Useful references: Why Neurons Have Thousands of Synapses and Continous Online Sequence Learning with an Unsupervised Neural Network Model
- What is first order vs high order sequence memory? What is high order vs variable order sequence memory? Why do we need high order sequence memory?
- Can the temporal memory learn an arbitrary sequence of inputs?
- Once the temporal memory learns a sequence, how does it make predictions?
- How long of a sequence can the temporal memory learn? If you have 10,000 inputs into the spatial pooler, 500 columns, and a “reasonable” set of temporal memory parameters (choose some), can you learn a sequence that is 1000 elements long? What about 100,000 elements long?
- What is “bursting” and what does it accomplish?
- What is the difference between one cell per column vs multiple cells per column (in terms of what sequences you can learn)?
- Suppose a temporal memory has learned the following two sequences: ABCDE and ABCDF. You now present the sequence ABCD - what will be predicted next? What is the representation of the temporal memory at this point in time?
- Suppose a temporal memory has only learned the following two sequences: ABCDE and FGCDH. You now present the sequence FGCD - what will be predicted next? Suppose you present the sequence CD - what will be predicted next? What is the exact representation of the temporal memory at this point in time (predicted and active cells).
- What would be the difference between enforcing one distal segment per cell vs allowing multiple distal segments per cell?
- How does online learning happen in the temporal memory?
- What are the disadvantages of the temporal memory?