Вы находитесь на странице: 1из 5

1

The Automation of Automating Automation: Automation in the AI Industry


James Steinhoff
The University of Western Ontario
jsteinh@uwo.ca
Presented at Canadian Communication Association Annual Conference, University of British
Columbia, June 2-6, 2019.

The contemporary form of artificial intelligence called machine learning (ML) is driving a new
wave of workplace automation. Two centuries ago, Karl Marx showed that automation is an
expression of the integral dynamics of capital, which is compelled to replace human labour with
machines. His analysis was prescient as contemporary technologies such as machine learning
enable the automation of cognitive, as well as manual, labour processes.
Today I’m going to talk about how the high skill data work that goes into producing machine
learning tech is itself being automated – through the recursive application of machine learning to
itself – a technique known as AutoML. I learned about this emerging kind of automation during
interviews I conducted for my dissertation, in 2017 and 2018, with data scientists and engineers
working in the commercial production of machine learning products. There’s a lot that could be
said about this, but I’m going to try to make only three points.
First, I’m going to talk about why machine learning is a different type of automation from
previous forms. Then I will show how the labour of producing machine learning tech is being
automated and briefly discuss the relations of workers towards it. I will argue that AutoML
complicates critical theorizations of automation from two Marxian schools of thought: labour
process theory and post-operaismo. Finally, I will suggest that AutoML indicates the need for a
conceptual shift from automation to autonomization.
1
I don’t have enough time to adequately discuss it, but it’s important to note that classical AI is an
automation technology. In the late 1980s and 90s, the first form of the AI industry was built
around classical AI “expert systems” in which the knowledge of workers was encoded in the
form of logical rules by “knowledge engineers”. However, it proved to be an impossible task to
manually code the knowledge necessary for dealing with complex situations, so AI producers
recognized a need to make machines learn on their own.
The machine learning approach to AI has existed since the earliest days of classical AI but it was
widely commercialized only in the 2010s. Machine learning is a novel approach to automation.
In traditional programming and classical AI, you build a program to solve a problem. This
requires knowing precisely how to solve that problem or obtaining that knowledge from
someone who does. With machine learning, you build a model which, if it is properly trained on
good data, “learns” a solution to that problem itself. The canonical example of this is a spam
filter: rather than attempting to write a program with a rigorous definition of spam, you can feed
a machine learning model many examples of spam and non-spam, and eventually, it will
recognize patterns by which the two can be distinguished. So rather than automating a given
2

process, with machine learning we have the automation of the process of automating a process: it
has thus been dubbed “automating automation” (Raschka 2016) and the “inverse of
programming” (Domingos 2015, 7).
However, producing this technology still requires plenty of human labour. The workers in this
field are highly remunerated; a data scientist with only a Masters’ degree averages nearly one
hundred thousand USD/year within the first three years out of school and PhDs pull in even more
(Press 2018). This is no doubt due to the very limited number of them and a consistently growing
demand, as more and more companies begin to incorporate AI into their operations. But it is also
due to the rather occult nature of machine learning work. One software engineer I interviewed
called it a “black art” because of how mysterious and intuitive much of this work remains. There
is of yet no “one best way” to do a lot of it and therefore it requires a lot of experience,
experimentation and creativity – in other words, this labour process has yet to be Taylorized. As
one software engineer I interviewed put it: “we know how to make [machine learning] really,
really bad. We know what not to do ... we can tweak it a little bit. But we don’t actually have a
firm … theory behind it”.
2
Despite this lack (or because of it), the labour of producing machine learning is itself being
automated – with machine learning. This is called AutoML or (it’s a mouthful) “the automation
of automating automation” (Mayo 2016). To see how this is taking place we can look at the three
stage machine learning labour process, where AutoML is being applied to at least two stages:
The first stage is data processing, which includes the laborious work of gathering, cleaning and
formatting data, as well as “feature engineering,” which is the essential task of deciding which
bits of data are important to the problem you are trying to solve. It requires considerable
knowledge about the domain the data represents as well as intuition and a lot of trial and error. It
is being automated via feature learning networks which can automatically pull out the most
relevant features (Hardesty 2015; Koehrsehn 2018). This technique has produced systems which
perform better than ones designed by human experts.
The second stage involves making the machine learning model. It includes choosing a learning
algorithm, the architecture of the model and parameters that govern its behavior. It’s also being
automated, in at least two ways. The first is model design. Designing machine learning models
also requires intuition and experience. It’s often not possible to say in advance what type of
model will work best for an unfamiliar problem (Zoph and Le 2017). Despite this, model design
is being automated via “neural architecture search” which uses machine learning to design and
test thousands of options in the time humans could try a few (Elksen, Metzen and Hutter 2019).
The second is hyperparameter optimization or “configuring the internal settings that govern the
behavior of a [machine learning] model”. It is being automated in a way similar to neural
architecture search, via the automatic design and testing of configurations at inhuman speeds (Li
and Talkwalkar 2018).
These three AutoML techniques are being applied to aspects of machine learning work which
have been resistant to Taylorist codification due to a lack of knowledge which not even expert
3

