The workshop will focus on three interconnected topics of particular relevance to Machine Learning:
Biological Inspiration and Inductive Biases: To what extent must human-level learning be innately constrained---through bias or biology---as opposed to more fully data-driven? Recognizing which biases humans bring to learning tasks could provide insights into more sample-efficient machine learning.
New Datasets for Comparing Humans and Machines: While machine learning algorithms and AI systems have continued to achieve human-level performance on a growing list of benchmark tasks, humans are still the gold standard in many respects. To understand the differences between human and machine strategies in more detail, large new datasets are needed to expose deeper representational and behavioral characteristics of humans.
Robustness and Generalization: What are the underlying mechanisms that make human perception invariant/robust to distortions, rotations, scaling, and occlusion? Identifying relevant mechanisms and representations may provide insights into human perception which consequently lead to more robust AI systems.
Data-Driven vs Biologically plausible Generative Models: As computational efforts continuously increase in producing highly realistic rendering samples from classical GAN or VAE model based approaches -- a next frontier is reducing both the sample complexity of the network architectures and amounts of training data via different inductive biases, that can also possibly mitigate dataset biases.