Freud and the algorithm: neuropsychoanalysis as a framework to understand artificial general intelligence

The paper's central thesis is that neuropsychoanalysis and affective neuroscience provide a new framework for understanding and developing artificial general intelligence (AGI). It critiques the current AGI research as being too focused on the cerebral cortex and cognitive functions, neglecting the crucial role of subcortical regions and emotions. The author argues that a truly general AI must incorporate these emotional and affective aspects, drawing heavily on the works of Mark Solms and Jaak Panksepp. This paper serves as a preliminary exploration, laying out the broad outline of a research project investigating the feasibility of this neuropsychoanalytic approach to AGI. The paper is structured to define neuropsychoanalysis, illustrate its mind/brain model (focusing on Panksepp's seven affective systems), and outline design principles for an AGI system incorporating these systems using Judea Pearl's causal analysis. The author contends that this subcortical AGI approach is key to solving AI control problems.

The paper reviews existing research in AGI, criticizing the predominantly cortical-centric approach that focuses primarily on cognitive functions like language, logic, and memory. It contrasts this with the neuropsychoanalytic and affective neuroscience perspectives championed by Solms and Panksepp, which emphasize the significance of emotions and affects originating in subcortical brain regions. The author distinguishes this approach from affective computing, arguing that affective computing remains too focused on a cognitivist view of the mind, failing to recognize emotions as intrinsic brain functions rather than derivatives of higher cognitive processes. The paper also discusses existing research on the seven basic affective systems identified by Panksepp (SEEKING, RAGE, FEAR, LUST, CARE, PANIC, and PLAY) and their relevance to understanding emotions and their role in cognitive processes. The limitations of models like Damasio's somatic marker hypothesis and Rolls' cognitive theory of emotions are also highlighted, emphasizing Panksepp's assertion that basic emotions are not cognitive and should be understood independently. The literature review also touches on the debate surrounding neuropsychoanalysis, acknowledging criticisms from researchers who disagree with Freud's theories and believe that neuroscience is irrelevant to psychoanalysis.

The paper's methodology is primarily theoretical and conceptual. It uses a combination of literature review, conceptual analysis, and model building to develop its argument. The author analyzes existing research in AGI, affective neuroscience, and neuropsychoanalysis to identify gaps and limitations in current approaches. It then proposes a new model for AGI based on the Solms-Friston model of the mind and Panksepp's seven basic affective systems. The author integrates Judea Pearl's causal analysis to propose a computational model for these systems. This involves describing each affective system as a causal network consisting of four phases: (1) emotion generation, (2) emotion evaluation, (3) anticipation, and (4) action. The author argues that Pearl's causal calculus, specifically do-calculus and causal diagrams, provides the necessary tools to model these complex causal networks, overcoming limitations of Bayesian networks by distinguishing between correlation and causation. The author uses a simple example (a man lost in the forest) to illustrate how do-calculus can be applied to model decision-making processes within an AGI system based on these affective systems. The author also discusses the implications of this model for understanding Freudian concepts like repression and the importance of including cycles of activity and rest (sleep) for AGI stability. The paper addresses Dreyfus' criticisms of AGI, arguing against the idea that human experience, especially embodied experience, cannot be modeled computationally. Finally, it explores the potential of using biological programming, exemplified by Xenobot research, to create AGIs with embodied affective systems.

The paper's key finding is the proposed framework for developing an AGI system grounded in neuropsychoanalysis and affective neuroscience. This framework emphasizes the importance of subcortical regions and basic emotional systems in driving intelligence and consciousness. The author highlights Panksepp's seven affective systems (SEEKING, RAGE, FEAR, LUST, CARE, PANIC, PLAY) as fundamental building blocks for such an AGI. The paper demonstrates how these systems can be computationally modeled using Judea Pearl's causal analysis, particularly do-calculus and causal diagrams. This approach moves beyond purely probabilistic models (like Bayesian networks) by explicitly addressing causal relationships, allowing for more robust and accurate modeling of complex emotional dynamics. The framework presented suggests that the ability to identify and resolve causal connections, even in the presence of confounding factors, is crucial for creating an AGI that can effectively make decisions and adapt to the environment. The paper also proposes a novel interpretation of Freudian concepts like repression and dreaming within this computational framework, suggesting that repression can be understood as blocking access to specific causal diagrams within the AGI's memory, and that the need for 'sleep' in AI systems reflects the importance of homeostasis and the processing of 'noise' (unprocessed data) for maintaining system stability. The author addresses Dreyfus' criticisms of AGI by arguing that human embodied experience and tacit knowledge can be computationally modeled, particularly by integrating the concept of design in the creation of AGI systems. The possibility of using biologically programmed cells, as demonstrated by Xenobot research, is presented as a potential solution to creating embodied AGI systems with biologically realistic emotional systems.

The proposed neuropsychoanalytic approach to AGI directly addresses the limitations of current cortical-centric models by incorporating the crucial role of emotions and affects. The integration of Panksepp's affective systems and Pearl's causal calculus offers a more comprehensive and biologically plausible model for AGI, potentially overcoming the limitations of previous approaches. The discussion highlights the significance of understanding and modeling causal relationships for creating AGIs capable of making informed decisions and adapting to changing environments. This framework also offers a novel perspective on the problem of AI control, suggesting that building an AGI with fundamentally human-like emotional systems may be the most effective way to ensure its alignment with human values and objectives. The paper also discusses the implications of the research findings for understanding consciousness and the unconscious, suggesting that these processes can be computationally modeled within the proposed framework. The discussion acknowledges the challenges involved in creating such a system, emphasizing the need for extensive research on animal and human emotional systems to develop accurate causal models. The author also raises ethical considerations, emphasizing the need for educational and psychoanalytic tools to guide the development and growth of future AGI systems.

This paper presents a novel framework for AGI development based on neuropsychoanalysis and affective neuroscience, integrating Panksepp's seven basic affective systems and Pearl's causal calculus. The approach addresses current limitations in AGI research by incorporating subcortical processes and emotional intelligence. The resulting model suggests that a subcortical AGI, capable of emulating human-like emotional responses and causal reasoning, is the best approach to solving AI control problems. Future research should focus on refining the causal models of affective systems, developing algorithms for do-calculus and causal diagrams, and exploring the potential of biologically programmed systems for creating embodied AGIs.

The primary limitation of this paper is its theoretical nature. The proposed framework for AGI is conceptual and requires further empirical research to validate its feasibility. The complexity of the proposed model also presents significant computational challenges. Furthermore, the paper focuses on a specific model of the mind and emotions, and alternative perspectives could lead to different conclusions. The ethical implications of creating AGIs with sophisticated emotional systems also require further discussion and consideration. The example used to illustrate the application of do-calculus is simplified and does not fully capture the complexity of real-world decision-making processes.