workers possess. AutoML overcomes or sidesteps that lack of knowledge through brute force
techniques of experimentation and learning.
Now to be clear, I’m not suggesting that the automation of software work is happening for the
first time here. Wendy Chun has demonstrated how automation really goes back to the very first
use of the stored program and is inherent in all higher level programming languages. However, I
think there is something distinctive about the type of automation AutoML enables.
Classical AI is automation. It eliminates the worker by capturing her knowledge and skills about
how to solve a problem (via interviews or time-motion studies). Machine learning is the
automation of automation. It eliminates the need for a worker who first knows the solution to a
problem. It extracts the solution from data automatically without it first existing in the worker’s
head. But, it requires skilled machine learning workers.
AutoML is the automation of automating automation. It eliminates not only the need to know the
solution to a problem in advance, but also the need for knowledge of the “black art” of machine
learning. It replaces aspects of high skill machine learning work with brute force
experimentation. AutoML thus represents for capital the possibility of dispensing with labour in
two senses: in terms of both conceiving of the problem and its solution, as well as understanding
and creating the tools by which the solution might be produced automatically. Ideally, AutoML
dispenses with both the need for someone to know how to solve the problem as well as for a
skilled creator of an automatic problem solver.
Since I conducted the interviews, AutoML has gone from an interesting idea to a commodity.
Today most AutoML techniques consume huge amounts of computational power and are
therefore available only to huge tech firms and research labs, but new methods are already
making them accessible via cloud platforms like Amazon Web Services. There is now talk of
“deep automation” in machine learning (Lorica and Loukides 2018) and the tech giants are
working on “end-to-end” or “one-click” AutoML packages which aim to streamline the
variegated components of the machine learning labour process into one user-friendly app which
will require very little technical knowledge. Amazon, Microsoft, Google and others are
describing moves like this as the “democratization” of AI.
We might expect machine learning workers to be worried about this apparent deskilling of their
work. But not all are, at least according to my interviews. While some workers expected that
their work would eventually be automated, most were unconcerned, and some even expressed a
desire for increased automation. They desire it because it helps them dispense with what one
interviewee termed “busywork” and another “drudgery”. Importantly, it is not management
implanting AutoML – it is the workers themselves. One data scientist said:
if you keep doing the same thing for six months you have to start to think of
automating … And if you keep doing the same routine job for a year, you’re
probably doing it wrong … We should always move towards automating.
3
4

