Why A.I. doesn’t learn like humans: Part 3— Curiosity, Motivation, & Logic
Life has three developmental strategies to become better adapted for an environment: evolution, knowledge, and technology; each a prerequisite for the next. Knowledge was at first only stored inter-generationally through genetic material and then intra-generationally with the development of learning capabilities and a nervous system. This change sped up the rate and extent of knowledge acquisition. We (life) survived this way for most of our existence until we learned to manipulate our environment and create technology. It was a very long time ago that human ancestors first invented tools, then language, agriculture, machines, and now computers. Arguably, computers are the greatest survival strategy any species has ever devised. However, our most capable computers still fall short in comparison to the intelligence of the human mind [for now]. The artificial narrow intelligence (ANI) machines that we’ve created, though very capable in their narrow band of function, lack certain capabilities that have allowed the human mind to become so diversely proficient. Part 2 of this series conceptually described deep learning, a learning strategy that is inherently able to identify patterns. However, deep learning by itself can only enable ANI machines. This part will discuss most of the aspects that would enable the transformation from ANI to artificial general intelligence (AGI): curiosity, motivation, and logic.
In order to become diversely proficient, a diverse collection of knowledge is required. Curiosity is the means through which humans seek out this diverse knowledge. Without a mechanism that compels it to collect information, all ANI lacks curiosity. Instead, a human counterpart is required to present the relevant data, usually in a very specific structure. However, it is not helpful or even feasible for a curiosity system to compel the collection of any and all information. Curiosity, to be effective, needs to be directed. We’ll return to this point later.
It can be argued that having a system whose sole purpose is to acquire knowledge would lead to a never-ending pursuit of knowledge. Accordingly, in conjunction with systems that compel us to gain knowledge, we must have systems that balance our energy expenditure on knowledge acquisition. When we are directly fighting for survival, it is disadvantageous to expend energy on acquiring knowledge. If you’re approached by a dangerous animal, you generally don’t want to ‘get curious’. However, developmental anomalies can cause dysfunction in the curiosity system of the brain. Such developmental anomalies in the curiosity system would likely present with characteristics similar to autism.
Autism occurs as a spectral disorder coinciding with the manner in which the curiosity system’s balance is disrupted. With not enough drive to acquire knowledge, the non- to low- curious types are left without a mechanism to become knowledgeable. With too much drive to acquire knowledge, the high- to full-curiosity types are left fixated on a specific sensory modality or subject matter that can become “all-consuming”. The term “all-consuming” in this instance refers to two aspects: 1) the amount of time and energy that is devoted to acquiring knowledge, and 2) the areas of the brain being adopted in knowledge acquisition and storage that would otherwise be responsible for other functions, such as language processing, social relationships, or speech synthesis. Sometimes, as in the case of autistic savants, some abilities of the high- to full-curiosity type of autistic individuals can be remarkable and superior to normal-functioning humans. In the theory of autism presented here, this heightened “savant” ability stems from an increased potency of the brain to acquire knowledge, with the brains ability to find patterns in that information still intact. Therefore, autistic savants have impressive ANI-like abilities that foreshadow the capabilities expected from an AGI. Given that an AGI would not have the same physical limitations of the information processing resources as an autistic savant, the development of these impressive abilities in an AGI would not need to be “all-consuming”.
Curiosity is the medium through which the world is experienced and it occurs in at least two ways: sensing and direct knowledge transfer. When sensing, a sensor detects information that is interpreted by learning to reveal insights about patterns in that information. As discussed in the post about deep learning, these detected patterns become concepts and the relationship between them becomes insight/logic (more on insight/logic to follow). On the other hand, direct knowledge transfer is an automatic reveal of insight that bypasses the requirement of pattern detection. Although it still requires some form of sensory input, direct knowledge transfer is energetically favourable because it requires much less computation to produce a similar result. For example, when a bee is agitated, it releases a chemical pheromone into the air that makes other nearby bees behave more aggressively to a potential threat. Using a chemical signal, bees use direct knowledge transfer to reveal that there is a nearby potential danger. This method is much more effective and energetically favorable than sensing, which would require all nearby bees to independently sense the subject and reach the necessary conclusion.
