"Collective Intelligence" Is Dead In The Water

Original post: here.

This post is commentary on the survey paper Collective Intelligence For Deep Learning - A Survey of Recent Developments by Ha and Yang from Google Brain.

Overall, the paper is a bit of a mess. As is almost everything that comes out of Google (these days).

Motivation

The motivation of the paper is premised in the abstract:

State-of-the-art deep learning models are known to suffer from issues that range from poor robustness, inability to adapt to novel task settings, to requiring rigid and inflexible configuration assumptions.

Essentially, deep learning models are rigid and this rigidity makes generalization difficult.

Robustness

In fact, it makes adversarial robustness less likely because the limited data and pre-defined (often randomly-chosen) architecture defines inflexible boundaries on the subspaces of the input space that are sufficiently covered or even "understandable" by a model.

Flexibility

What about model "flexibility"? Considering that all deep learning models are function approximations (or really, they are just functions) then they are always restricted to, well, the function that they are. The greatest restriction being the function signature that they are. This means input and output types and encodings are both chosen ahead of time as fundamental constraints. Thus, it is not possible for a model to be "flexible" by definition of what it is. A function will still be just that function regardless of how it is implemented (or in this case, according to the exactness of it's particular implementation only).

Novelty

If a model were to be constantly in fluctuation of what it was modeling (adaptable to arbitrary novelty) then it would not be a model of some particular thing, or at least it would not remain a model of the starting particular thing that it first modeled. No model can model everything at one time. Every model is bounded by the information that it can model to the approximate degree that it can model such information. Thus, any model is always bounded by the information being attempted to be modeled, and the parameters (capacity) it has available to model said information.

This holds true even taking into consideration "re-training" between novelty. A particular set of parameters arranged in a particular way is not guaranteed to be universally approximating. Put in terms of computer science, arbitrary parameterized architectures are not necessarily Universal computing devices.

Complex Systems

The idea is that concepts from complex systems could be potentially useful in adding degrees of freedom to new types of models (... which they will not, since they will also still just be functions).

The simplest example is training an ensemble of models and making predictions based on the averaged outputs may be the simplest form of incorporating CI into DL

Oh boy, here we go. "If we build a model, by composing other models, then gasp we have more complexity, and then like that, more complexity is equivalent to more flexibility and learning power". Off to the races, and here we are already setting out on the "Oh, I know! Let's increase complexity!" foot.

Concepts

Which complex systems concepts does this paper advocate for?

Concepts and ideas from complex systems, such as cellular automata, self-organization, emergent and collective behavior

Of course, they do.

Cellular Automata are the go to for "we have no idea what we are doing so we chose one of the most complex and not-well understood, but still highly-studied, computer science topics".

Self-organization has yet to be defined in a meaningfully useful way by anyone, anywhere, at anytime. And no, pointing to nature and saying "that is self-organizing", or pointing to a minecraft animation of blocks being positioned based on a running algorithm, are not "definitions" or even proof of "self-organization". Especially in the case of the minecraft animations it most certainly is nowhere close to SELF-organization. And the "self" part is a critically important differentiation to harp on. Critically important because in biology entities really and truly are self-propagating. Nothing in computer software is. And yes, that includes any silly renderings being claimed to be.

And every time someone starts using words like "emergence" and "collective" without even making a small attempt at formalization you should close the tab. And no, this post will not be attempting such a daunting task as well. If you have read this far then you can quit reading. The paper being discussed really does not save itself.

Applications

The paper then begins the surveying and focuses in on four areas:

We have identified four areas of deep learning that have started to incorporate ideas related to collective intelligence: (1) Image Processing, (2) Deep Reinforcement Learning, (3) Multi-agent Learning, and (4) Meta-Learning.

Meta-Leaning? What? Nevermind, I am afraid of the answer.

Image Processing

Image Processing is fairly straightforward: utilize Neural Cellular Automata to generate images, classify images, etc. It is really just applying a neural network to every pixel (or small locale of pixels).

One of the biggest advantages noted is the discretization of image processing that this buys you.

with the advantage of the approach being able to scale up to incredibly large image sizes, a challenge for traditional methods which are bounded by GPU memory

Having a model operate in a way that is not memory-bounded from the ground up seems fairly useful.

It should be of note the importance of this concept: every cell can have arbitrary complexity "internally" (e.g., a multi-thousand parameter neural network) and make decisions independently (to some varying degree). The importance is individual cells carry the burden instead of a grandiose global model.

This is a purely biologically-inspired concept that is not being given proper credit in this "collective intelligence CA" research.

What would be more profound is diversity of cells, and increased independence of cells.

But, the core problem is the cell itself. Meaningful and useful collectives of advanced enough cells is relatively straightforward iff you can solve the complexity of "the cell"ular unit itself.

The cell is like the recursive closure of the collective program. It's constraints define the emergent (yes, I am using this word) constraints of the collective.

This is why we have embryogenesis. This "collective intelligence" research feels more like artificial embryogenesis research. That would be a better usage of CAs. The obsession with "machine learning" is completely unnecessary for this type of research.

Deep Reinforcement Learning

I won't touch much on this part of the survey as it is pretty straightforward. In fact, it seems the most incrementally insightful to the research field today and I would encourage a review of this section if you are interested in ALife/ML research.

The research surveyed is not per say groundbreaking but it does begin to (more properly) compose biomimetic concepts.

As one example:

Recently, soft-bodied robots have even been combined with the neural CA approach discussed earlier to enable these robots to regenerate themselves

Which is referencing this paper. Which has usefulness in designing simple algorithmic methods for localized robotic repair. Of course, to implement something like this in physical mechanical systems would be dramatically more complicated. But, I could imagine that the Xenobots could potentially find geometric repair mechanisms? Who knows? Probably not, if we are to be entirely honest.

Multi-Agent Learning

This section does not seem to address any serious advancements in multi-agent systems in regards to collective intelligence. It merely mentions the attempts at addressing scalability of such massive amounts of complexity.

I will note that most work around multi-agent systems would be more applicable to Game Theory research.

Of course, it is possible that neurons in the brain actually engage in a competitive-cooperative game that generates a more general collective intelligence. Pure conjecture and I lean towards the "unlikely" side on that, at this time.

Meta-Learning

Firstly, what even is "meta-leaning"?

Meta-learning is an active area of research within deep learning where the goal is to train the system to learn

Well, that would be cool, but the paradox is that if we knew precisely everything about what it means to "learn" then can't we just build an AGI instead that can, well, learn?

Regardless, what kind of things do we have in mind?

modeling each neuron of a neural network as an individual reinforcement learning agent

Ah, so just stacking on more complexity to existing models and concepts?

Nothing fundamentally different then?

A similar direction has been taken by Kirsch et al. 35, where the neurons and synapses of a neural network are also generalized to higher dimension message-passing systems, but in their case each synapse is replaced by an recurrent neural network (RNN) with the same shared parameters.

It sounds like the only thing we are heading towards is rebuilding the internet (or any computer network) with the humans and all.

If we just want networks with complex nodes and advanced message-passing abilities then why not just study collective intelligence on the internet? Or why not find an algorithm to train the internet?

Oh wait, Google already attempts to do this.

Conclusion

In this survey, we ...

explained much of nothing. But, we had a good go.

The ole adage really holds for Google: garbage in, garbage out.