Margins like the share price are dropping right down

You might say this is just a one-off result,

But I'd say there is more competition in town,

So growth will be harder; it's not Appen's fault.

Here are some competitive threats taken from this article about Appen's business model.

-----

One trend to watch out for is the advent of projects like Snorkel, a Stanford University initiative started in just 2016, after Appen listed. It “set out to explore the radical idea that you could bring mathematical and systems structure to the messy and often entirely manual process of training data creation and management, starting by empowering users to programmatically label, build, and manage training data.”

Today, Snorkel, which explicitly seeks to reduce reliance on human contractor data labelling, is “used by many of the big names in the industry (Google, IBM, Intel),” according to this machine learning specialist. The 2019 version of Snorkel brought with it an improved capability to automate data labelling.

The point here is not that Appen is a bad company, but that the public narratives around Appen are inadequate to explain its recent results. We live in a world where market participants are variously incentivised not to dig too deeply into the weaknesses of the business models of high flying stocks. That is why you still see unquestioned claims that it is not possible to realistically automate the data labelling services that Appen provides.

And yet, if stock market researchers dared to delve beyond the confines of company presentations, they would see this is a problem that smart people are actively attacking. Appen is an expense for big tech; do you really think no-one will try to undercut them?

Data Labelling Can Be (Partially) Automated, At Appen’s Expense

The authors of this study state that “As machine learning models continue to increase in complexity, collecting large hand-labeled training sets has become one of the biggest roadblocks in practice. Instead, weaker forms of supervision that provide noisier but cheaper labels are often used…” Their goal is to improve the weaker, cheaper labelling techniques, and they conclude that “our approach leads to average gains of 20.2 points in accuracy over a traditional supervised approach, 6.8 points over a majority vote baseline, and 4.1 points over a previously proposed weak supervision method that models tasks separately.”

Oh, and by the way, those authors are the same people running Snorkel, you know, the data labelling tool used by Google and IBM.

But it gets worse than that for Appen. You see, when Facebook was just starting out with deep learning neural networks, every new project would have started from just about zero, in terms of labelled data. These days, “A procedure called ‘transfer learning’ takes a neural network trained on a vast data set and specializes it with a small supplement”, according to Alex Krizhevsky, one of the early proponents of neural networks.

Put simply, “Transfer learning allows us to deal with these scenarios by leveraging the already existing labeled data of some related task or domain”, and “Andrew Ng, chief scientist at Baidu and professor at Stanford, said during his widely popular NIPS 2016 tutorial that transfer learning will be — after supervised learning — the next driver of ML commercial success.”

Finally, we have the spectre of unsupervised learning. I’m not quite sure how this will impact need for fully supervised data labelling like Appen provides, but I’m sure it will reduce demand. Below, you can see Facebook’s CTO boasting about a recent advance in image segmentation, without data labelling. Clearly, here, the change is that Appen was not required, whereas its services (or a competitor’s) would previously have been needed to achieve the same result.

https://twitter.com/schrep/status/1388189398496202752