The activation of episodic memory in a computerised model of an invertebrate brain paves the way for a future artificial intelligence capable of learning autonomously from its environment, being adaptive, and not requiring millions of training data.

The significant evolutionary push experienced by generative AI in recent years, exemplified by its two most prominent and popular manifestations, ChatGPT and DALL-E, has refocused attention on Artificial General Intelligence (AGI). This term refers to the grand, robust “general” artificial intelligence capable of closely replicating the functioning of the biological brain and all its cognitive processes.

Analysts at McKinsey & Company caution that, despite the progress made by AI over the past decade, the era of an artificial intelligence that can pass the Turing test – demonstrating an ability to go beyond predictive algorithms and be indistinguishable from human intelligence – is not yet upon us. It may still take decades or even centuries before we can speak of “general artificial intelligence.” However, they add, we are on the right path. Among the capabilities that AI must master to reach the AGI level, they highlight excellent visual and auditory perception, high-level motor skills, precise natural language processing, memory, problem-solving, navigation (with simultaneous localization and mapping), creativity, social skills, and emotional competencies.

«Unlike predictive machines like ChatGPT» they comment, «which can predict a specific response with a high degree of accuracy because they are trained on vast amounts of data, AGI tools could exhibit cognitive and emotional abilities – such as empathy – that rival those of humans. They might even be able to consciously grasp the meaning of their actions» [source: “What is artificial general intelligence (AGI)?” – McKinsey & Company, 21 March 2024].

Let’s now attempt to provide an overview of the research activity in the direction of Artificial General Intelligence to understand where we currently stand.

---

TAKEAWAYS

---

Metamemory and multimodality: insights from international research

Among the various fields of study within Artificial General Intelligence (AGI), the endeavour to simulate human brain activity by integrating cognitive processes that govern memory of the past into machines is one of the most compelling.

Addressing this topic, a group of researchers from Nagoya University in Japan presented a model of an artificial neural network capable of performing a range of metamemory-related tasks, as detailed in “Evolution of metamemory based on self-reference to own memory in artificial neural network with neuromodulation” (Scientific Reports, 26 April 2022). Metamemory encompasses the set of skills used to recall learned information over time, allowing adaptation of behaviour to the environment.

In contrast, a team of scientists from Renmin University of China in Beijing tackled the challenge of developing a machine that surpasses the limitation of possessing a single cognitive skill, a constraint characterising most current AI research methods.

The team developed a large-scale, pre-trained AI model using an extensive dataset of multimodal data(specifically, 650 million image-text pairs). This model can be adapted to perform various cognitive tasks. The experiments conducted to test the model across a range of tasks highlighted its ability to transfer knowledge inter-domain via multimodal learning. The authors particularly noted that this approach enables the model to acquire imagination and reasoning capabilities [source: “Towards artificial general intelligence via a multimodal foundation model” – Nature Communications, 22 June 2022].

Perception of occluded objects and humanoid locomotion in open spaces

The challenge of machine visual perception, particularly in identifying objects that are close together,overlapping, or within crowded spaces, is the focus of the work described in “Persistent Homology Meets Object Unity: Object Recognition in Clutter“, by the Department of Mechanical Engineering at the University of Washington, Seattle, published in IEEE Transactions on Robotics in December 2023. In an attempt to replicate the human cognitive mechanism where partially visible elements are recognised by associating discernible features (shape, colour, position) with a whole object stored in memory, the authors developed a system combining computational topology and machine learning techniques. This system creates 3D representations of occluded objects based on their shapes, classifies them, and compares them with a library of previously stored representations [for further details, we recommend reading our article “Navigating crowded spaces: a fresh look at how service robots understand what they see“].

In the realm of general artificial intelligence, a research group from the University of California, Berkeley, presented their work on humanoid locomotion in open spaces in “Real-world humanoid locomotion with reinforcement learning” (Science Robotics, 17 April 2024). They proposed a reinforcement learning AI modelfor controlling humanoid locomotion in real-world environments (as opposed to research labs). This model was tested on a life-sized humanoid robot without an onboard vision system, through both indoor and outdoor experiments. The robot successfully mimicked human walking and gait patterns in direct contact with the environment [for more information, see our article “The role of reinforcement learning in controlling humanoid robots’ locomotion on natural terrain“].

