Suppose you woke up one fine morning and decided to create an artificial intelligence. Where would you begin?
You could begin by reading up on all the relevant literature, the state of the art of what people have invented so far. That would be a reasonable approach. But suppose also that on this fine morning you were feeling especially confident and prideful and wanted to start from scratch. Where would you begin?
One natural place to look would be to natural intelligence. Pick some examples of intelligence in nature and try to reverse engineer them. Alternatively, you could look to the process that created all forms of intelligence in the first place: evolution. Without evolution, there would certainly be no intelligence. Following this line of thinking, you may even wonder: how could we possibly create a truly intelligent machine without evolution? It’s an important question. Many researchers have answered that we can’t, and that has led many of them to consider evolutionary algorithms as a supplement to deep learning.
But this line of thinking is missing a crucial fact, which is the fact that every piece of artificial intelligence that already exists—deep learning, expert systems, you name it—is already the result of an evolutionary process. Not an evolutionary process that is happening via algorithms but an evolutionary process that is happening via the tinkering of tens of thousands of engineers and scientists around the world.
This collective tinkering is a biological process, and thus we can think of the field of artificial intelligence—or more specifically, machine learning—in biological terms:
- The species are the different types of models (CNNs aka Machina lecunus, LSTMs aka Machina schmidhuberus, etc.)
- The symbionts are humans
- The ecological niches are problems and datasets
- There are forms of lateral gene transfer, for example, a new component or training process, such as ReLUs or batch norm, can be transferred to many species of models
- There is macroevolution and speciation events driven by drift (new niches, i.e. problem domains) and symbiosis (the combination of multiple species of models)
- The energy supply of the whole system is money/funding (from corporate and political entities)
Viewing the ML ecosystem in this way can tell us a lot about how the field operates. For one, we can see that, in some sense, AGI already exists. There are numerous species of models that collectively solve a very general set of problems. There isn’t one single metal-shelled, red-eyed robot that solves every problem, and humans always need to coexist as symbionts, but nevertheless, a meaningful form of AGI already exists.
Coming back to evolutionary algorithms, we can see that when EAs are used for deep learning (as in Evolutionary Strategies or Neural Architecture Search, for example), it’s as a tool for optimization rather than as a tool for innovation. Even when EAs are used to evolve intelligent agents (in some simulated environment), that is still not driving evolutionary innovation in the whole-field sense. In our biology analogy, it’s equivalent to the learning that happens within a single organism’s life, which is meaningful to the whole process of evolution insofar as its learnings gets passed on via cultural or epigenetic means (in the case of ML, this could be the agent providing information to its researcher, or perhaps passing on its knowledge to another agent via transfer or imitation learning).
More importantly for researchers and business leaders in the field, though, viewing the ML ecosystem in this way can help answer practical questions such as where do breakthroughs come from? and what does the future hold?
To try to answer the question of where breakthroughs come from, let’s talk about evolutionary fitness. There are roughly three kinds of “fitness” for ML models: (1) convenience of use, (2) efficiency/scalability, and (3) better performance. Every improvement to any ML model corresponds to an improvement in one or more of these dimensions. Practical applicability of a model could be considered another dimension, but that’s really a question of whether a niche (problem/dataset) has practical applicability, which is a separate question from whether models within a niche adapt to have higher fitness.
Next, let’s talk about evolutionary novelties. We can think of big breakthroughs in machine learning as “novelties”. There are at least three kinds of novelties: (1) transferable component improvements (e.g. dropout, batch norm, ReLUs, or even pytorch or keras, anything that can be transferred to many species of models), (2) new variants of existing models (i.e. speciation, e.g. LSTMs from existing RNNs), (3) the combining of multiple models and/or techniques (i.e. symbiosis, e.g. AlphaGo/AlphaZero).
The evolution of AlphaGo and AlphaZero, which is a great example of symbiosis, can be seen as akin to the evolution of lichen. Lichen, an incredibly successful organism—one that can thrive even in the most inhospitable of environments—is believed to be the result of two species of fungi combining with one species of cyanobacteria. “Fungi that discovered agriculture,” as one eminent botanist called them. Similarly, AlphaGo resulted from the combination of CNNs, supervised learning of human go games, value and policy reinforcement learning via self-play, tree search, and the advent of TPUs, which over time, became one single unified organism (AlphaZero).
Now whether this technology will be helpful or harmful for humanity is undecided (and completely up to people to decide). In other words, is this new species a parasite? If it’s used in parasitic ways, it will be. If it’s used merely to get rich, it will be. If, on the other hand, people are trying to make useful tools (ignoring, as much possible, all other incentives), and this technology can be utilized in a natural way to make those tools more useful, it won’t be.
But let’s take a step back now and return to our protagonist.
Suppose it’s a few weeks after that one fine morning, and you managed to invent an approach that is giving exciting results. You’re fairly confident that your approach is unique, and you’re wondering: if I turned this AI approach into a business, how large and profitable could this business become? Invention for the sake of invention is great, but cash is great too…
It’s impossible to answer this question for any particular business, but if we consider the dynamics and equilibria of the field as a whole, it may lend some insights.
Considering the field as a whole, the first thing we can say is that although ML has numerous real world, wealth-creating applications, the field as whole is most certainly not profitable. Instead, it operates like an arms race. Companies invest in ML because other companies are investing in ML, and if they don’t, what could that mean for their future? Countries invest in ML for the same reason. The US thinks, China is investing in ML. We better invest too, or else… The growth, in other words, is fueled by speculative competition and fear. It’s no wonder, then, that marketing masterminds who are heavily invested in AI, such as Elon Musk, stoke the flames of fearful speculation by reminding us constantly of the dangers of AI.
This arms race can be seen also by the fact that many of the biggest new ML businesses—such as C3 AI—are selling services for creating models, rather than selling the products of models themselves.
Of course, every new, exciting technology begins with more potential energy than kinetic energy. So the fact that the field of ML is a bubble is not surprising. Also, not all bubbles have to pop. They can reach a state of equilibrium. It would be a different equilibrium from one where R&D is funded purely by profits, and more fragile in the sense that the ecosystem’s energy (funding) comes from outside rather than inside. But in the end, the bubble will pop (and the ensuing AI winter will come) only when the field stops seeing growth, since only then will speculation flip.
So the questions—as with any bubble, organism, species, or ecological environment—are: can we keep the growth going? And for how long?