Neural Networks Around Us
15 Jun 2021 | 9 minutes to readWhen I think of a neural network, words like machine learning, AI, pytorch, big data, image identification come to my mind. I’ve previously always associated neural networks with technology trying to replicate the design of a brain as a way to solve problems that traditional computing has trouble with. For this post, I’ve tried to look beyond specifically-created neural networks and look for networks that exist on a greater scale. In the end, a neural network is just a graph of inter-connected nodes that each receive inputs, react to them, and then produce an output which can be used both inside and outside the network.
Social media is an example of a neural network, with individual people being the nodes and the posts and content they share is how those nodes communicate with each other. A social media site could be a like a brain, where ideas eventually come to some sort of consensus. Can analyze the activity of a single neuron in the brain, which is like monitoring a single person’s interaction with the social network.
When measured on a greater scale, individuals are clumped together based on “the algorithm” into groups. Groups generate and share content among members of the group, with only a portion of it being shared outside of the group. Groups are connected to other groups, etc, etc, etc, growing magnitudes of clustering.
The “thoughts” that fly around in this social network brain are based on the events experienced by the users. That’s a lot of input, so I feel that it would be difficult for this “brain” to form single, coherent thoughts without some large event experienced and posted about by many users.
It might be possible that it is capable of some sort of consciousness, but I don’t know if we would recognize it. Because it’s created through people’s interactions, it would “live” in a warped reality different from ours. How would it perceive it’s environment? How would it perceive? Maybe it would be a consciousness with no central thoughts
This “social media intelligence” is structured very differently than our own brains. For one thing, the “neuron” response time is much slower. Single posts can be amplified to thousands of other people very quickly, but it is still measured in the matter of minutes and hours. Everything seems to move at a fast paste, with people not dwelling too long on any single post.
It may take an even larger network than what is possible. Animal brains can have many billion neurons. But, number of neurons doesn’t seem like a very good measure of intelligence and a “user neuron” interacts with other users in a completely different format.
The input into the social network is not done through synapses and chemicals. Individual users create new content as input into the network. They also recycle or pass on other posts people have made by sharing, re-posting, cross-posting etc. As these are passed around, people might make changes or tweaks to it, morphing the idea as it is shared. Instead of having a few senses from which to draw information, this brain has millions of distinct inputs/repeaters. Each user is experiencing something different, and interact with the platform differently.
I imagine this could create creates a lot of “noise” in the system, though it depends on the organization of the social network. Site like Reddit have a very tight category system through subreddits, which helps control the flow of information. Users are connected to other users through the subreddit. User’s aren’t really connected to each other directly, unless through the “follow” system or sending messages.
Facebook has a more hierarchal approach. Users are connected to each other, and posts are shared based on the radius from the user: only me, family, friend, friend of friend, everyone. In addition to directly connecting users to each other, Facebook also has groups which congregate users into rough groups around ideas and topics which act similar to a subreddit.
I think because of how Facebook users are directly connected it experiences more noise than Reddit. The “neurons” of Reddit seem to be able to more easily focus on a single idea, at least across groups of subreddits. Focusing the interaction through subreddits helps direct the input a little better.
But maybe that would make it harder to be intelligent? The added randomness and inter-connectivity between users of Facebook could be closer to the “special sauce” of intelligence.
An example with an extreme amount of randomness could be Twitch chat. This network is different, in that all the viewers share the same stream. In this way, the “input” problem that Facebook and Reddit has is simplified and much easier to control. Users in chat interact with the streamer and each other. In this way, the streamer can ask chat questions and use chat to synthesize a response. Unlike a “ML” neural network, there isn’t really any tiers of nodes; all of twitch chat is a 1-d network, with inputs being the stream and other responses in chat.
The trick would be figuring out a way to determine what chat as a single entity is thinking. Maybe a word cloud where word/phrase size corresponds to frequency in chat?
The number of viewers determines how fast chat goes, and how much the streamer responds. As more people participate in chat, it makes it more difficult to read and follow a single person. The messages in chat also get more “reactionary”, where users are just posting emotes or short phrases. Instead of having more nuanced messages, they become much more copy-pasty and repeated, which might make it easier to combine chat into a single, morphing response to what’s happening on stream.
The big problem I can think of is how we are supposed to interface will would face in determining the intelligence of a social network is how to interface with it. How would you pose a question to the social media intelligence? Would you create a popup to quiz some or all the users about a question? You would need to figure out a way to format your input in a way that it could be consumed naturally by the users. Maybe a single post, or a variety of posts tailored to the audience in some way. Social network intelligences may just experiences things as more general “feelings” over much longer timescales than we’re expecting.
Maybe people settle into dense groups with a few connections to other dense groups, mimicking how the brain has different “clusters” of neurons to perform specific tasks (sight, sound, memory etc.). These sub-groups of social media users could be used to types of groups’ reactions to external stimuli, wether it be events the users experience on their own or through the platform.
Nature has been around much longer than we have, and has experimented with innumerable ways of “being”, from single individuals to large groups that congregate and work towards a common goal as one. Animals definitely communicate with each other, forming groups working together to survive. Plants have also been known to communicate with each other through methods like releasing volatile organic compounds into the air or shared mycelium networks underground 1. They all communicate in ways that are very different than our own methods and it makes it difficult to find a shared method of communication. We don’t know how to ask questions, but maybe we need to figure out what to ask first.
A parallel to figuring out how to communicate with non-human entities could be slime mold path finding 2. There have been experiments testing how slime molds solve simple mazes and recording the method it used to solve the maze. In this case, we are posing a “shortest path” problem to the slime mold. It’s not a big jump to also apply this type and format of question to other insects like ants or bees.
Maybe it’s better to consider what questions the networks would be asking themselves, and try and determine how they are answering it. Then we don’t have to worry about how to ask a question, or what to ask. We only need to focus on ways to measure the response. For example, a question that would matter to a bee hive could be “What will the weather be like tomorrow?”. For the average flower bee it’s much more difficult to gather pollen in the rain than in fair weather. If the bees are able to sense the weather, then we would expect them to change their behavior depending on what they’re sensing. This change in behavior is something we can measure!
In 2015, researchers in China attached RFID chips to 300 worker bees from 3 healthy colonies (~6,000 bees total) and monitored their exit and entry from the hive 3. They also recorded the weather for each day, noting the temperature, precipitation, wind, and sunrise/sunset times. Armed with all this data, they found that the bees worked harder to gather more on days before rain. The pollen collectors would spend more time foraging, and would forage later in the day, when the following day was rainy (>5mm of precipitation). If enough hives were scattered throughout a geographic region, you could create a rudimentary “bee weather network” to help predict upcoming weather conditions. While it wouldn’t be as accurate as our own satellite weather predictions, it shows that we can glean useful information from networks even when we don’t have a way to directly pose questions.
There a networks of entities all around us reacting to stimuli on large and small scales. If we can find ways to measure their reactions to stimuli in a meaningful way, we can potentially reach useful conclusions. With technology, it has become easier and easier to quantify the world around us. It can be difficult to organize, and process it in such a way to get a valid, understandable result. The key is having a good understanding of the type of data collects, and how different types inter-relate to each other. The better we understand the input data, the better we can tune the algorithms and tools we use to try and make sense of the data.
Polish Rating: 4
