Human-centric AI requires not only a fundamental shift in the way AI systems are conceived and designed but also a reorientation in basic research in order to figure out how AI can come to grips with meaning and understanding. Meanings are made up of distinctions to categorize and conceptualize an experience at different levels, from directly observable factual meanings to expressional, social, conventional and intrinsic meanings. Meanings get organised into larger-scale narratives that conceptualize experiences from a particular perspective. Understanding is the process of constructing and then integrating these narratives into a Personal Dynamic Memory that stores narratives from past experiences. This memory plays a crucial role to construct more narratives and thus works intimately together with inferences, mental simulations, and the analysis of experiences in terms of syntactic and semantic structures. This paper outlines this approach to meaning and understanding by clarifying what it entails, outlining technical challenges that must be overcome, and providing links to earlier relevant AI work as well as new technical advances that could make Personal Dynamic Memories a reality in the near future. 2 What is human-centric AI?
“Human-centric AI focuses on collaborating with humans,
enhancing human capabilities, and empowering humans to better achieve
their goals.” [
Achieving human-centric AI requires a number of changes in
focus compared to current AI:
(i) Human-centric AI systems should be made aware of the goals and
intentions of their users and base their own goals and dialog on
meanings rather than on statistical patterns of past behavior only, even if
statistical patterns can play a very important role, for example for
drastically reducing search or carrying out approximate inference.
Human goals and values should always take precedence. Respect for
human autonomy should be built into the system by design, leading
to qualities such as fairness and respect.
(ii) Human-centric AI requires that a system is able to explain its
reasoning and learning strategies so that the decisions are
understandable by humans. Only by emphasizing human understandability will
human-centric AI achieve proper explainability and transparency.
(iii) Human-centric AI should not only learn by observation or
theorizing about reality but also by taking advice from humans, as
suggested in John McCarthy’s original 1958 proposal of the Advice
Taker [
Today the dominating perspective on AI is not human-centric. It focuses primarily on achieving high predictive performance on predefined benchmarks, trying to exceed human performance so that humans can be replaced in the task being considered. This approach is machine-centric rather than human-centric. It emphasizes numerical (subsymbolic) techniques (from neural network research, pattern recognition, information retrieval, and data science), often ignoring valuable contributions from symbolic AI that are needed to achieve explicability and robustness.
Admittedly the machine-oriented focus has recently lead to a jump in performance on chosen benchmarks, particularly in the domain of pattern recognition and computer vision, but unfortunately also to a kind of AI that is opaque, cannot explain or defend its decisions, is unable to take human advice, is not robust against adversarial attacks, has no understanding of the motivations of its users, and requires vast amounts of data and computing power. Although for a large, growing class of applications these shortcomings are not an issue, for AI applications that touch on human lives and are socially consequential, these disadvantages are highly problematic.
Different approaches to human-centric AI have been proposed
recently. They are all valuable. Some researchers have advocated
guidelines and design methodologies to make AI more trust-worthy
and responsible by emphasizing safety, privacy, data governance,
transparency, diversity, fairness, and accountability [
Here I focus on the idea that human-centric AI requires above all another kind of AI, namely AI which has meaning and understanding at its core. The present paper is a position paper, trying to clarifying this point of view and reflecting on the key issues and possible technical solutions. But first, what do we mean by meaning and understanding? 2
The notion of meaning is related to how we try to understand how humans make sense of an experience. An experience can be a behavior or the observation of a behavior, an image or a sequence of images, sounds, soundscapes, smells and tastes, spoken or written text, and more generally cultural artefacts like scenes in a theatre play. In the real world, there is a flow of experiences that we need to interpret and cope with quickly. For example, if we are driving a car there is a quick succession of situations that we have to gauge correctly in order to act appropriately, even in unusual situations: Why is the car behind mine honking its horn? Is the woman with a baby stroller going to cross the street or has she seen me coming? Why is everybody slowing down? What does this red light on the dashboard mean?
Meanings are built from categorisations of reality, for example,
colors, actions types, temporal and spatial relations, etc.
Categorisations are distinctions that are relevant for the interaction between
humans (or agents more generally) and their environment,
including other agents [
Categories are the building blocks for constructing different levels
of meaning for an experience, The following levels are often
discussed in the appreciation of art works [
The base level of an experience details the external formal properties directly derivable from the perceived appearance of the experience, for example, the lines, shapes, color differences in hue, value (brightness) and saturation, textures, shading, spatial positions of elements, etc. in the case of images.
The first level of meaning is that of factual meaning. It identifies and categorises events, actors, entities and roles they play in events, as well as the temporal, spatial and causal relations between them. In the case of images they require a suite of sophisticated processing steps, starting from object segmentation, object location, object recognition, 3D reconstruction, tracking over time, etc.