Focus on episodic memory: what it is and the cognitive skills it enables

A recent study by a team from the University of Illinois Urbana-Champaign revisits the cognitive processes related to memory. The study, titled “Cognitive mapping and episodic memory emerge from simple associative learning rules” (Neurocomputing, print issue forthcoming on 28 August 2024, with an online preview available), particularly focuses on episodic memory in the context of artificial intelligence.

According to the American Psychological Association, episodic memory in mammals (including humans) and all vertebrates, in general, refers to the encoding of the spatial-temporal context of past events and experiences, supported by the hippocampus region of the brain.

«In the hippocampus, experiences are represented as an integrated ‘cognitive map,’ where events are presented as ‘objects’ organised within a context, and memory is stored in episodic sequences of events across time and space. Through the cognitive mapping of contextually organised objects, episodic memory provides the associative substrates for past awareness and creative divergent thinking in humans, other mammals, and birds», explain the authors.

Regarding the animal kingdom, the University of California Irvine highlights in “The evolution of episodic memory” (Proceedings of the National Academy of Sciences – PNAS, 2013) that «fundamental characteristics of episodic memory are present in mammals and birds, and the main brain regions responsible for episodic memory in humans have anatomical and functional homologues in other species».

Current AI systems, lacking human episodic memory and cognitive maps to generalise across multiple environments and tasks, cannot encode the spatial and temporal contexts of past events and experiences.

Building on the thesis that we currently lack artificial intelligence algorithms with the flexibility to enable the range of social and creative activities found even in the smallest insects, researchers at the University of Illinois Urbana-Champaign aimed to explore how even the simplest conditioning mechanisms introduced in simulating the brain of an invertebrate (which, unlike vertebrates, do not possess episodic memory) could trigger sequential learning and thus episodic memory. Let’s examine how this works.

Software simulation of an invertebrate’s neural network, with episodic memory activation

The invertebrate in question is the California sea slug, *Pleurobranchaea californica*, roughly the size of a grapefruit, residing in the Pacific Ocean where it spends its time crawling along the seabed in search of food.

In a previous study (“Implementing goal-directed foraging decisions of a simpler nervous system in simulation” – eNeuro, 2018), the same research team from the University of Illinois Urbana-Champaign examined the functioning of the brain of this small creature, which consists of only a few thousand neurons. They mapped it to create a software simulation of its neural circuits responsible for decision-making processes. 

«Its brain is small and simple. Yet it is capable of executing relatively complex decision-making processes and can even learn from experience», the authors observe.

Specifically, the computer model of Pleurobranchaea (named Cyberslug) simulated its foraging for prey, which involved three different types of smaller slugs: Hermissenda, which are nutritious and lack natural defences (thus its preferred food), Flabellina, equally nutritious but with toxic spines (Cyberslug will approach them only if very hungry and without alternatives), and finally, a fake food (introduced for the experiment) that emits a chemical scent similar to Flabellina but lacks toxic spines.

The simulations revealed that, with only a handful of neurons, the Cyberslug model learned, through trial and error, to select nutritious food over harmful options. By making the correct choices, it optimised its use of food resources, obtaining sufficient nourishment to survive while avoiding as much damage as possible (toxic spines and fake food with a pleasant scent but no nutritional value).

Introduction of reward-based conditioning scheme

In the 2024 study, the team integrated a computational module for episodic memory into Cyberslug, called the Feature Association Matrix (FAM), modelled on the architecture and functions of the hippocampus in the human brain, which is a key area for learning and memory.

«Although Cyberslug could learn from experience, its memory and ability to integrate information from past experiences were limited», stated the researchers, explaining the rationale behind this integration.

In the new study, Cyberslug-FAM explored a new simulated environment with various paths, some of which were associated with reward-prizes by the researchers.

It was observed that, utilising the computational module for episodic memory, the computerised model of Pleurobranchaea developed cognitive maps to learn its spatial environment. This enabled it to remember the most functional paths (rewarded with prizes) for finding nutritious food and shortcuts to traverse the environment more quickly and efficiently, resulting in increasingly greater rewards. «This is an example of spatial reasoning, with event memorisation», the researchers conclude.

