Deciphering human intelligence for better machine intelligence

Article information

Abstract

A Book Review on

A Thousand Brains: A New Theory of Intelligence. Jeff Hawkins (New York, NY, United States: Basic Books), 2021, 288 pages, ISBN: 978-1541675810

If you are interested in acquiring knowledge about the functioning of artificial intelligence and its potential similarities to the human brain, this book is an excellent resource for you. The book presents a surprisingly simple yet comprehensive theory regarding the functioning of the human brain, and it subsequently offers profound insights into the fundamental mechanisms of machine intelligence. Jeff Hawkins offers a distinctive perspective on the neocortex, having transitioned from the field of computer science to neuroscience. He maintains that true machine intelligence cannot be achieved without a comprehensive understanding of the functioning of the human brain. He was the founder of Palm Computing, which previously manufactured one of the smallest computing devices and the most advanced smartphones, namely the Treo with Palm OS, in the early 2000s. Subsequently, he resumed his erstwhile intellectual pursuit, which he had left unfinished in the 1980s, by establishing a brain research laboratory, Numenta, and conducting self-funded research on the neocortex and machine intelligence technology.

The Thousand Brains Theory

At the core of Hawkins’ theory is a 2.5 mm thick cortical column, which represents the fundamental unit of the neocortex. Hawkins extends Vernon Mountcastle’s theoretical proposal by arguing that the human neocortex comprises approximately 150,000 cortical columns that are nearly identical. The distinguishing factor between these columns is not their intrinsic function, but rather the connections they form. If a specific region of cortical columns is linked to the eyes, the result is the provision of vision. If another region of cortical columns is connected to other language regions, such as Wernicke’s area and Broca’s area, it serves the functions of language comprehension and language production. This versatile functionality of cortical columns resembles the versatility of the neural networks in machine learning. Just as cortical columns can adapt based on their sensory input connections, neural networks can be trained on different data types to perform different tasks, from computer vision to text generation. If you feed image data and train image-recognition neural network algorithms, you get the computer vision function. If you feed text data and train the transformers, you get text-generative artificial intelligence. This versatility and adaptability might explain why neural networks are such a powerful, general-purpose technology.

Prediction and Learning Mechanism

A particularly innovative aspect of Hawkins’ theory is his conceptualisation of the neocortex as an instantaneous prediction machine. These predictions are made for every sensory modality, including low-level vision and high-level concepts. In order to make a prediction, the neocortex must have a model on which to base this prediction. Each instance of sensory interaction entails the continuous creation of a model, the making of a prediction, and the implementation of updates. To illustrate, the brain possesses a cognitive model of an object, such as a stapler. The model encompasses the visual representation of the stapler, its chromatic attributes, the location where it is situated, the tactile sensations it elicits, and the auditory cues associated with its utilisation. When there is a change in the location of the stapler, for example if it is misplaced from its usual desk drawer, the brain will promptly register this and update its model accordingly, indicating that the stapler is no longer present. In the event of the stapler malfunctioning, the model of the stapler is updated to reflect this new information, indicating that the stapler is no longer functioning correctly.

This process of prediction and model update is analogous to the backpropagation algorithm, which permits neural networks to learn from errors and adjust their parameters in order to make more accurate predictions. The crucial distinction is that humans develop these intricate world models through experience and movement, whereas machines learn through data training.

Location and Reference Frames

Another insight of his theory is the role of location/position data in understanding the world. He posits that the brain requires two fundamental pieces of information in order to construct models of the world and make predictions: the identity of the object and its location. In order to identify the location, the brain requires the presence of reference frames. Reference frames may be conceived of as invisible coordinates (x, y, and z axes), which permit the specification of the locations of objects. The generation of reference frames by each cortical column enables the brain to specify relative positions and structural relationships with respect to observed objects. Furthermore, reference frames can be utilised to organise knowledge of entities that cannot be directly sensed. Hawkins illustrates how one can visualise a DNA molecule in order to organise the knowledge about it. Furthermore, he illustrates this concept with the memory technique known as the “method of loci,” which refers to the idea that recalling information becomes easier when it is mentally attached to familiar locations.

