This glossary contains the most common terms used in the NIML model. For more insight into specific parameter values and functions, refer to our API documentation.
| SDR, iSDR, oSDR | Sparse Distributed Representation--the vector containing the encoded pattern. The prefix "i" indicates an input SDR created by encoding data, and the prefix "o" indicates an output SDR created once an observation has been processed by the NPU |
| Set bits | A set of positions within an SDR where the value at this position is "on" or has a "1" |
| sdr_set_bits | The total number of set bits contained in an iSDR |
| sdr_width | The total number of positions in an iSDR in it's expanded format. This can be computed as set_bits / sparsity * num_features |
| sparsity | The number of set bits divided by the total sdr width. |
| NPU | Neural Processing Unit. This is comprised of digital neurons that are exposed to the inputs iSDRs from the data and learn and pool together similar inputs |
| synapse | The positional "link" between an iSDR and a neuron |
| input_pct | Input percentage: The proportion of locations in a neuron where a synapse is initialized. |
| SSV | Synaptic Strength Value. The numeric value that increases and decreases as the synapse becomes more or less strong |
| connected synapse | a synapse whose strength value is above a given threshold. Typically this synapse will have strengthened during training as it positionally aligns with certain types of iSDR patterns |
| disconnected synapse | a synapse whose strength value is below a given threshold. Typically this synapse will have weakened during training as it doesn't positionally align with most iSDR patterns |
| non-synapse | a position in a neuron where no synapse is placed due to that position not being selected when the neuron was initialized |
|
Neuron |
The individual learning components within the NPU. Each neuron has a unique configuration of synaptic connections, and is selected to learn based on whether it is naturally closely aligned with any of the input patterns |
| active neurons | The group of neurons that most closely align to a given iSDR. These are the neurons whose synapses will strengthen and weaken for a given input |
|
F34 Classifier |
The proprietary NIS classifier. This classifier functions well for smaller datasets and in scenarios where anomaly detection is key |
| pfv | Positional field vector. An expanded representation of SDRs that includes all the "0" positions as well as the "1" positions |
| Boosting | A mechanism to keep neurons from being over or under active as the system trains. Neurons are boosted if they are underactive and suppressed if they are overactive to discourage premature convergence |