AI has dramatically changed in the past few years as many AI models are available via low-cost APIs. This article discusses our learnings from the Applied AI Conference and how they can be helpful for your business.
We all know that AI has been a huge area of innovation for years but many of us have yet to use it professionally in a project. This has now dramatically changed in the past few years as many of the AI models are being made available via low-cost APIs. Gone are the days of having to stand up your own AI and then spend tremendous resources training and refining the models. Many modern models are simply pick up and go.
The Applied AI Conference (11/5/2022) covered the following topics: graph machine learning, sound event search, document search, utilizing assistive AI, computer vision, ML for agriculture, AI in manufacturing and extracting data from forms. You can find out more about the conference here.
The following are my notes about each session and what I thought was interesting or useful from an architect’s point-of-view. These may apply to your project or at least spark an idea!
Intro to Graph Machine Learning
- Reference Frames are the key to understanding human intelligence
- Embeddings is the way to add fast comparison (using cosine similarity) between nodes – helps calculate distance between nodes (knowledge)
- Use ML to build the Embeddings (ints, doubles, floats)
- ML can find distances between nodes in the graph to calculate the embedding, can also provided extra data to enhance the graph data (ala broken left/right arms are coded differently in a health care provider graph but they are very similar)
- Vespa is taking over from ElasticSearch for document search (uses semantics)
Sound Event Search
- Helps humans quickly scan large/long audio files for interesting sounds.
- The presentation compares distances between classification of sounds and then allows a user to find & rank other similar sound events in a file.
- You can pick a sound to ignore as well, and the system will no longer flag it as interesting.
- Very fun presentation – but I’m not quite sure of the implications and use cases yet in our customer base.
Lucy – Bringing Internal Data Resources to Life with AI
- Enterprise internal IT document search, enabling users to find details without having to bother subject matter experts.
- The tool Indexes video as well so a searcher can go right to place in the video where their topic is discussed.
- Provides analytics to the business about how people are using the tool.
AI is emerging from the playground
Role of AI in Telehealth during Pandemic
- Virtuwell’s AI helped answer questions about how to provide better customer satisfaction (that’s their key metric).
- Different forms of AI & ML were applied to help provide that better customer experience.
- For example, during an interview of a patient with a skin problem, it knows how to ask around 30 questions out of 5,000 possible questions to help the doctor provide the right diagnosis and treatment plan.
Computer Vision Panel
- Discussion on where computer vision is today, some of the benefits and pitfalls of using computer vision and what lies ahead (exciting stuff).
- Roboflow is like the Microsoft Computer Vision API Custom Vision
- Generative Art – DALL-E – words to art
- A great resource to explore on how it may enhance our customers existing technologies.
AI in Manufacturing
- Drift happens
- Transformer models are facilitating move to single/few shot training
- Much less training = less effort = more productivity
- Success requires planning…
- Team – both internal and external capabilities (usually capacity problem)
- Identify the … Where do I start?
Spatial Feature Matrix: Barrier for ML in Agriculture
- What agronomic problems lack solutions? I just like the phrase.
- Weather is the problem! Soil is a close second.
- Trying to solve spatial (soil) and temporal (weather) variability problems.
- Data definition & robustness is very important and standards need to be implemented.
Forms vs Natural Language: Why is it more challenging to process forms with AI
- Semantics needs to be combined with Geometry
- NLP = Semantics, Vision = Geometry
- Transformer AI models
- Pre-training then fine-tuning
- LayoutLM v3 is the model that implements NLP & Geometry
- Combine this with knowledge graphs – knowledge graphs can understand the values (by key) in the form. The knowledge graph also helps us throw away detected form values that aren’t important to us.
AI is coming out of the sandbox and into production
AI usage is exploding and easier than ever to use. You can see by my session notes that from agriculture to sound analysis and automated assistants, that AI is now becoming a very useful tool in your tool belt. Did a light bulb go off when you read over some of the use cases above? Drop us a line and we can show you how to make it a reality!