Humans have especially optimized direct knowledge transfer through the use of language. Through it, we can request and communicate insight without requiring the recipient to sense the information directly. Sentences enable us to track the progression of logic to reach a deep conclusion (i.e. recognize a complex pattern). Within sentences, words are used to signify concepts while grammar identifies the relationship between those concepts. However, the downside to direct knowledge transfer is the difficulty of forming new concepts with guaranteed accuracy. For example, describing the experience of color to a congenitally blind person is less than ideal in enabling that person to understand the description. It is almost impossible to transfer concepts of raw sensory experience. Therefore, direct knowledge transfer is most effective in the exchange of insights based on a common sensory experience. With this conceptualization of curiosity, it becomes feasible to gauge the extent that something is curious as the extent to which it has the ability to add to and modify its knowledge. Therefore a hypothetical AGI would need to be able to add to and modify its knowledge.
The ultimate purpose of curiosity is to build a semantic recreation of the world within the mind: a representation of knowledge in the context of other knowledge and the manner in which the knowledge was acquired. However, this effort would not be worthwhile if the information being learned is random. Instead, the curiosity system directs the pursuit of knowledge by measuring energy expenditure versus the amount of new information acquired. New knowledge that is similar to pre-existing knowledge requires less energy to be incorporated into previous semantic recreations of the world than the energy required by fundamentally new knowledge. For this reason, the curiosity system compels us to explore avenues of knowledge acquisition that are familiar, building new knowledge on top of previous knowledge. However, after a certain level of complexity, finding new information becomes difficult, slow, and energetically unfavorable. It is at this point that people typically become disinterested and experience boredom, moving onto another pursuit.
The curiosity system helps to explain the entertainment phenomenon that has overtaken human culture, such as watching television or Netflix. The learning technique employed by the entertainment industry is possibly the most energetically favorable way to gain new knowledge. Considering that it typically happens in a low-energy sitting position (i.e. on a couch), that the most popular entertainment content uses episodes (engaging the energetically favorable method of building new knowledge on top of pre-existing knowledge), and that many people consume food while watching (turning the act of watching TV or Netflix into an energetic surplus), it is no surprise that these services have become as popular as they have. The entertainment phenomenon of human culture has hijacked the curiosity system of the brain because the world that is presented by these media is typically fictional. Through television and Netflix, people are building a semantic recreation of fictional worlds within their minds. And sadly, this new knowledge is rarely applicable (though some of it is undeniably beautiful). Essentially, these services have turned vital biology into a vestigial system of complacent time investment with futile hope of returns on that investment.
While curiosity concerns the acquisition of the semantic representation of the world in the mind (semantic because everything is organized in relation to everything else), the motivation system is used to select and pursue the best next course of action. During this selection process, the action that can potentially be taken is selected from the same actions that were used to produce the knowledge of the action’s outcome via the curiosity system. The motivation system cannot select for an action that has never been experienced (and imagining an action is indeed a way to experience it — I may describe imagination in part 5 of this series).
Part of the context that makes the information semantic is the connection between the sensory input and the available actions, in addition to the expected outcome of each action. Sensory input flows into the mind and after being combined with internal drives that promote survival, such as hunger, reproduction, social connectedness, curiosity, alertness, and emotional states, the motivation system outputs an action. The particular action that is selected corresponds to the strongest internal drive in proportion with the available opportunities to satisfy that internal drive, as detected by sensing. If an opportunity for a particular action is not sensed, the possibility of that action being selected is essentially non-existent. If you do not see any rocks, you do not contemplate picking up a rock.
Upon selection of an action, an expectation is created to determine whether the outcome of the action satisfied the internal drive by the expected amount. The amount of satisfaction expected equals the strength of the internal drive estimated to be attenuated by the action as learned from previous experience. If the result is unsatisfactory, the strength of the inputs that led to the selected action are reduced. If the result is exactly as expected, the strength of the inputs are left unchanged (but the connections can become more fixed). And, if the result is better than expected at satisfying (i.e. attenuating) the internal drive, the strength of inputs for selecting that action in the future are enhanced.