However, the use of reference frames does not require the presence of physical anchoring. The reference frames associated with concepts may possess a greater number of dimensions than those pertaining to physical objects. While the location of an object, such as a coffee cup, can be specified in relation to its surroundings, a theoretical concept may have reference frames that are multidimensional. This possibility permits an understanding of knowledge representation in super high-dimensional spaces. This theory bears a curious resemblance to the manner in which large language models (LLMs) store concepts in high-dimensional spaces. The location data attached to the objects/concepts is of particular importance for the Transformer Model, which requires two pieces of information about each word: the meaning of the word and its position in the sequence.

What constitutes intelligence?

In Part II, Hawkins presents a discussion of the nature of intelligence and consciousness in the context of machine intelligence. He employs a methodology that may be regarded as somewhat traditional, and which calls into question the prevailing approach within the AI industry. While contemporary AI enterprises quantify intelligence through benchmark achievements and particular human-like capabilities, Hawkins proposes a more principled definition of AI. In his view, the mere ability to outperform humans in specific tasks does not constitute a sufficient criterion for defining true intelligence. For example, he notes that despite achieving victories against the leading human players, deep learning networks such as AlphaGo lack the capacity to be considered truly “intelligent” in the traditional sense. This is because they are unable to contextualise their understanding of Go within the broader historical and cultural significance of the game. Additionally, Hawkins posits that true artificial general intelligence (AGI) should adhere to foundational principles, including continuous learning, learning through movement, the utilisation of multiple models, and the employment of reference frames for knowledge storage.

Hawkins also offered a critique of two primary discourses surrounding the existential threats posed by machine intelligence. The first threat is the “intelligence explosion” scenario, in which advanced AI may subjugate or eliminate humans when machines’ capabilities surpass those of humans and humans are no longer as useful. He posits that this scenario is improbable and that machine learning may not be as rapid due to the necessity for physical interaction with the world. However, in contrast to his hypothesis, the replacement of humans by machines may not be the result of technological superiority, but rather of economic efficiency. As he notes in Part III, human perceptions can be erroneous. Should employers and capital investors perceive machines as more effective and efficient, the replacement of the human workforce will become inevitable. It is possible that an intelligence explosion may be driven by market forces rather than the genuine supremacy of machine intelligence.

The second existential threat, namely goal misalignment, whereby AI might pursue objectives contrary to human well-being, is dismissed by Hawkins as groundless. He posits that machines are devoid of intrinsic motivation and cannot self-replicate in the absence of human intervention. Nevertheless, although machines may not develop motivations independently, humans can and will use machine intelligence strategically. This may entail implanting motivations and desires into machine intelligence with the objective of eliminating human competition.

In Part III, Hawkins examines a fundamental constraint on human cognition. The human brain constructs internal models of the external world. Such models may not necessarily correspond to the external reality. In this regard, Hawkins’ perspective is analogous to that of Walter Lippmann. The model is updated when the prediction made by the model does not correspond with the moment-to-moment sensory input that is directly experienced. It is possible for individuals to make different choices, maintaining a false model when the knowledge in question cannot be directly verified through experience. In the event of encountering information that is inconsistent with their existing cognitive structures, some individuals tend to disregard such contradictory data rather than modifying their existing mental models. Those who adhere to the belief in a flat Earth are a case in point, rejecting any evidence that they are unable to directly sense or comprehend. Hawkins suggests that individuals should proactively seek evidence that could refute their beliefs, given the inherent cognitive limitations.

Conclusion

In his book, A Thousand Brains, David J. Hawking presents a compelling theory of intelligence that bridges neuroscience and artificial intelligence. The Thousand Brains theory, as outlined by Hawking, helps us understand biological human intelligence and machine intelligence. Furthermore, it encourages readers to contemplate the nature of intelligence, consciousness, and learning. Despite the rapid technological advances in AI since the book’s publication, Hawking’s insights continue to provide a thoughtful framework for understanding the ongoing evolution of intelligent systems.

Article information Continued