Table of Links
Abstract and Introduction
-
Extents and ways in which AI has been inspired by understanding of the brain
1.1 Computational models
1.2 Artificial Neural Networks
-
Embodiment of conscious processing: hierarchy and parallelism of nested levels of organization
-
Evolution: from brain architecture to culture
3.1 Genetic basis and epigenetic development of the brain
3.2 AI and evolution: consequences for artificial consciousness
-
Spontaneous activity and creativity
-
Conscious vs non-conscious processing in the brain, or res cogitans vs res extensa
-
AI consciousness and social interaction challenge rational thinking and language
Conclusion, Acknowledgments, and References
1.2 Artificial Neural Networks
Historically, from the early times of AI, Pitts and McCulloch referred to networks of idealized artificial neurons (McCulloch & Pitts, 1943). The connectionist program was subsequently negatively affected by Marvin Minsky’s seminal critique (Minsky & Papert, 2017) to then gradually come back until the explosion in use and popularity of Artificial Neural Networks (ANNs) in recent times (LeCun, Bengio, & Hinton, 2015). Decades of computer research developed on this basis moving from simplistic to complex architecture, from a single to multiple layers of artificial neurons – from the perceptron up to deep learning (LeCun et al., 2015) and the billions of parameters of Large Language Models (LMMs) (e.g., ChatGTP). Not only the symbolic approach (“Good Old-Fashioned AI” or GOFAI, including the “Logic Theorist” created by Newell and Simon), which was prevalent at the beginning of AI research and that aimed to reproduce the logical aspects of intelligence at a high functional level while neglecting the underlying brain mechanisms, but ultimately also the ANNs program does not fully account for the complexity of brain architecture (Moulin-Frier et al., 2016), if any reference to it is included.
\
In fact, the notion of neuron referred to in ANNs (i.e., a mathematical function which models biological neurons) is much simpler than its biological counterpart. For instance, basal and apical dendritic compartments of cortical pyramidal neurons connect to feedforward and feedback information processing respectively (Aru, Suzuki, & Larkum, 2020). This complexity intrinsic to biological neurons is partly taken into account in most modern ANNs, even if structured neurons for AI applications is an active field of research (Haider et al., 2021; Max et al., 2023; Senn et al., 2023). Other characteristics of the brain, like the role of GABAergic interneurons and the modulation by neurotransmitters like acetylcholine and dopamine (Changeux & Lou, 2011), or the capability of pyramidal cells to have two functionally distinct sets of inputs (one about which the neuron transmits information, and one that can selectively amplify that transmission when it is useful to do so in the context of information being transmitted by other neurons) (Phillips, 2023) are translated into large scale brain models like functional patterns (Eliasmith et al., 2012; Humphries, Khamassi, & Gurney, 2012). As mentioned above with reference to the multiple realizability thesis, the question is whether this functional emulation captures the right elements, processes and properties of the world to produce an empirically adequate description of the target object (e.g., the brain) and on this basis to possess its same properties (e.g., the same causal effect of the neuronal underpinnings of human features, like conscious processing). In other words, the question is whether it is possible to reproduce the features of real world objects, like the brain, in computing systems, including those built on digital electronic principles, different from the analog biological principles of the brain.
\
The recent developments of AI (e.g., ChatGPT, Sora, Dall-E, Stable Diffusion) illustrate further the extent to which such a basically algorithmic approach can be successful, even if the discussion is open about its limitation (Mitchell, 2023; Mitchell & Krakauer, 2023; Shanahan, 2024; Shanahan, Crosby, Beyret, & Cheke, 2021). Similarly, for years, computational “functionalist” descriptions of cognitive processes deliberately excluding any reference to the human brain were abundantly produced. Again, this connects to the question about multiple realizability of cognitive processes: can the same outcome be achieved with completely different biological, cognitive, or computational mechanisms? (Melis & Raihani, 2023)). Are we facing analogy rather than homology? The challenge is whether a formal algorithmic based computational approach, up to now successful in a number of applications, in particular the so-called neuromorphic hardware (Petrovici et al., 2014; Poo, 2018), will ever be sufficient to approach human-like conscious processing, an alternative form of conscious processing, or eventually no consciousness at all (Kleiner, 2024).
\
:::info
Authors:
(1) Michele Farisco, Centre for Research Ethics and Bioethics, Department of Public Health and Caring Sciences, Uppsala University, Uppsala, Sweden and Biogem, Biology and Molecular Genetics Institute, Ariano Irpino (AV), Italy;
(2) Kathinka Evers, Centre for Research Ethics and Bioethics, Department of Public Health and Caring Sciences, Uppsala University, Uppsala, Sweden;
(3) Jean-Pierre Changeux, Neuroscience Department, Institut Pasteur and Collège de France Paris, France.
:::
:::info
This paper is available on arxiv under CC BY 4.0 DEED license.
:::
\
