This also illustrates the power of Insight from a custom model perspective. No generic service, no commodity cloud offering could possibly identify this company’s products with any level of granularity or accuracy.
By the way, Alan is right. Your accuracy and results depend on the quality of the training set you use. This is one of the reasons why we are adding capabilities to Nuxeo Insight to help users best identify training sets to achieve the outcomes they want. More on this in a moment.
Another aspect of Nuxeo Insight is worth highlighting: its ability to identify which decisions are made by the AI and which are human-generated. Each content bot is versioned and saved, along with the associated models and data sets. This versioning process provides, in theory at least, a strong basis for a governance model.
We talk about Nuxeo Insight as delivering business-specific outcomes, but we also talk about Insight as enterprise-ready. On one level, this means delivering a cloud service that will scale and perform as our large, enterprise customers demand it to. On another, though, this is really about enabling proper governance for our AI services.
As Alan noted, we are able to distinguish machine-generated data from human-generated data. We not only version our ‘content bots,’ we also store our training data sets so that we can demonstrate how a particular model was trained and why it is producing the results it produces. Fundamentally, we want to provide a service that is fail proof and fully auditable. If a bot becomes corrupted, or begins to show bias, or if the performance of a particular model begins to decrease, we give you the ability to not only roll back to a previous version of the model, but also to roll back all of the data values the model produced. And, should a regulator or auditor ever inquire as to the source of data or how a model was trained, we give you the tools to provide an informed response.
Bridging the Gap
Lastly, I wanted to come back around to Alan’s comment about bridging the gap between data scientists and business users.
Though all of the machine learning and AI capabilities in Nuxeo Insight worked well in the demonstrations we observed, what caught our eye was the new UI (currently in a Beta release, with full release expected in Q1 2020). This is used to access and program the AI. The UI is wizard-driven and has been designed both to simplify the work of the data scientist and to be of use to business analysts. Like any such interface, there will be a learning curve, but the Nuxeo Insight UI is intuitive and cleanly designed. It is one of the best we have seen and bridges the gap between data science and business. The first release is impressive, and it will presumably improve even further over time.
First, we do offer various on-ramps and points of connectivity for organizations that are very savvy with AI/ML. Data scientists can build their own TensorFlow models and load them into Nuxeo. They can export their own datasets and train completely outside of the Insight environment. We don’t want to preclude our customers from using their own experience and expertise with artificial intelligence.
But what we also want to do – for our less AI-savvy customers – is provide tooling that makes it easy for them to both train and administer their own custom ML models. We felt that the best approach here is to employ a ‘point-and-click’ or wizard-driven experience to help a more casual user to configure and train a new model. We are also introducing a set of dashboards that will enable less-technical users, at a glance, to determine how their models are performing over time and to quickly and easily identify models that either aren’t working or are beginning to degrade. And, as I mentioned earlier, we even want to provide some intelligence to help guide less-technical users through the process of identifying the right training set to produce the results they are after. After all, an ML model can’t guess and it can’t deliver a value that it hasn’t been properly trained to produce.
Our new UI and user experience for Nuxeo Insight will begin to roll out in Q1’20 with further enhancements scheduled for Q2’20 and beyond. If you are interested about other enhancements we have planned for Nuxeo Insight, I will write another blog post about this later. But, for now, please refer to my previous blog post that talks about our product strategy and key deliverables for the first half 2020.
Hopefully, this has helped to give you some understanding of our AI strategy and how truly different Nuxeo Insight is from other offerings in the market. Again, I’d to thank Alan for his thoughtful analysis and positive comments about our Insight offering. And, if you’d like to learn more about the Nuxeo Platform or Nuxeo Insight, please don’t hesitate to schedule a demo or to reach out directly. I am always happy to answer questions about our products and services.