I study how the brain might work.
How do you describe the novel in terms of the familiar, in such a way that proves useful?
This question comes up in a lot of AI domains. If a vision system can describe a novel object as an arrangement of familiar parts, then it can predict what that object will look like from other viewpoints, and hence can comprehend new objects from a few quick observations. If a robot can describe a novel environment using elements of other environments, then its behaviors can be informed by previous experiences from those other environments.
Long-term, I want to understand how the brain solves this problem. In the shorter term, I want to understand the bag of tricks that the brain plausibly might use.
One essential tool from machine learning is distributed representations. By learning to map inputs to a high-dimensional vector, artificial neural networks learn a function for describing novelty, and these output vectors can prove useful for classification, prediction over time, or behavior generation. Another potentially essential tool is taking input and describing it using a variable-sized graph, where each node and edge has its own distributed representation, and each node signifies a part. This latter approach is powerful but requires more engineering. What is possible with these tools? Are other tools needed?
This set of questions is my playground.
Some "Causal Inference" intuition
Grid cells: Visualizing the CAN model
Appendix: The classic TP
The Life and Times of a Dendrite Segment
The column SDR that wasn't random enough
See your HTM run: Stacks of time series
HTM time series: Now add synapse learning
HTM time series: Two charts, one scale
HTM time series: Number of dendrite segments
How many coin-flips till heads?
Three snapshots of Temporal Pooling
Data flow: Visualizing big remote things
Using Bézier curves as easing functions