Recently I attended the NVIDIA GTC Developers’ conference and spent a little time learning more about the deep learning products. A branch of machine learning that often aims to make better representations and create models to learn these representations from large-scale unlabeled data. Many of the use cases have been associated with text and image search, facial recognition or medical diagnosis (identifying specific cancers, medical problems).
Deep learning is a very generic (in the broad-reaching sense) technology, companies like NVIDIA have developed the GPU technology and on top of them frameworks have been built to allow researchers and developers to rapidly construct products without actually developing bespoke image recognition technologies or similar. At GTC we had a bit of an employee hackathon and a group of our staff with no experience or machine-learning background built a solution to classify whale songs in an afternoon. Supplied with the data set from a kaggle competition and a framework it was apparently fairly straightforward (http://danielnouri.org/notes/2014/01/10/using-deep-learning-to-listen-for-whales/).
Coming from a CAD/CAE/PLM background the perfect application in my mind for this type of technology is part recognition and PDM (product Data Management). For every designer actually using Siemens NX or Dassault Solidworks/Catia there is a vast ecosystem of engineering users using and sharing the CAD data via PDM/PLM applications such as Enovia, Delmia, Teamcenter or cloud collaboration tools.
CAD is undoubtably getting cloudy – with onshape, outscale, fra.me and GrabCAD as well as numerous VDI users using VMware/Citrix etc. CAD though always carries a lot of legacy constraints and you have to deal with what you’ve got. The PDM/PLM and indeed BIM (in AEC) communities spend vast amounts of effort on part labelling conventions, search and part translation. Basically trying to find out what they’ve already got! (You only have to read the main PLM/PDM blogs such as Oleg Shilovitsky – see: http://beyondplm.com/2015/11/04/intelligent-part-numbers-might-be-a-good-idea-in-connected-world/ to see the PDM/PLM folk spend a lot of time and effort indexing and trying to find their parts!)
So how do you find stuff
There are a few 3D CAD search technologies around but strangely whilst everyone is talking about putting stuff in the cloud and sharing these technologies don’t seem to get much press.
Many products use some of the features and techniques outline in this seminal paper: https://www.uni-konstanz.de/mmsp/pubsys/publishedFiles/BuKeSa06.pdf; and use one or more of the techniques to make a very lightweight “fingerprint” of the CAD part and then look for sections of that fingerprint: same size, same hole pattern, material attributes (bits of metadata on part), sometimes you want same part but rescaled, could be re-oriented/affine transform etc ….
Companies/products I’ve come across in the 3D search space:
This isn’t a space I’ve followed particularly closely for a few years but I suspect it might finally be getting some end-user traction and visibility. As more and more folk put stuff in the cloud and datacenter the need to find it again over ever more disperse infrastructure grows!
- Geolus (Siemens PLM); Geolus must be a decade old or more. When I worked at Siemens they fascinated me and I never understood why this gem was never more promoted amongst the vast range of Siemens portfolio. I guess with so many technologies the exciting cooler ones that would be the VCs/start-ups sparkle are simply taken for granted? Mature and established in use in many vast real manufacturers that are Siemens bread-and-butter.
- Cadenas; This has always been a fascinating company, a German CAD parts catalogue rather than a modeler. Quietly for many years they have been developing simply very interesting technologies. They added some degree of modelling capability allowing a designer to define a simplified part against which to search. Partnered with more familiar names e.g. Delcam (now Autodesk) and won some serious acclaim from the manufacturing sector. Novel technologies being built on-top of a real established traditional CAD business, slowly and steadily.
- ShapeSpace – made a fair amount of impact in 2013 and got some good partners but seem to have gone quiet recently, a small UK software company I’m not quite sure what they are up to currently.
- Yowza; Where did they come from? Established in just 2015, this year they made a splash at Develop3D live, with a polished booth, talks and demos. Of all the companies I am familiar with in the 3D/CAD search.
There are huge financial drivers for decent CAD search and categorization:
- Opportunities for advertising around search services: https://www.catalogdatasolutions.com/3d-cad-search-engine.html
- Find all instances and duplicates of a screw and find an alternative 0.01c cheaper can lead to huge savings.
- Eliminate duplication in design and costly parallel varients
So where is Deep Learning?
Now this is where it got a bit disappointing. It seemed such a natural match of technology to problem and yet when I started googling I was surprised at the very limited material of the area. A few papers mainly behind pay-for (booo!!!) journal firewalls http://link.springer.com/article/10.1631%2Fjzus.C1300185
CAD / PLM does tend to be a closed industry. The bucks are big in PLM and many technologies are developed in-house and the expertise kept in house. Although many of the big players license their components and APIs (e.g. Geolus and Parasolid from Siemens), you won’t find a developer community or information on the technologies lying around the internet.
So will we see deep-learning in the CAD / BIM headlines soon? My guess is probably not, my feeling is that this type of technology will creep in behind the scenes.
Although with start-ups like Yowza popping up at end-user events like D3DLive! Perhaps there will be a little disruption and end-user interest in the overlooked area of CAD indexing/part retrieval.