Conference Vision
“Everything should be made as simple as possible, but not any simpler.”
– Albert Einstein
One of the most fundamental reasons for the emergence of intelligence and science is that the world is not fully random: its observations, mechanisms, and laws exhibit structure and regularity. A central purpose of intelligence is therefore to discover parsimonious descriptions of those structures from data—models that capture what matters without reproducing every degree of freedom of the ambient world.
Over the past decade, machine learning and large-scale computing have transformed how we process, interpret, and predict with data in engineering and science. Classical approaches were often built around explicit structural models—sparsity, low rank, subspaces, manifolds, symmetry, latent variables—and optimization algorithms designed to exploit them. Modern data-driven systems instead learn representations and mechanisms at very large scale, often through pretraining followed by adaptation to many downstream tasks. Yet the success of both paradigms continues to depend on finding and exploiting simpler structure inside apparently high-dimensional problems.
For this reason, learning and compression are deeply connected. Structure may be assumed in advance, induced by an architecture or regularizer, discovered from data, or emerge implicitly from optimization dynamics. Parsimony can appear as sparsity, low intrinsic dimension, symmetry and invariance, modularity, compositionality, simple mechanisms, conditional computation, or other forms of compressibility. Understanding these structures can improve statistical efficiency, computation, memory, energy use, interpretability, robustness, and scientific understanding.
Several rich research directions have developed around this theme. Some seek theoretical explanations for feature and representation learning, generalization, identifiability, and implicit bias. Others design interpretable architectures that explicitly incorporate known structure. A complementary line studies pruning, quantization, distillation, low-rank adaptation, mixtures of experts, and hardware/software co-design for efficient learning. Scientific machine learning, inverse problems, neuroscience, robotics, and control provide settings in which the right structure can be as important as model scale.
The emergence of foundation, generative, multimodal, and agentic models makes these questions even more pressing. Modern systems operate in enormous parameter and data spaces, but their successful representations, reusable skills, reasoning patterns, and computational pathways may be far more structured than their raw dimensionality suggests. CPAL welcomes work that makes this connection explicit: what is the parsimonious structure, is it assumed, learned, or emergent, and what does it enable?
These communities have often developed in parallel despite sharing a common scientific purpose. CPAL is intended as a general forum where researchers in machine learning, applied mathematics, signal processing, statistics, optimization, intelligent systems, neuroscience, hardware, and the sciences can exchange ideas and work toward a modern theoretical and computational understanding of intelligence through the perspective of parsimony and learning.