**7. Conclusions and Future Work**

The info-computational approach, developed by the author, with natural morphological computation as a basis, is used to approach learning and learning to learn in humans, other living organisms, and intelligent machines. This paper is a contribution to epistemology of the philosophy of nature, proposing a new perspective on the learning process, both in artificial information processing systems such as robots and AI systems, and in natural information processing systems like living organisms.

Morphological computation is proposed as a mechanism of learning and meta-learning, necessary for connecting the pre-symbolic (pre-conscious) with the symbolic (conscious) information processing. In the framework of info-computational nature, morphological computation is information (re)structuring through computational processes which follow (implement) physical laws. It is grounded in the notion of agency, with causality represented by morphological computation.

Morphology is the central idea in understanding of the connection between computation (morphological/morphogenetical) and information. Morphology refers to form, shape, and structure. Materials represent morphology on the underlying level of organization. For the arrangements of molecular and atomic structures, material are protons, neutrons, and electrons on the level below.

Morphological computation, represented as information communication between agents/actors of the Hewitt actor model, distributed in space, where computational devices communicate asynchronously and the entire computation is generally not in any well-defined state [3]. Unlike Turing computation, which is a mathematical–logical model, Hewitt computation is a physical model. For morphological computing as information (re)structuring through computational processes which follow (implement) physical laws, Hewitt computation provides consequent formalization. On the basic level, morphological computation is natural computation in which physical objects perform computation. Symbol manipulation in this case is physical object manipulation, in the sense of Brooks "the world is its own best model". It becomes relevant in robotics and deep learning that manage direct behavior of an agen<sup>t</sup> in the physical world.

In morphological computation, cognition is the restructuring of an agen<sup>t</sup> through the interaction with the world, so all living organisms possess some degree of cognition. As a result of evolution, increasingly complex living organisms arise from the simple ones, that are able to survive and adapt to their environment. It means they are able to register inputs (data) from the environment, to structure those into information, and in more developed organisms into knowledge. The evolutionary advantage of using structured, component-based approaches is improving response-time and e fficiency of cognitive processes of an organism, which drives the development from organisms with learning on the System 1 level, to the ones that acquire System 2 capabilities on top of it. In more complex cognitive agents, knowledge is built upon not only reaction to input information, but also on internal information processing with intentional choices, dependent on value systems stored and organized in agents' memory.

Knowledge generation places information and computation (communication) in focus, as information and its processing are essential structural and dynamic elements which characterize structuring of input data (data → information → knowledge → metaknowledge) by an interactive computational process going on in the agen<sup>t</sup> during the adaptive interplay with the environment.

In nature, through the process of evolution and development, living systems learn to survive and thrive in their environment. Interactions present forms of reinforcement learning or Hebbian learning that make previous successful strategies preferred in the future [70]. That happens on a variety of levels of organization. On the meta-level, the meta-morphological computing (as a Sloman's virtual machine) [96] governs learning to learn.

In the case of human learning, the brain as a network of computational agents processes information obtained through the embodied communication with the environment as well as internal information from the body. Consciousness is a process of integration of information in the brain [35], and it gets a huge amount of data/information that would be overwhelming for the brain to handle in real time, so it uses the mechanism of attention to focus on a specific subset of information, typically regarding agent-based processes in the world. There changes in the scene are the consequence of the agent's interactions, and they are the unfolding of physical processes of morphological computations. Causality, or rather stable correlations between structures and processes in the world (from an agent's perspective) follow from what humans learn/memorize, as they ge<sup>t</sup> organized internally through the Hebbian principles where neurons that fire together, wire together.

Sloman, who developed theory of Meta-morphogenesis [74], started with the idea that changes in individual development and learning of an agen<sup>t</sup> produce new forms of information processing [74]. His approach o ffers new insight, that variation is algorithmic. The interplay between structure and process is essential for learning, as past experiences stored in structures a ffect the possibility of future processes and strategies of learning and learning to learn. To Sloman's morphogenetic approach, I would propose to add that steps in variation are results of morphological computation, which means physical computation, capable of, e.g., modifying genes, and executing morphological programs which do not present smooth incremental changes, but jumps in properties of structures and processes. Morphological computation acts also through gene regulation, which is one more process that was unknown to both Darwin and to proponents of evolution as Modern Synthesis.

Since contemporary deep-learning-centered AI (dealing with human-level cognition and above) is gradually developing from the present state System 1 (connectionist, sub-symbolic) coverage towards the System 2 (symbolic), with agency, causality, consciousness, and attention as mechanisms of learning and meta-learning [5,114], it searches for mechanisms of transition between two systems. An inspiration for technology development, the human brain is of interest as the center of learning in humans, that is self-organized, resilient, fault tolerant, plastic, computationally powerful, and energetically efficient. In its development, like in the past, deep learning is inspired by nature, assimilating ideas from neuroscience, cognitive science, biology, and more. The AI approach to understanding, via decomposition and construction, is close to other computational models of nature in that it seeks testable and applicable models, based on data and information processing. Bengio's proposal of agent-based perspective [5], necessary to proceed from System 1 to System 2 learning, can be related to the model of learning based on morphological computing.

For the future, more interdisciplinary/crossdisciplinary/transdisciplinary work remains to be done as a way to increase understanding of connections between the low level and the high level cognitive processes, learning, and meta-learning. It will also be instructive to find relations between (levels of/degrees of) cognition and consciousness as mechanisms helping to reduce the number of variables that are manipulated by an agen<sup>t</sup> for the purpose of perception, reasoning, decision-making, planning acting/agency, and learning.

The goals of artificial intelligence, as well as robotics, di ffer from those of the computing nature and morphological computing. AI builds solutions for practical problems and in that it typically focuses on the highest possible level of intelligence, even though among the AI fields inspired by computing nature, there is developmental robotics, which has more explorative character.

The priority of info-computational naturalism is understanding and connecting knowledge about nature, while a lot of current technology is searching for inspiration in nature in pursuit of new technological solutions. Paths of the two are meeting, and mutual exchange of ideas is beneficial for both sides. Specialist sciences and philosophies also need close communication and exchange of ideas. Learning and meta-learning within computing nature is such a topic of central importance that calls

for more knowledge from a variety of fields. This paper is not only the presentation of how much that is already known, but also an attempt to indicate how much more remains to be done.

**Funding:** This research is funded by Swedish Research Council, VR gran<sup>t</sup> MORCOM@COG.

**Acknowledgments:** The author would like to thank anonymous reviewers for very helpful, constructive, and instructive review comments.

**Conflicts of Interest:** The author declares no conflict of interest. The funders had no role in the design of the study, in the writing of the manuscript, or in the decision to publish the results.
