A new method to query the inner workings of neural networks promises to increase our trust in them.

If we could understand, exactly how customers think we’d be able to predict their behavior and we’d then be able to cater to what they need and what they want, in advance. That’d be an amazing business model, one that Google co-founder, Larry Page, described when he was detailing his own vision of the perfect search engine.

We don’t yet have that perfect search engine and neither do we have businesses that can understand their customers that well. Even the neural nets we create to generate unguided real-world behavior (like self-driving cars, large language models such as ChatGPT, machine learning modules in Google search and certain diagnostic tools in various other industries) are not transparent to us so we cannot really tell why they come up with the results they do even though we understand the mathematics that drive them. 

The reductive approach to understanding the world is to break everything down into its component parts and then seek to understand the individual contribution of each component to the whole. This works in some cases but the moment we get to dynamic, multi-part systems the process fails. Quantum mechanics tells us that the whole is, indeed, greater than the sum of its parts and this also stands when it comes to examining neural networks. 

The smallest part of them we can examine, the neuron, just like neurons in the human brain, has a poly-semantic aspect to it that makes it impossible to predict its contribution to a specific situation and, therefore, the computational outcome that derives from it before it is produced. 

There Is More Than One Dimension To Every Neuron

Polysemantic is the attribute ascribed to neurons that respond to a mixture of unrelated inputs making it hard to predict what they’re actually processing and therefore what they are computing. This apparent complexity happens when there is a need to describe external input values that are significantly greater than the number of neurons available. In the human brain, for instance, this is called a “continuous semantic space” and it is a space where the representations of objects and action categories is orders of magnitude greater than the number of neurons found there. 

Rather than use the inefficiency of a single neuron representing a particular object or state that becomes evident on the activation of that neuron, (monosemanticism) human brains and, evidently, neural nets that are trained on massive amounts of data; employ activation patterns which in turn represent an object or an action category. 

This is where it gets interesting.     

Those activation patterns are both discrete and finite. That is to say they can be identified and cataloged and, also, be expected to appear in other neural nets that encounter the same or similar representation even if their particular architecture is somewhat different. Think of it a little like the neural signature of an emotion like joy or sorrow. While the circumstances that will engender each of these emotional states in each person will vary according to each person’s memories, experiences and knowledge which shape their perception, the emotion will have an identifiable, universal neural pattern signature that makes it easy to recognize. 

The neural excitation pattern signature that represents each emotional state in my example then has only one, reliable, interpretation in each case for each person. That’s the transition from polysemantic neurons to monosemantic neural states. 

Identifying Meaning In Neural Network Activation Patterns

Artificial intelligence startup Anthropic PBC has come up with just that way of isolating and identifying what goes on in a neural network. As they state in their blog on the subject: 

“In a transformer language model, we decompose a layer with 512 neurons into more than 4000 features which separately represent things like DNA sequences, legal language, HTTP requests, Hebrew text, nutrition statements, and much, much more. Most of these model properties are invisible when looking at the activations of individual neurons in isolation.”

The approach offers real hope for making meaningful progress in Mechanistic Interpretability, the capability, to better understand the internal work of neural networks of ever increasing complexity and achieve a satisfactory amount of trust in their operation. 

Anthropic have published a paper on the subject that makes for really interesting reading not least because it says: “If we view each feature as a vector over the neurons, then the set of features form an overcomplete linear basis for the activations of the network neurons.” Vectors are used in semantic search both as word representations (i.e. word vectors) and in concept attributes that define the data itself. 

The fact that we are using them to better understand the inner working of neural networks is of critical importance to marketers. 

Attributes trend towards monesemanticism. Their meaning is narrowed down to a single parameter or value. Since they then build everything up from there it makes sense to apply the same logic to marketing messages and messaging and content written with a view to SEO. 

The New Rules Of Marketing

Back in 2012 maybe we could get away with calling the marketing approach necessary “new”.  

It isn’t. We should always focus on the experience of what we offer for the person our product or service is intended for. By aiming at the attributes instead of the underlying neurons we create stable outcomes predicated on real human needs and wants instead of chasing the ever-shifting landscape of machine-logic algorithmic and architectural updates. We should seek to serve as opposed to persuade. We should want to help as opposed to convert. 

The difference is a significant one. The approach is supported by behavior that is dependent on specific, identifiable attributes. The same attributes search engines and recommendation engines and, eventually, AI-driven personal digital assistants will use to rank websites, make recommendations and find answers to questions we have and solutions to problems we experience. 


Go Deeper: 

Take Control Of Your Actions.    Make Better Decisions.