3. AI Technology Stack Flashcards
What is an AI Platform?
An AI platform is software that allows an organization to develop, test, deploy and refresh AI applications.
Platforms can:
-Centralize data analysis
- Streamline development and productions workflows
- Facilitate collaboration
- Automate systems-development tasks
- Monitor models and systems in production
Examples: Google Cloud Platform, Microsoft Azure, Amazon Web Services
What are common uses of AI applications?
- E-commerce
- Education
- Health care
- Autonomous vehicles
- Navigation
- Facial recognition
- Robotics
- Human resources
- Marketing
- Social media
- Chatbots
- Finance
What are common AI models?
- Linear and statistical models
- decision trees
- Machine learning models
- Neural networks
* Computer vision model
* Speech recognition model
* Language models
* Reinforcement learning models - Robotics
Describe Linear and Statistical Models:
- Models the relationship between two variables (ex. how sales of a product are related to changes in pricing based on historical data)
- Linear statistical models are not black box algorithm and more explainable
Describe the Decision Tree Models:
- Predicts an outcome based on a flowchart of questions and answers
- Explainable and not a black box
- Disadvantage: changing the training data (even in a small way) can significantly impact the algorithm; subject to security attacks and hacks
Describe the Machine Learning Models:
- Have black box capabilities
- Have a lack of transparency and explainability
- Neural networks (based on the human brain) * Contain nodes, like neurons, in a layered structure and continuously improve the
ability to find the right answer - Do not need to be trained to make complex nonlinear inferences in unstructured
data - Commonly behind technology, such as facial recognition
Give examples of Neural Networks:
- Computer vision models: used to recognize images in videos
- Speech recognition models: used in products like Alexa, transcription software (analyze speech across factors such as pitch, tone, language, and accent)
- Language models: natural language processing; allow computers to understand human language using machine learning, deep learning models and linguistics (used to process and respond to large amounts of communications data ex. customer service chatbots)
- Reinforcement learning models: train models to optimize their actions within a given environment to achieve a specific goal
- Guided by feedback mechanisms of rewards and penalties
- Conducted through trial and error; interactions or simulated experiences that do not require external data. (Ex. an algorithm trained to earn a high score in a video
game by having its efforts evaluated and rated according to success towards the
goal. - Disadvantage: lack of explainability and transparency
Explain Robotics:
- Multidisciplinary field encompassing the design, construction operation and programing of robotics
- Allows AI systems and software to interact with the physical world without human
intervention. (Ex. Roomba using machine learning to navigate a building)
How does technology stack propose challenges to AI and has it driven AI to the heights that we
see today?
The algorithmic innovation has been one of the true advances in pushing forward
Started with phenotypic and image data capture systems in response to genomic research
and the accumulation of data:
* Supervised data
* Data collected in interactive environments
* Structured and unstructured data
What are the main areas of AI infrastructure?
- Compute
- Storage
- Network and software development
How does compute infrastructure advance AI?
- GPUs have thrust the AI movement forward. They have specialized chips that offload from
CPUs.
* Provide better performance and match to algorithmic advances
* Better in matching hardware to the AI model for optimal performance - Serverless: not limited to a particular server or piece of hardware; running code on multiple
hardware devices, providing two important services or functionalities for AI:
* Loose coupling: taking data from a variety of sources
* Scalability: running multiple instances of the code because it’s not tied to a given server, which helps drive AI forward - High performance compute: create isolated clusters of compute power; high-speed networking, specialized chipsets
4.Trusted execution environments: good from a privacy perspective for AI, as human influence is taken out of the equation
What is Quantum Computing?
Processing data in three dimensions horizontally and vertically.
What are the 4 general stages of AI storage?
- Ingestion
- Preparation
- Training
- Output (inference)
Each stage has different storage requirements that must be adhered to in order to avoid project failure.
What are storage infrastructure considerations?
- Expense of a storage solution for massive amounts of data
- Different storage for a variety of storage types (file, object, image, etc.). (Each require different storage subsystems and may also affect expenses)
- Storage types for structured vs. unstructured data
* Easier to process structured data than unstructured data
* AI must be done at scale; flexible storage allows the ability to do an AI at scale
How does network infrastructure advance AI?
- High-speed networks needed to support AI models: complex AI models, deep learning
models, natural language processing, large language models like ChatGPT - Deliver to training data in time to the AI algorithm as well as training and inference at scale
through high-speed networks - High-performance compute. Underlying infrastructure is housed in the same data centers, usually in the same rack and connected via fiber connections
- Edge computing: the Internet of Things (It is estimated that within the next three to five years,
each individual will have five connected devices) - Communication or network protocols: based on a congestion-free design, especially for larger language models and neural networks
- Transmission control protocol (standard). (However, this requires a packet to be sent)
