Is a Modular Architecture Enough?

Abstract

Inspired from human cognition, machine learning systems are gradually revealing advantages of sparser and more modular architectures. Recent work demonstrates that not only do some modular architectures generalize well, but they also lead to better out-of-distribution generalization, scaling properties, learning speed, and interpretability. A key intuition behind the success of such systems is that the data generating system for most real-world settings is considered to consist of sparsely interacting parts, and endowing models with similar inductive biases will be helpful. However, the field has been lacking in a rigorous quantitative assessment of such systems because these real-world data distributions are complex and unknown. In this work, we provide a thorough assessment of common modular architectures, through the lens of simple and known modular data distributions. We highlight the benefits of modularity and sparsity and reveal insights on the challenges faced while optimizing modular systems. In doing so, we propose evaluation metrics that highlight the benefits of modularity, the regimes in which these benefits are substantial, as well as the sub-optimality of current end-to-end learned modular systems as opposed to their claimed potential. 1

Evolutionary trends in the use of artificial intelligence in support of architectural design

“…Soluzioni mirate a fornire più struttura alle reti neuronali sono state perseguite con interessanti risultati da Stanley attraverso l'applicazione di strategie evolutive (Stanley and Miikkulainen, 2002); ma questo indirizzo, per quanto molto promettente, non ha ancora generato una massa critica di ricerche affini. Tra gli esperti di deep learning è oggi diffusa la sensazione che l'avanzamento dello stato dell'arte possa avvenire, oltre che attraverso la creazione di maggiore struttura nelle reti (Mittal et al, 2022), attraverso una sintesi tra approcci "tradizionali", "simbolici" all'intelligenza artificiale, e approcci "connettivistici", sub-simbolici (Goyal et al, 2022). La medesima convinzione è anche diffusa tra studiosi operanti nel campo dei sistemi simbolici: non a caso, per vari di questi sistemi sono stati, negli anni, creati moduli sensoriali basati sull'impiego di reti neuronali (Rosenbloom et al, 2016).…”

“…Solutions aimed at giving more structure to neural networks have been pursued with good results by Stanley through the application of evolutionary strategies (Stanley and Miikkulainen, 2002), but this very promising direction has not yet generated a critical mass of related research. Among the experts of deep learning, there is now a diffuse sensation that the advancement of the state of the art may entail, besides the creation of greater structure in artificial neural systems (Mittal et al, 2022), a synthesis between "traditional", "symbolic" approaches to artificial intelligence and connectivistic, sub-symbolic ones (Goyal et al, 2022). This belief has also matured in scholars operating in the field of symbolic systems for arti-ficial general intelligence.…”

Having “multiple selves” helps learning agents explore and adapt in complex changing worlds

“…The action of the agent is selected based on an arbitration of the individual module preferences. Such modular architectures have received increasing attention in RL (30)(31)(32)(33)(34)(35), and so has modularity in machine learning more broadly, because this 'divide-and-conquer' strategy tends to improve learning speed, scaling properties, and performance robustness (36).…”

Higher Cognition: A Mechanical Perspective

“…These approaches are a foundation for emulating the modularity and breadth of function in higher cognition. For achieving this aim, meta-learning methods can create a formal, modular [44], and structured framework for combining disparate sources of data. This scalable approach would lead to building complex information systems and reflect the higher cognitive processes [45,46].…”