In short, opportunities are sensed, the information is combined with internal drives to select an action, and the value of each opportunity is adjusted through feedback. It is this chain of processes that comprises the motivation system.
This motivation chain is dependent on previously-defined values gained through sensing and direct knowledge transfer. These values represent an estimation of the extent to which a concept contributes to attenuating an internal drive. It is the goal of the advertising industry to manipulate these values in peoples’ brains to get them to carry out a specific action or to change their perception of a concept. Ultimately, the value system is responsible for determining the likelihood of selecting one action over another to satisfy internal drives.
It is noteworthy to point out that the physical distribution of actions in the mind is also not random. Actions that have a similar intended output are closer together. It is possible to produce an intention-map of the mind (located in the posterior parietal cortex), with similar geographical areas producing similar actions. This proximity allows the cell assemblies that represent these actions to be redundant while enabling the selection of one action to inhibit related actions. For example, it is possible to lift a cup with a hand in many ways, such as by pinching the rim using the thumb and the pointer-finger, wrapping the side with the thumb and fingers, clawing the top using the thumb and fingers around the rim, etc. However, it is not possible to undertake more than one of these actions simultaneously with the same hand. Accordingly, as one action is selected, other related actions are inhibited to prevent confusion of the output. The intention-map helps to keep output actions efficiently represented and focused.
All of the systems I’ve described so far can operate without any integrated comprehension of the relationships between the sensory inputs, internal drives, or selected actions. Curiosity and motivation, by themselves, represent the way an organism with little ability for logic would function, such as a mouse. A mouse does not understand why it inevitably always gets hungry (relationship between energy expenditure and hunger) or why it wants to reproduce (relationship between reproduction and survival). It simply either gets hungry or sexually motivated.
A strong logic system enables humans to theorize the relationship between concepts and to rewire sematic knowledge so that each concept is meaningfully related to other concepts. This process is the backbone of language. Every word is a concept and the rules of grammar identify how the concepts are semantically related. By engaging the logic system’s ability to rewire the brain, concepts get connected in meaningful ways to produce insights. For example, as you read this sentence, you are absorbing each word to activate specific neuronal assemblies that represent each word. Simultaneously, you are using the grammar to define the semantic relationships between those words/assemblies. The outcome of that activity results in the creation of connections between specific neuronal assemblies that are meaningful. When those neurons are connected in your brain, you are empowered with a new understanding. Short connections typically produce superficial insights while connections that span several groups of neurons connected in a particular sequence produce more powerful insights (because they expose patterns in deeper layers, as covered in the previous post on deep learning). However, the more ‘orders of logic’ that are required to form a connection, the more difficult the connections are to establish.
The logic system also governs our ability of math, or any other ability that depends on specific rules for defining the relationships between concepts. It is intimately responsible for our heightened ability to make tools and to manipulate our environment. It enables us to derive hypotheses about the relationships of things so that we can develop experiments and carry out science. Most importantly, however, the strength of our logic system is the main reason we have our level of consciousness. It is only by developing a strong ability for logic that humans have become as self-aware as we have, enabling similarly powerful volition. Without the ability of logic, sentience is not possible.
In this post I’ve described three systems that exist in the human brain: the curiosity system, the motivation chain, and the logic system (more on the logic system in part 4). A computer that is empowered with these three systems will be substantially more versatile and capable than any ANI that exists today. It would instead act as an AGI that is able to acquire semantic knowledge on its own and gain a meaningful understanding of relevant information. It would be able to use its semantic knowledge to self-select actions that most closely align with its programmed goals (i.e. internal drives) and with each outcome, to adapt its internal values for selecting those actions in the future. Such an AGI could theorize about the relationships between concepts to derive insights that further its semantic understanding of the world. It could predict events and forecast activity without requiring as much data as deep learning requires today, and it could communicate this information using language. This AGI would be the most powerful tool ever created by humans. However, like any tool, this AGI would only perform as it is programmed to. It will not have the ability of sentience or volitional output. It is not capable of having internal goals that it was not programmed to have. The next part of this series will further the description of the logic system, how consciousness operates (aka self-awareness, the subjective experience, etc.), and how such a system would be conceptually built into an AGI.
