Category Archives: Learning

The Differentiable Organization

Neural networks aren't just at the fast-advancing forefront of AI research and applications, they are also a good metaphor for the structures of the organizations leveraging them.

DeepMind's description of their latest deep learning architecture, the Differentiable Neural Computer highlights one of the core properties of neural networks: they are differentiable systems to perform computations. Generalizing the mathematical definition, for a system to be differentiable implies that it's possible to work backwards quantitatively from its current behavior to figure out the changes that should be done to the system to improve it. Very roughly speaking — I'm ignoring most of the interesting details — that's a key component of how neural networks are usually trained, and part of how they can quickly learn to match or outperform humans in complex activities beginning from a completely random "program." Each training round provides not only a performance measurement, but also information about how to tweak the system so it'll perform better the next time.

Learning from errors and adjusting processes accordingly is also how organizations are supposed to work, through project postmortems, mission debriefings, and similar mechanisms. However, for the majority of traditional organizations this is in practice highly inefficient, when at all possible.

  • Most of the details of how they work aren't explicit, but encoded in the organizational culture, workflow, individual habits, etc.
  • They have at best a vague informal model — encoded in the often mutually contradictory experience and instincts of personnel — of how changes to those details will impact performance.
  • Because most of the "code" of the organization is encoded in documents, culture, training, the idiosincratic habits of key personnel, etc, they change only partially, slowly, and with far less control than implied in organizational improvement plans.

Taken together, these limitations — which are unavoidable in any system where operational control is left to humans — make learning organizations almost chimerical. Even after extensive data collection, without a quantitative model of how the details of its activities impact performance and a fast and effective way of changing them, learning remains a very difficult proposition.

By contrast, organizations that have automated low-level operational decisions and, most importantly, have implemented quick and automated feedback loops between their performance and their operational patterns, are, in a sense, the first truly learning organizations in history. As long as their operations are "differentiable" in the metaphorical sense of having even limited quantitative models allowing to work out in a backwards faction desirable changes from observed performance — you'll note that the kind of problems the most advanced organizations have chosen to tackle are usually of this kind, beginning in fact relatively long ago with automated manufacturing — then simply by continuing their activities, even if inefficiently at first, they will be improving quickly and relentlessly.

Compare this pattern with an organization where learning only happens in quarterly cycles of feedback, performed by humans with a necessarily incomplete, or at least heavily summarized, view of low-level operations and the impact on overall performance of each possible low-level change. Feedback delivered to humans that, with the best intentions and professionalism, will struggle to change individual and group behavior patterns that in any case will probably not be the ones with the most impact on downstream metrics.

It's the same structural difference observed between manually written software and trained and constantly re-trained neural networks; the former can perform better at first, but the latter's improvement rate is orders of magnitude higher, and sooner or later leaves them in the dust. The last few years in AI have shown the magnitude of this gap, with software routinely learning in hours or weeks from scratch to play games, identify images, and other complex tasks, going poor or absolutely null performance to, in some cases, surpassing human capabilities.

Structural analogies between organizations and technologies are always tempting and usually misleading, but I believe the underlying point is generic enough to apply: "non-differentiable" organizations aren't, and cannot be, learning organizations at the operational level, and sooner or later aren't competitive with other that set up from the beginning automation, information capture, and the appropriate, automated, feedback loops.

While the first two steps are at the core of "big data" organizational initiatives, the latter is still a somewhat unappreciated feature of the most effective organizations. Rare enough, for the moment, to be a competitive advantage.