So with AutoML we have workers automating their own work. What is significant about that?
Well, according to post-operaismo theorists such as Michael Hardt, Antonio Negri and Paulo
Virno, in the age of “immaterial labour” automation can be turned against capital. Post-
operaismo theorists argue that the proliferation of information technologies gives labour new
capacities for autonomy from capital. They are not overly concerned about the threat of
automation because they believe that immaterial labour, being comprised of the communicative
capacities of networked humans, “cannot ever be set within fixed capital” (Virno 2004, 106).
Thanks to the essentially human core of labour, automation holds promise rather than peril.
Thus, left accelerationists like Srnicek and Williams (2015) make full automation a primary goal
of their political program.
I suggest that post-operaismo’s optimistic take on automation is complicated by AutoML. While
machine learning workers are indeed capable of automating their own work, it is questionable
whether they exhibit the autonomy posited by theories of immaterial labour since their work
takes place within capitalist firms and the products they produce remain commodities. Not to
mention that the production of machine learning systems requires immensely powerful hardware
which is also owned by capitalist firms.
More significantly, if machine learning work, which arguably represents the epitome of
“immaterial labour”, is already being automated, the question remains open whether there are in
fact human capacities which will remain impossible to automate. AutoML’s new form of
automation dispenses with the Taylorist process of knowledge and skill capture elucidated by
Harry Braverman. Braverman described automation as the “progressive elimination of the
control functions of the worker… and their transfer to a device which is controlled … by
management” (1998, 146).
Instead of a transfer of knowledge from worker to management, with AutoML we have a move
towards the autonomous generation of skills and knowledge by machines. In other words, a
continuation of the change begun by AutoML could require a conceptual shift from automation
to autonomization. But this would be the autonomization of capital, not labour.
If this is the case, then automation will require a new Marxian orientation. Rather than doubling
down on an optimistic appraisal of the irreducible human component to capital, as post-
operaismo does, we should take seriously the possibility of capital achieving an increasingly
machinic form. It might be the case that, as Nick Land puts it: “Capital only retains
anthropomorphic characteristics as a symptom of underdevelopment” (2012, 445-446).
What does a critical orientation to machines under capital look like if a shift from automation to
autonomization is indeed underway?

References
Braverman, Harry. Labor and monopoly capital: The degradation of work in the twentieth
century. NYU Press, 1998.
5

Dettmers, Tim. (2015) “Deep Learning in a Nutshell: Core Concepts”. Nvidia Developer Blog.
https://devblogs.nvidia.com/deep-learning-nutshell-core-concepts/
Domingos, Pedro. The master algorithm: How the quest for the ultimate learning machine will
remake our world. Basic Books, 2015.

Dong, Catherine. (2017) “The evolution of machine learning”. TechCrunch.


https://techcrunch.com/2017/08/08/the-evolution-of-machine-learning/

Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. "Neural architecture search: A survey."
Journal of Machine Learning Research 20, no. 55 (2019): 1-21.

Hardesty, Larry. 2015. “Automating big-data analysis”. MIT News. October 16.
https://news.mit.edu/2015/automating-big-data-analysis-101

Hardt, Michael and Antonio Negri. 2000. Empire.

Hardt, Michael and Antonio Negri. 2017. Assembly.

Koehrsen, William. 2018. “Why Automated Feature Learning Will Change the Way you do
Machine Learning”. KDNuggets.com. https://www.kdnuggets.com/2018/08/automated-feature-
engineering-will-change-machine-learning.html
Li, Liam and Ameet TalkWalkar. 2018. “What is neural architecture search?” O’Reilly.
https://www.oreilly.com/ideas/what-is-neural-architecture-search
Lorica, Ben and Mike Loukides. 2018. “Deep automation in machine learning”. O’Reilly.
https://www.oreilly.com/ideas/deep-automation-in-machine-learning
Mayo, Matthew. (2016). “The Case for Machine Learning in Business.” LinkedIn. Accessed 20
November 2018. https://www.linkedin.com/pulse/case-machine-learning-business-matthew-
mayo/
Press, Gil. 2018. “The salaries of data scientists remain steady but still sexy with AI on the
horizon”. Forbes. May 4. https://www.forbes.com/sites/gilpress/2018/05/04/the-salaries-of-data-
scientists-remain-steady-but-still-sexy-with-ai-on-the-horizon/#a162b69733df
Raschka, Sebastian. 2016. “How to Explain Machine Learning to a Software Engineer”.
KDnuggets. https://www.kdnuggets.com/2016/05/explain-machine-learning-software-
engineer.html
Srnicek, Nick, and Alex Williams. Inventing the future: Postcapitalism and a world without
work. Verso Books, 2015.
Virno, Paolo. 2004. A Grammar of the Multitude. London: Semiotext (e).
Zoph, Barret and Quoc V. Le. 2017. “Neural Architecture Search with Reinforcement Learning”.
Under review as conference paper at ICLR 2017. https://arxiv.org/pdf/1611.01578.pdf

Вам также может понравиться