A recent MIT tech review article suggests that the era of deep learning is coming to an end. According to the article:
Our study of 25 years of artificial-intelligence research suggests the era of deep learning may come to an end.
Is this the case? We say…. eh. Probably not.
The review raises some solid points, but we’re not entirely sure we’d draw the same conclusions. Numbers are numbers, but as is often the case with deep learning, machine learning, and AI; the data itself is not necessarily significant; nor does it necessarily direct the conclusions being drawn from it.
So, where do we look?
What Is Deep Learning?
Deep learning is a broad concept. Like AI, ML, DevOps, and many other techno-buzzphrases, there’s not necessarily one universally recognized application or definition.
In and of itself, that makes defining whether or not deep learning is “coming to an end” difficult. What does that mean, exactly? Are we going to put an end to machine learning algorithms that focus on learning data representations as opposed to task-based or task-specific algorithms?
No, probably not.
So What’s Ending?
We don’t think deep learning is “ending.” We think we’re changing the way we perceive and apply it. Here’s what we mean.
The Concept is Solid
Deep learning as a concept (i.e. learning models as opposed to task-oriented models) isn’t going anywhere. The concept is solid and we’ve mostly figured out the core necessity of running these programs. What’s more, they work.
General Interest May be Waning, But That’s Expected
As with any “new” concept, it’s all the rage for a while. This is true of the general public’s interest as well as that of the skilled professional class which make use of the “shiny new object.”
A spike in interest, is therefore normal. If it’s waning, that could also be normal. Interest may be cyclical, but it often ebbs and flows for a reason. For example, if we apply a simple search on Google trends for “The Superbowl” after the past five years we get this:
Google Trends - Superbowl (2014-2019)
Not surprisingly. the interest spikes each year around, well, The Superbowl. Likewise, if we look at the trends for “Deep Learning” over the last 5 years, we see this:
Google Trends - Deep Learning (2014-2019)
We don’t think anyone would suggest that interest in American football or The Superbowl is coming to an end. Rather, interest follows a not unexpected pattern, i.e. people aren’t as interested in The Superbowl when The Superbowl is not being played.
Likewise, the trends in deep learning above are not unexpected either. It ramps up steadily over the last five years; with the largest spike seeming to be between 2016-2017. The trend appears to follow an up and down pattern while generally experiencing a steady incline overall. The last year or so seems relatively steady.
Does this mean deep learning is coming to an end (at least according to Google Trends?). We’d posit a different conclusion and that is that the steady increase is normal as interest and adoption grows. The plateau, or even a decline is normal as things regress to their more normal distribution.
Language May Be Evolving
As previously mentioned, certain techno-buzzwords go through several applications and iterations. Is Amazon’s “Alexa” really “artificial intelligence?” We’d say no. However, the label is frequently applied and there’s little we can do to prevent it.
To that end, how much is deep learning or AI “coming to an end” and how much is a question of labeling v.s. actual progress? The concepts that fueled the rise in interest and adoption are still there; they just may be called different things.
The Normal Growth Phase
We think that deep learning has simple gone through it’s first phase of growth. The innovators and early adopters have done their thing and we’re now looking at more widespread adoption and implementation.
And why shouldn’t we see this? It’s easier and less costly to experiment, implement, and adopt than ever before. On its way out? Sure, out of the early adopter phase.
To be fair, deep learning may not go significantly “deeper” than it is now in terms of the tech involved. However, it will most certainly cast a wider net into more industries and applications. This next phase will almost certainly be more about new adoption and implementation as opposed to technical innovation.
We Don’t Yet Know The Winner
The MIT research provides a good snapshot of where we’re at. Though we reach different conclusions from the data, it’s a good reminder of a few things.
- Tech is moving quickly. By leaps and bounds.
- It’s more about the number of categories that exist as opposed to the quantity of programs, methods, and algorithms.
- We don’t yet know what the “next big thing” is. In fact, there will always be a next big thing to be excited about.
- There is likely to be a “winner” in the realm of AI, ML. What is “the thing” (or several) that emerges from all of this that is truly innovative and new? The new process or application or algorithm? What new capabilities are we driving? And what does that mean for society and community as a whole?
The building blocks are in place for a unique combination of these emerging capabilities to work in concert to drive outcomes. And at the end of the day, isn’t that what we all want from technology? To drive real-world implementation and innovation?
Synergy is an oft used buzzword that is frequently defined as “the value of the whole is greater than the sum of their parts.” Whatever the next iteration of AI or ML happens to be which drives the market and brings about new capabilities; the cluster of tech which surrounds it will almost certainly create something which is far greater than the sum of its parts.