ProtoDash - Fast Interpretable Prototype Selection by Karthik Gurumoorthy

Explore a 32-minute conference talk on ProtoDash, a fast and interpretable prototype selection method for machine learning. Learn about extracting compact representations of large datasets, understanding prototypes and criticisms, and their applications in causal reasoning. Dive into the mathematical foundations, including set function maximization and submodularity properties. Compare ProtoDash and ProtoGreedy algorithms, examining their performance in classification accuracy, sparsity, and computation time. Discover how to visualize selected prototypes and criticisms, and explore real-world applications such as finding relationships between datasets and improving prediction accuracy through training data selection. Gain insights into theoretical guarantees and engage with the Q&A session to deepen your understanding of this innovative approach to interpretable machine learning.

ProtoDash: Fast Interpretable Prototype Selection
Objective: Extract compact synapses of large data sets
Prototypes and Criticisms
Some applications:
Causal reasoning
Metric to quantify best for prototype selection
Reformulation as maximizing a set function
Properties of set function: Submodularity
Weak submodularity
ProtoDash algorithm: Greedily build the set based on gradients
ProtoGreedy algorithm: Greedily build the set based on maximum function increment
ProtoDash Vs ProtoGreedy
Classification accuracy Vs sparsity
Classification accuracy Vs skew
Prototype selection quality
Computation time
Visualizing selected prototypes
Visualizing selected prototypes from the same data set
Theoretical guarantees
Set function value Vs sparsity
Criticism selection
Greedy algorithm for choosing criticisms
Visualizing selected criticisms
Application 2: Finding relationships between data sets
Approach to identify relationships
Visualizing relationships
Human expert based evaluation
Application 3: Improving prediction accuracy by training data selection
Training data selection
Prediction accuracy
Interpreting selected prototype
Interpreting selected prototypes
Thank You
Q&A