Module 1: Foundations of AI: AI Technology Stack: Infrastructure Flashcards
What are the 4 main areas of the AI Technology Stack?
1) Compute
2) Storage
3) Network
4) Software
What are the 2 main elements of Compute?
1) CPUs - Central Processing Units
2) GPUs - Graphical Processing Units
What is the key to AI performance?
Match hardware to AI model requirements.
What is the definition of “serverless”?
Not limited to a particular server or piece of hardware.
What is the definition of “loose coupling”?
The ability to take data from a variety of sources - not limited to a particular network storage or a physical file share.
Describe how AI is scalable.
It enables you to run multiple instances of code without being tied to a particular server.
What is “high-performance compute”?
An isolated cluster of compute power. It includes high speed networking and specialized chip sets.
What is “quantum computing”?
Data can be processed in 3 dimensions.
What are “trusted execution environments”?
Humans are not in the loop.
What are the 4 general stages of AI?
1) Ingestion
2) Preparation
3) Training
4) Output
Name some storage considerations for AI.
- Expense of storage solution
- Different storage types (file, object, image, etc.)
- Structured vs. unstructured data (structured data is much easier to process)
- Flexibility (in order to do AI at scale)
What aspects of networks are important to AI models?
- High speed network
- Understanding that different AI data types have different performance requirements. For example, an image file will take longer than a text file to get to the training algorithm.
- Protocols based on a congestion-free design.
What is edge computing?
Edge computing is a distributed computing paradigm that brings computation and data storage closer to the sources of data, so that a user of a cloud application is likely to be physically closer to a server than if all servers were in one place. This is meant to make applications faster.
Example: The Internet of Things (IoT)
Name some software implications for AI.
- Democratization of AI
- Tuning: Tuning an AI system allows you to customize models
- Transformation
What are some tuning challenges?
- Number of hyperparameters
- Varies on model type and complexity
- Scaling trial and error tuning
Why does data need to be transformed for an AI model?
For data compatibility.
Methods of data transformation.
- Internal to the model
- External to the model (pre-processing)
- Output (post-processing)
What is labeling?
Enriching data to use it for deployment, training and tuning.
What are some considerations for labeling?
- Data labels must be of a high quality and standard. Validation of the data quality must be performed.
What are hyperparameters?
Parameters or values adjusted for an AI model or algorithm to tune it toward desired outcomes. The adjustment is mostly accomplished through trial and error.
List some common hyperparameters.
- Learning rate
- Number of epochs (one cycle through the full training dataset)
- Momentum (the amount of history included in the equation)
What is AI observability?
Monitoring the AI algorithm and metrics of the AI system.
Why are Data and AI observability key to the success of any AI project?
They provide indices and metrics for performance, in-depth analysis of AI data and models, and the capability to investigate, resolve and prevent AI model issues.
What are some challenges with AI observability and monitoring?
- Data integrity
- Data drift
- Bias and discrimination
How can AI observability and monitoring challenges be mitigated?
Outcome validation
What is a benefit of open source AI frameworks?
They give companies greater freedom to innovate with AI.