Glimpses of Futures

The software simulation of the neural circuit of a sea snail, engaged in simple tasks of food foraging and route selection, has confirmed an approach that, if validated through more complex experiments, could potentially be used in the future to develop advanced artificial intelligence algorithms.

These algorithms could enable machines to autonomously perform a range of tasks by using cognitive maps that allow them to learn from their environment and retain what they have learned.

Let’s now try to anticipate possible future scenarios by analysing the impacts, both positive and negative, that the evolution of the described approach might have from various perspectives using the STEPS matrix.

S – SOCIAL: the future prospects of this work are undeniably fascinating. They envision, in the coming decades, machines that can continually learn from their experiences in their operational environments and retain this knowledge to later recall, relive, communicate, and utilise it to develop new models and more effective policies for their tasks. This scenario could be driven by the episodic memory of artificial intelligence, making the behaviour of its agents more akin to that of humans and, more generally, mammals. Consider robots involved in rescue operations, those employed in hospitals, and those assisting the elderly and disabled – these are just a few of the most critical and impactful applications in terms of human interaction. This helps to better understand the significance (and value) of a machine that remembers what it has done and can act, adapt, and decide in the present based on this memory.

T – TECHNOLOGICAL: for the proposed approach to an AI agent equipped with episodic memory and cognitive maps to materialise, the development of an architecture that combines artificial neural networks with techniques for storing and retrieving episodic memories will be necessary. The cognitive skills related to episodic memory, allowing for generalisations in different environments and tasks, suggest the development of simple network architectures that, unlike deeper and more complex neural networks, do not require large amounts of data for their training. We are talking about a future AI that will learn autonomously from the environment and thus only require basic pre-training. This advanced AI is expected to impact Large Language Models, particularly improving the performance of ChatGPT, transforming it from a “prediction machine” into an AI agent that learns through conversation.

E – ECONOMIC: Tìthe drawbacks of an artificial intelligence increasingly close to human intelligence, such as one that integrates episodic memory, are far from minimal. In the aforementioned article “What is Artificial General Intelligence (AGI)?” dated 21 March 2024, McKinsey & Company analysts, while acknowledging the uncertain timelines for the emergence of Artificial General Intelligence, note that «when it arrives – and it likely will, sooner or later – it will be a major issue for every aspect of our lives, activities, and societies». The labour market will be among the first to experience the negative impact, in addition to potential benefits, of machines that learn by doing, remember their actions, and adapt to any department they are placed in. «Executives worldwide can begin to work now to understand the phenomenon that will lead machines to achieve a level of human intelligence, resulting in even more automated workplaces than we have today», advises McKinsey & Company. This is supported by the Forrester Job Forecast 2020-2040, which, analysing prospects up to 2040 for countries like the UK, Germany, France, Italy, and Spain, estimates the disappearance of 12 million jobs, with retail, hospitality, and catering sectors being the most at risk.

P – POLITICAL: in a future scenario where AI performance increasingly resembles that of a biological brain, critical aspects of this phenomenon must be considered with utmost care. For future robots that learn autonomously from their environment and remember past experiences to the point of recalling and using them to learn new ones, concerns regarding their safety levels, given their interactions with users, are pertinent. How do they adapt to the environment? What do they learn? Who controls and supervises what they learn? And what if they learn something erroneous and harmful? The mind immediately turns to the world of work. Recalling the EU regulation that replaced the Machinery Directive 2006/42/EC – namely, Regulation 2023/1230, published on 29 June 2023 and applicable from 20 January 2027 – it is important to emphasise the focus imposed on manufacturers regarding new safety components and on users concerning new protection requirements in human-machine collaboration, as well as the proper management of robots within companies.

S – SUSTAINABILITY: From an environmental sustainability perspective, it is positive that future AI models equipped with episodic memory and cognitive maps, learning from the environment and past experiences, do not require thousands of data points for their training. This reduces the hours dedicated to training algorithms and processing tasks, consequently lowering CO2 emissions and carbon footprint, two indicators often at high levels when referring to artificial intelligence techniques and, more generally, the digital technology sector.