Summary and Contributions: This paper is concerned with the automated discovery of interesting patterns in morphogenetic systems. It introduces the idea of meta-diversity search and shows that a dynamic and modular neural architecture allows for a more controllable search for diversity. =============== After rebuttal ================== The rebuttal addresses part of my issues and the new overview figure should make it easier for the reader to follow. Therefore I'm changing my score from a 7 to an 8. However, for a strong accept, I think the authors would need to apply their approach to at least one other (ideally functional) domain. I'm very much looking forward to the next steps in this research direction.
Strengths: The work introduces important new methods to controlling self-organizing morphogenetic systems. It combines a variety of different techniques such as exploratory search methods, the automatic and incremental discovery of different behavioral characterisation spaces, and model architecture. The results show that the system allows to quickly adapt the type of created diversity to a simulated end-user.
Weaknesses: Because of the diversity of different techniques, it was initially difficult to understand the paper's contributions and how everything fits together. The abstract and introduction could be updated to further help the reader understanding the paper's most important contributions. Additionally, a more high-level overview figure than the one presented in Fig. 1 could help to convey how all the different parts fit together. The second weakness is that the system is currently only applied to one particular domain. It would be very interesting to see how the approach would scale to a domain that requires functional content, such as the self-organization of robot morphologies.
Correctness: The claims made in the paper appear correct.
Clarity: The introduction could make the important contributions of the paper and how the different parts fit together more clear.
Relation to Prior Work: Prior work is described in succinct but in sufficient detail. Work on involving user-feedback, such as interactive evolution, could be added as well, if space permits.
Summary and Contributions: This paper gives interesting evidence for the need of modularity for better exploration and diversity in goal-driven systems. In particular, the papers shows that a single neural network is worse than a dynamic+modular architecture of sub-networks on diversity metrics. The paper formulates the notion of meta diversity, a search algorithm to hierarchically (tree structured) build modules that route decisions to either create new networks to handle incoming observations or decode it outcomes. A continuous version of the game-of-life system is used (Lenia) as the experimental platform. In this system, different starting states, rules and interventions lead to vastly different outcomes. The diversity metrics are measured on such outcomes. Additionally a human user can guide the system's diversity in particular directions by intervening on the decision boundary between different sub modules. This enables a subjective notion to the exploration process. ---------- After rebuttal ------------ The rebuttal clarifies most of the concerns raised during the review process. For this reason I am increasing my score for this paper.
Strengths: This paper studies a very interesting, complex and dynamical test system with a large and intuitive space for exploration. The outcomes from such a dynamical system is combinatorial but has an ecological and intuitive basis -- making it an interesting test bed to study the influence of subjective interventions on exploration. This paper has also interesting implications for continual learning of neural networks in the reinforcement learning context. This paper is related to the progressive neural networks paper but goes beyond it by learning a tree-structured hierarchy of modules. It would have been interesting to see how the structure between modules affects exploration outcomes. How would the original progressive neural networks like topology for the modules work?
Weaknesses: "Because β-VAE can poorly represent high-frequency patterns, another variant trained on cropped patches is proposed" - patch beta VAE is introduced without any explanation or motivation in the main text How does this method compare to all other species that have been discovered in Lenia by others methods or manual interventions? Is there a metric to evaluate novelty? Why is a monolithic VAE better for diversity in some cases but for others? "when their existence remained an open question raised in the original Lenia paper" - is there a quantitative novelty metric to validate this claim?
Correctness: The claims and methods look accurate and reasonable.
Clarity: - The text to explain how the goal space interacts with Sec 3.1 can be more simply explained. It would be very helpful to pictorially show the exploration process to get better/faster intuition - Section 3.2 is dense and having simple intuitive explanations for each step would go a long way. For instance, 1) how is the parameter mutation process done, 2) why are particular goal sampling strategies chosen over others and what is the general space here? - The paper is currently dense and hard to follow. The underlying proposal and experiments are very interesting but it is difficult to understand due to a large number of unexplained choices and moving parts.
Relation to Prior Work: I would like to better understand what kind of diversity other approaches have produced in the Lenia system.
Summary and Contributions: The paper proposes a novel way of exploratory search that combines hierarchical clustering with goal-based intrinsic motivation techniques in a learned latent space for finding a diverse set of morphogenetic systems. The automation of diversity-driven discovery is an important area of research as human-based manual tuning is too expensive and slow. The novel systems [e.g. 50] tries to mitigate it by introducing human-made representation or learned representation (e.g. with VAE) and automate diversity-driven discovery in this space. The authors proposed to make one more step of automation and learn representations mapping that are diverse calling this meta-diversity search and showed that this search leads to diverse clusters of patterns. Main contributions: - The novel objective of meta-diversity search and architecture suitable for learning organized (possibly guided by human preference) hierarchical representations. - Empirical evaluation of diversity of the discovered patterns and possible guidance by a human. === after rebutal === Thank you for the informative response. No major concerns remain from my side.
Strengths: - Novel formulation of the diversity that aims to discover diverse clusters of patters. - Empirical evaluation of the previously proposed approaches to show that they discover patterns that are diverse only in one behavioral characterization (BC). - As the task is not fully specified without human preferences, the authors proposed to use human guidance to search patterns that are more interesting for people.
Weaknesses: - As the main novelty lies in the clustering of different patterns, the usage of K-means in embedding space G could be explained better. Maybe also a discussion of other possible choices and their comparison would be helpful. - As several previous approaches were also restricted only to Continuous Game of Life, another application could make the main topic of diversity search (and meta-diversity search) more task-independent. For example, it is not clear how useful the same clustering approach would help in different complex dynamical systems.
Clarity: The paper is well-written, however, the main novel clustering part (3.1) of the paper should be explained and described better.
Relation to Prior Work: yes
Additional Feedback: The webpage with the additional images and explorer for all the generated patterns is really nice.