A quantitative measure of mode-seeking vs. mode-covering preferences
Another view
argmax(blog)
Recent content on argmax(blog)
Latest posts
Stein discrepancy is mode-seeking
Stein discrepancy is more mode-seeking than reverse KL
Sampler evaluation: No mode-covering without importance sampling
An impossibility result
Defining mode-covering behavior with gradient analysis
What does mode-seeking vs. mode-covering behavior in distributional divergences really mean?
Outranking leaderboard models on test sets, with only their predictions
A brainteaser
On diversity and many-model ensembling: AI government & AI-augmented public goods funding
When AI agents outnumber voters, how should we elect an AI government?
The deep retrieval + remixing hypothesis
How do modern deep generative models produce high-quality outputs?
What is data? The classical and postmodern views
Is data sacred?
On Creating AGI within Capitalism -- The Greatest Threat to AGI is ANI.
A speculative argument
Implicit regularization of SGD in a linear model
A brief mathematical result
Treating high-dimensional prediction/generation as optimization
Exploring connections between diffusion models, alphafold's recycling, structured prediction energy networks, and manifold learning.
Primer on Score-based Generative Models
An introduction
Implicit Differentiation through Equilibria
An introduction.
On KL Divergence in Discrete Spaces
KL behaves differently in discrete spaces than in continuous spaces.