Latency Flashcards
Latency
Refers to the time delay between the input being fed into the NN and the output being produced (inference).
7 factors that affect latency
- Number of layers in NN
- Number of parameters
(These both increase computation time) - The type of hardware used.
- The memory bandwidth
- Size of batch (the larger the batch, higher the latency)
- Data transfer speed between storage, memory, and processing units.
- Specifically in distributed systems, the transfer speed between different nodes on a network
______ & _______ can _______ latency by using their __________ processing capabilities
TPUs, GPUs, reduce, parallel
Batch size
The number of data points from the training dataset processed together in a single iteration before updating the model’s parameters.
4 factors in how the training dataset can affect latency
- If the training dataset is very large, it may lead to a complex model with many parameters, increasing latency
-The noisier the data, the more post-processing steps required to filter low-quality predictions, increasing latency
- Lack of diversity in the dataset will mean the model struggles with unfamiliar inputs, potentially requiring additional processing so will increase latency
- Complex features such as complex patterns and dependencies can require more computational resources and therefore, increase latency
Problem of Machine Learning Dependencies in feeding data
This is an issue in machine learning that arises from the fact that when feeding data from data sources into the machine learning model, there are a series of components/processes the data must travel through before reaching the model, and we cannot do these things in parallel. This creates latency.
These components/processes include:
Data source –> Ingestion/Integrated –> Data store -> Cleaning and pre-processing –> Machine Learning Model
Critical Path Algorithm
This is used to mitigate the problem of machine learning dependencies when feeding data.
It is essentially a project management tool that will identify the longest sequence of dependent tasks (the critical path), and help in optimizing it by identifying bottlenecks to produce the lowest completion time for a project.
Natural Language Understanding (NLU)
Structured sequence of machine learning models and algorithms that work together to gradually transform and understand user input.
Each step in the sequence, referred to as the pipeline, contributed to a deeper linguistic understanding of the input, allowing the chatbot to generate accurate and contextually appropriate responses.
Natural Language Understanding 3 advantages
- Accuracy. Each process in the NLU pipeline enhances understanding of the input data leading to a better output
- Efficiency. Since the pipeline is modular, each section can be individually focused on for optimization.
- Adaptability. The modular nature of the pipeline allows for easy updates to individual components without having to make drastic changes to the entire system.