When there are actors involved, a second level, that of expressional meaning becomes relevant. It identifies the intentions, goals, interests, and motivations of the actors and their psychological states or the manner in which they carry out actions. The next level is that of social meaning. It is about the social relations between the actors and how the activities are integrated into the local community or the society as a whole.
The fourth level is that of conventional meaning, based on figuring out what is depicted or spoken about and the historical or cultural context, which has to be learned from conversations or cultural artefacts, like books or films.
The fifth level is known as the intrinsic meaning or content of an experience. It is about the ultimate motive of certain images or texts, or why somebody is carrying out a certain behavior. It explains why this particular experience may have occurred. We define a narrative as a coherent reconstruction of the different levels of meaning of an experience or a set of experiences based on one or more perspectives. It contains categorised entities at each of these levels, links between the levels, and possibly additional crosslevel categorisations. The perspective, which is often the perspective of the agent itself, is unavoidable because categories are most of the time observer-dependent. For example, an object which is to my left is for a person opposite of me to the right. I may categorise a gesture as aggressive whereas the person making the gesture may have performed it to defend herself. I may not know a particular historical figure and believe it is just the representation of an old man, whereas you may recognize the figure and be repulsed by the atrocities that were conducted under his command. Transforming a narrative from one perspective into a narrative for the same experience from another perspective is a critical component in handling meaning. Even to communicate properly in language we often have to look at the viewpoint of the interlocutor and categorise spatial and other relations accordingly.
Understanding is a process with three functions: (i) Reconstruct the different levels of meaning by casting them into coherent narratives that explain the events underlying the experience, (ii) predict how the experience will unfold in the future and reconstruct what has happened in the past and (iii) integrate these narratives into a Personal Dynamic Memory. A Personal Dynamic memory is an active store of past experiences which may include partly some of the original data but mostly the webs of meanings and the narratives that have been constructed during the interpretation of earlier experiences. A Personal Dynamic Memory is crucial for supporting the construction of narratives of new experiences but it is today missing from existing AI systems.
Here is a simple example to illustrate these ideas. Consider the image in Figure 1 (left). This is from a poster that used to be employed in French and Belgian schools to teach children about daily life and to learn how to talk about it. We instantly recognize that this is a scene from a restaurant, using cues like the dress and activities of the waiter and waitress or the fact that people are sitting at different tables in the room. Current image recognition algorithms would be able to segment and identify some of the people and objects in the scene and in some cases label them with a fair degree of accuracy, see Fig. 1 (right).
However a normal observer would see a lot more than that. For example, when asked whether a person is missing at the table on the right, the answer would be straightforward: Yes, because there is an empty chair, a plate and cutlery on the table section in front of the chair, and a napkin hanging over the chair. So there must have been a third person sitting there, probably the mother of the child. Moreover nobody has a lot of difficulty to imagine where she went. There is a door marked ‘lavabo’ (meaning ‘toilet’ in French) and it is quite plausible that she went to the toilet while waiting for the meal to arrive. Any human viewer would furthermore guess without hesitation why the child is showing his plate to the waitress arriving with the food and why the person to the left of the child (from our perspective) is probably the father looking contently at the child. We could go on further completing the narrative, for example, ask why the cat at the feet of the waitress looks eagerly at the food, observe that the food contains chicken with potatoes, notice that it looks quite windy outside, that the vegetation suggests some place in the south of France, and so on.
Fig. 1. Left. Didactic image of a scene in a restaurant. Right. Image segmentation identifying regions that contain people (based on Google’s Cloud Vision API).
Clearly these interpretations rely heavily on inferences reflecting knowledge about restaurants, families, needs and desires, roles played by people in restaurants (waiter, waitress, bar tender, cashier, customer). These inferences are not only necessary to properly interpret the visual image in Fig. 1 but also to answer questions such as ’Who is the waitress?’, ’Why is she approaching the table?’, ’Where is the missing person at the table?, ’Who will get food first?’, etc., We can also make predictions and reconstructions, for example, that the waitress will reach the table, put the food on the table, cut the chicken into pieces, and put them on the different plates, or that the mother of the child will come back from the toilet, sit down again at the table, and start eating herself.
Each of us has a vast Personal Dynamic Memory that stores narratives based on prior experiences: from visiting restaurants, seeing images in pictures or movies, reading about them, etc. Our daily life is filled from morning to evening with activities to feed and reorganise our Personal Dynamic Memories and the richer they become the more we are able to make sense of new experiences. What is truly amazing is that by the time we reach the adult stage these memories must already contain a massive number of facts, which are nevertheless searched at an incredibly fast rate with relevant parts of memory becoming primed and ready for use for handling novel experiences.
