How the human brain differs from deep learning approaches in AI

How the human brain differs from 

deep learning approaches

 There is an emerging view of the human brain as an engine of probabilistic prediction. Statistically-driven perspectives of the mind such as predictive processing and the Bayesian brain view have been gaining prominence in the last decade. Incoming data is interpreted through the lens of expectations we model the world around. We extract patterns from the world around us and use them to update the models we will use in future interpretations. When framed like this, there are surface similarities to some of the techniques used in recent AI applications. The deep learning (DL) approaches currently dominating the field of AI are in essence large-scale probability-driven algorithms, adept at finding patterns in large data sets. Given this surface similarity to DL training methods, is the human brain really so different? The simple answer is “yes”, but the “why” behind that requires some unpacking.

Broadly speaking, machine learning technologies recognize patterns and based on past observed data predict future outcomes. Deep learning is a subset of machine learning that relies on artificial neural networks (ANNs) to encode patterns in data. ANNs are networks of nodes divided into connected layers. Signals travel between nodes and assign weights to those nodes. More heavily weighted nodes exert more influence on the next layer of nodes. The final layer compiles contributing node weights to produce an output. The ANN “learns” by comparing its output to a success criterion (such as human-labelled images in the case of image recognition tech), marking the error between its own output and the target criterion, and making adjustments to its weighting to try to reduce this error with each successive iteration. This process continues with the machine gradually coming to produce output similar to the target criterion. Information is fed to the machines in the form of data sets, which the ANNs then learn to classify through this process.

The ANNs themselves are loosely based on a simplified view of the way neurons in the brain code information. It’s worth emphasizing that this view is not only heavily-simplified, but also heavily-outdated. The ANN approach was conceived of and developed many decades ago, over which time neuroscience has progressed considerably while the same ANN principles have gone on to inform future AI technologies. While it is true that the summation of neural inputs contributes to the activation of connected neurons (not unlike the basic idea behind ANNs), the similarities between the two systems do not extend much beyond that. In ANNs, data is represented in a centralized rather than distributed form, data is processed sequentially rather than in parallel, and the inbuilt feedback systems present in the brain have no equivalent in ANNs. Furthermore, the brain may actively modify its architecture, producing entirely new neural connections (or eliminate old ones), but ANNs cannot go beyond adjusting their signal weightings. Their architecture itself is fixed from the beginning, which places greater limits on the restructuring that can potentially occur. What this all adds up to is a much less versatile machine than the brain.

To take on a specific example of an implementation of the DL approach, let’s focus our lens on large language models (LLMs). These models are essentially probability maps that model the likelihood that a given word will follow another given word, not entirely dissimilar to the autocomplete function you find on any smartphone. A large corpus of text is fed to the model (the training set), and when given a prompt the model will produce the words mostly likely to follow that sequence. For LLMs to move beyond being only high-powered autocomplete functions, they need to be implemented in a system that will frame these prompts in a particular way. Chatbot implementations of LLMs, for example, may present the prompts to the model within the framework of a conversation in which their role is to provide a direct and useful response, with the result that the model will generate words in the form of a reply to the prompt. The reason it does not produce gibberish or even something ungrammatical is because ungrammatical nonsense is somewhat less likely to occur in the corpus that it has been trained on than more sensible responses, the result being that the machine more often than not gives the impression of competence.

Disregarding any appearance of interacting with an intelligent agent, it cannot be emphasized enough that the model has no understanding of anything that it is producing. If one were to ask it how plants produce energy from sunlight and the machine were to reply with an explanation of photosynthesis, it is because in the context of a conversation in which one actor asks another how plants produce energy, the most likely way that their conversation partner would reply is to give an explanation of photosynthesis. An offshoot of its complete lack of understanding of what it is producing is that it has no means of measuring truth from falsehood. The only metric that the model itself uses when crafting its reply is the likelihood that words will appear paired together in its corpus and that gives no measure of the truth of any given statement. The machine makes no distinction between Donald Duck and Donald Trump in terms of truth. Reality holds no special privilege over total fabrication, which accounts for the reputation of chatbots as occasional generators of misinformation. But even when it is outright fabricating information, it does so with the same surface appearance of competence it would use for anything else. This results in a situation where one needs to fact check its replies, which rather defeats the entire point.

A further deficit of LLMs is that they lack communicative intent. A conversation with another individual occurs in a context in which the speaker knows that they will be heard by another and their words may go on to affect the beliefs and behavior of the listener. This has an effect on the way that we communicate information. Furthermore, when we are interacting with another individual, especially in the case of a known individual, we may bring to the interaction beliefs about the intent and knowledge state of our interaction partner. We may have some ideas about what they already know, so we don’t need to reiterate things that we can take as a given. We may have some idea about what their motive is in asking a question (perhaps they’re only interested in a specific aspect of the answer) which allows us to leave out irrelevant information. In short—we are able to communicate relevant information more efficiently and effectively because of outside knowledge that we bring to the conversation and apply with communicative intent. This is another feature that LLMs totally lack, which inhibits their effectiveness as a conversation partner.

A further distinction between the human mind and DL-based AI is that we have the capacity to organize the information we have extracted into modifiable concepts that can be reflected upon. The meaning of words can be represented abstractly and generalized to new situations. Statistical regularities extracted from the environment go on to form higher-level concepts (such as beliefs) that exert a top-down influence on our behavioral output. These higher-level concepts may be accessible for self-reflection and deliberate modification. This results in a feedback loop between a reflective agent, its own internal workings, and its environment. Our ability to actively reflect upon and modify aspects of our own mental architecture is a key feature that strongly distinguishes us from pure applications of the DL approach.

Knowledge transfer, cross-domain reasoning, and flexible problem solving are abilities where the human mind far outshines anything DL could produce. Training in DL amounts to building a predictive model out of a data set well-suited to solving a particular problem. The architecture that emerges from this training may perform well enough when solving the task it has been trained for, but this architecture has no means of transferring its knowledge beyond the bounds of that domain. This may be sufficient when the developed architecture is being used as a tool for a closed-environment task, but it can be a problem when these architectures are applied (or misapplied, depending who you ask) to real-world environments. The training data set cannot encompass all possible exceptions that may be encountered, and without flexible problem solving abilities, the developed architecture cannot effectively react to unforeseen occurrences (try letting your search engine autocomplete “self-driving car crashes into…” and take your pick to see some examples). This is more than just a problem to be overcome; it lies at the very core of the deep learning approach and cannot be improved on by further extensions of the same logic. There may still be a future for the DL approach as a single component of a more multi-faceted AI architecture (possibly as a module for statistical learning), but the industry’s current overreliance on this approach is misguided and overly optimistic. The technology is already approaching its limit.

Let’s look at a potential AI application – removing the second pilot on commercial airline flights.

We load up the system with instruction manuals, wiring diagrams, short courses on aerodynamics and electronics. Fault behaviour is very unusual, so we are unable to cover the field with many examples in any training set we could provide.

During a flight, the aircraft begins to manifest a fault. There is nothing to account for the issue in the manual – what happens now?

We ask the system to hypothesize as to the cause of the fault. It begins to search around the general location of the fault, looking for out of spec operation. It tracks the fault down to a specific module, asks for and receives approval to take it offline.

How did it accomplish this? It used the meanings of words, and the realization of those words in the architecture of different systems - it could not just regurgitate a string of words that already existed – it used AGI. The DL approach cannot produce outcomes like this because the problem is unique and not represented with high enough frequency in any training set that could be provided to the model. DL may have use as an updated ELIZA, where its users are not critical of its performance, but deploying these technologies in any serious applications is outright irresponsible.