Understanding uses information both from syntactic and semantic parsing of the experience and from inferences based on a Personal Dynamic Memory, in order to fill in unexpressed or un-observable information, e.g. via logical reasoning and mental simulation. Moreover the understanding process changes the contents of Personal Dynamic Memory, not only because the new experience, its interpretation, and links to other experiences are stored, but also because earlier experiences are revisited and their storage may be affected by newer experiences. Memory needed for understanding is therefore highly dynamic, unlike computer memory that remains unchanged once something has been stored.
This leads to the proposal for a general architecture for AI systems that handle understanding depicted in Fig. 2. It shows the flow from experience to syntactic and semantic structures, and from there towards the construction of narratives, integrated into a Personal Dynamic Memory. The flow of information is not only bottom-up but also top-down, shown with the green arrows. The narrative under construction is partially guiding semantic analysis and cutting down combinatorial search in syntactic analysis, whereas the narratives already contained in the Personal Dynamic Memory are guiding the construction of narratives of new experiences.
analysis experiences and data
integration personal dynamic memory syntactic and semantic structures interpretation narratives
Before putting some more technical flesh on this architectural skeleton, I want to emphasize that current techniques and AI design methodologies are not handling meaning and understanding. Current techniques fall into two classes: numerical (or subsymbolic) techniques and symbolic techniques with shades in between.
Simplifying, numerical (or subsymbolic) AI techniques translate problems into a numerical form (real numbers and vectors) and perform numerical operations over them. The numerical representations are constructed using information-theoretic considerations, specifically, their ability to help predict or complete patterns. Most neural networks fall into this class, but also other techniques like Latent Distributional Semantics, which associates a vector representation known as an embedding with words, images, or actions. The embeddings capture the syntactic and semantic contexts in which an element appears and can be used to compute similarities, predict the next word or image, relate an image to a label, answer textual queries, or perform many other useful subfunctions for building intelligent applications. Embeddings are computed either by statistical methods or by using deep learning algorithms.
Importantly, and as pointed out clearly by Claude Shannon [
Thus, ‘Neural image labeling’ associates rather directly labels with images (sometimes even using only pixel-based image representations), without attempting to discern individual objects, actors, or events, and without trying to figure out the situation underlying the image, the nature of the action, the motivations of the actors depicted in an action, the historical setting, the reason why the image is made, and many other aspects which human viewers spontaneously come up with. ‘Neural translation’ does not try to perform a syntactic analysis using grammars and parsers nor semantic analysis using interpreters building conceptual representations of what is being said. Rather, they associate n-grams in the source language with n-grams in the target language based on word vectors that capture statistical co-occurrences in dual (source/target) corpora.
Circumventing meaning has made the current wave of deep learning based AI applications possible but it is also responsible for the brittleness of image labeling, the nonsensical nature of translations, failures in answering questions that fall slightly outside of the statistical patterns in the corpora used to train them, the success of adversarial attacks for interpreting images or texts that do not confuse humans but throw off AI systems, the non-transparency of decision making, and many other features that human-centric AI considers undesirable.
Intuitively a kind of hybrid or integrated AI that combines the virtues of numerical with those of symbolic AI is a possible way out and has indeed been proposed by several researchers. Symbolic AI maps problems into symbols and symbolic structures and performs transformations over these symbolic structures, for example guided by rules of sound logical inference. This approach flourished in the 1970s and 1980s leading to expert systems built for interactively supporting experts, large-scale ontologies and domain models as now used in the semantic web or in encyclopedic knowledgegraphs, computer-assisted theorem proving, constraint solvers for scheduling or design, precision language processing, and much more.
The symbolic approach has tried, at least in principle, to get closer
to handling meaning. It has used terms like semantic information
processing [
Nevertheless, the symbolic approach has its own limitations with
respect to handling meaning and understanding. A key criticism,
reflected in Searle’s Chinese Room argument and known as the symbol
grounding problem, is that symbolic AI operates in a world of
symbols with no systematic connection to the real world. To solve this
problem requires an integration of a symbolic and a numerical
approach, because the latter starts from the (real) numbers delivered
by sensors and actuators that are directly connected to the world,
so that the categories that constitute the meaning of symbols indeed
become properly grounded. However, it is important that the
grounding of symbols is based on what is meaningful, i.e. relevant, to the
agent, which is different from grounding based on success in
prediction tasks. When agents cooperate on tasks in a shared environment,
particularly if they have to communicate about tasks, they
implicitly have to coordinate the way the categorize reality and how these
categorizations are expressed.[
In addition, the transformations of symbolic structures are formal operations, similar to a set of axioms and rules of logical inference as in mathematics. But the problem is that it is very hard, if not impossible, to define axioms exhaustively for real world open-ended domains due to the unavoidable exceptions, lack of knowledge, and the problem of making clear-cut definitions. These problems have been discussed widely under the title of the frame problem. Also here an integration of numerical and symbolic techniques is a way to go forward so that the flexibility of pattern recognition and action selection based on neurally inspired models, which gives only approximate answers, can be married to the precision and compositionality of symbolic reasoning. 4
The ideas proposed here are certainly not new. For a long time it has
been commonly accepted in cognitive science that the construction
of narratives is an essential ingredient of cognitive intelligence
because it allows us to make sense of reality [
In the psychological literature there has also been extensive work
on personal memory, often based taking Tulving as a starting point
[
Meanwhile various important technological advances have been made in other areas that make a renewed effort towards the experimentation with Personal Dynamic Memories and narratives a realistic prospect. Among these advances I just want to highlight the following:
Very large knowledge bases. One of the critical bottlenecks for
effective Personal Dynamical Memories is the sheer size of the
knowledge that has to be represented and processed. If we
express this in terms of facts, then we must expect to handle at least
tens of millions, if not billions. This was totally impossible two
decades ago but very significant progress, pushed by the
development of the semantic web, has changed the situation. It is now
possible to represent fact-bases up to 100 billion triples using
standard knowledge representations (RDF statements and OWL) and
perform inferences over them fast enough to be used in
interactive applications[
Finally there have been significant advances recently in
Computational Construction Grammar. Most linguistic formalisms, such
as Chomskyan generative grammar, remain close to the
morphosyntactic structure of a language. Construction Grammar in
contrast focuses on capturing the systematic ways in which grammar
expresses meaning [
In my opinion, the most critical bottleneck at the moment is: How should a Personal Dynamic Memory be organised at the microlevel and what kind of basic computations (including inferencing and learning) should be supported. Obviously a linear list of facts, possibly represented in RDF, will not do, we need higher level structuring devices, partly for managing inferential and combinatorial complexity, partly for dealing with the frame problem, and partly for achieving fast access to the most relevant prior experiences that will help to make sense of a new experience. What will also not work is to blindly store the vast amount of information generated by an experience, the complete sensori-motor data streams, the data from the mental simulations that are triggered, the language descriptions and their semantic interpretations, or all the facts relevant for an experience. If eveything is stored this is not only costly from an energetic point of view but will certainly get in the way of fast retrieval and inference.
The cognitive science and AI literature already contains various
proposals for the organisation of memory. Many of them start from
Bartlett’s original idea of a schema also called a frame. It was
formulated in the 1930s and revived again in the 1970s by psychologists
such as David Rumelhart [
A schema is a way of framing a particular situation in terms of a set of entities, roles for these entities, constraints on the kind of entities that can fill these roles, and relations between the entities based on their roles. Each schema has various associated cues to recognize quickly whether it applies to the current situation. Once it is triggered, a schema casts a web over the sensori-motor inputs and facts associated with the situation and it makes us see or infer certain aspects of the experience more clearly at the expense of others. Schemas impose a bias and perspective on a situation and often also an emotional reaction. They come with a lot of defaults. These are facts which can be expected to be the case if a particular schema matches well with an experience, but are not explicitly mentioned or observable. Sometimes these defaults even override perception or fly in the face of obvious facts.
The notion of a schema was introduced into AI by Minsky[
The paper argued that human-centric AI, with its implications of explainability, transparency, robustness, etc., is only going to be possible when AI comes to grips with meaning and understanding. This requires that we go beyond the numerical AI paradigm that is currently dominating AI, where meaning is captured only very indirectly in embeddings and operations over embeddings, but also beyond the symbolic paradigm, which focuses on formal operations over nongrounded symbols.
First of all we need at the very least a form of integrated or hybrid AI that combines numerical and symbolic AI. But we need to go beyond both. The paper argued that a central characteristic of understanding is the ability to build a coherent narrative of an experience based on narratives of past experiences stored in a Personal Dynamic Memory, and integrate this narrative in memory. The big challenge for AI is partly technical, to solve problems of computational complexity to handle the very large knowledge bases and huge inferences that are required. But it is also conceptual. We need to understand much better how new experiences and the narratives built for them get integrated into a Personal Dynamic Memory in such a way that they get triggered again on the most relevant new experiences, and how facts or narratives that are deemed no longer relevant can be forgotten or simply not stored in the first place.
Acknowledgement The author is funded by the Catalan Institute for Advanced Studies (ICREA) embedded in the Institute for Evolutionary Biology (UPF/CSIC) in Barcelona. This work was made possible by H2020 grants within the frame of the Humane AI Flagship preparation project and the AI4EU project.