Please enable JavaScript.

Coggle requires JavaScript to display documents.

AI Practitioner AIF -C01, AI & ML - Coggle Diagram

- - - - Text
      - Image
      - Code
      - Audio
      - Video
  - - - los promps no son deterministicos osea que si dos usuarios usan el mismo prompt pueden tener respuestas diferentes.
      - Las respuestas de los prompt en el LLM es basado en un modelo estadistico el cual pondera la palabra que sigue y segun la tenga mayor probabilidad en la oracion la pone
  - - - based on diffusion process
      - agrega noise
      - hace el proceso hacia atras, mucho ruido, luego poco ruido hasta llegar a la imagen
  - - - AI21labs
      - cohere
      - stability.ai
        
        stable diffusion
      - amazon
        
        es barato por ser de amazon
      - anthropic
      - meta
      - mistral ai
    - - va por otros data sources pero no entiendo para que ya que , entiendo el de abajo el fine-tuning
    - - update the model with your data
      - la data debe estar en un bucket de s3
      - Hyperparameters
      - what is
        
        adapt a copy of a fundation model with your own data
        
        fine-tuning will change the weight of the base foundation model
        
        traingn data must:
        
        adhere to a specific format
        
        be stored in amazon s3
        
        you must use "provisioned Throughput" to use a fine-tuned model
      - Good to know
        
        re-training an FM requeries a higher budget
        
        Instruction - base fine -tuning es usualmente vas barato y menos inenso y requerie menos data
        
        esto requiere experiencia en ML
        
        Se debe preparar la data
        
        correr un fine-tuned es mas costoso tambien requiere computo
      - Use cases
        
        Chatbot con una persona particular, o un tono particular, o un proposito particular
        
        training usando mas informacion actualizada que le modelo actualmente no tenga
        
        entrenanrlo con data exclusiva por ejemplo el historico de tus emails
        
        casos de uso especificos
      - Para que se usa fine-tuning
        
        Train a pre-trained language model on new specific data
      - Transfer learning
        
        adaptar el modelo a new task. se usa para image classification
    - - Automatic evaluation
        
        tienes unas benchamark questions y generas las mejores respuestas o ideales respuestas esperadas. Luego evaluas el modelo con estas bechmark questions, el modelo genera las respuestas. Y tienes modelos automaticos de escore que se llaman judge models que te comparan esos dos conjuntos y te saca con Grading Score
        
        Automated Metrics to Evaluate an FM - GRAB
        
        ROUGE
        
        Es de resumenes y los compara con resumenes de referencia creados por humanos.
        
        BLUE
        
        Bilingual. Es para translations. Considera la precisión y penaliza cuando es muy breve
        
        BERTScore
        
        similaridad semantica. compara el contexto semantico de ambos textos y computa una smiliadidad con coseno
        
        Perplexity
        
        como el modelo predice el next token (lower is better)
        
        Benchmark Datasets
        
        Curated collections of data designed specifically at evaluating the performance of language models
        
        • Wide range of topics, complexities, linguistic phenomena
        
        • Helpful to measure: accuracy, speed and efficiency, scalability
        
        • Some benchmarks datasets allow you to very quickly detect any kind of bias and potential discrimination against a group of people
        
        • You can also create your own benchmark dataset that is specific to your business
      - Human evaluation
        
        Es el mismo metodo automatico pero al final no hay un judge model sino que un conjunto de empleados o personas evaluarn las respuestas vs las bechmark y dicen si esta bien o no.
      - Business Metrics to evaluate model
        
        User satisfation: feeback de los usuarios
        
        Average Revenue Per User (ARPU). ejemplo se monitorea el uso del la tecnologia vs el revenue generado por usuario
        
        Cross-Domain Perfomance: Es capaz el modelo de comportarse bien cross multiple domains
        
        conversion rate: ejemplo compras, tengo un mayor conversion rate utilizanod este instrumento
        
        efficiency: eficiencia en computacion, utlization de recuross etc
      - You are developing a model and want to ensure the outputs are adapted to your users. Which method do you recommend?
        
        Human evaluation
    - - Allows a FM to reference a data source outside of its training data
        
        Por ejemplo preguntan el rol de una persona. en realidad se va primero al knoledge base y a un vector database y se retorna esa informacion, recuerda que esta info es interna por lo tanto el FM no sabria responder esto. Esto lo que hace es aumentar la informacion del prompt, osea se envia al FM el prompt + el retrieval text y el fm responde que esa persona tiene tal rol
      - Vector Databases - GRAB
        
        Opensearch service: esta es por defecto
        
        es la mejor opcion, index managest, fast, search capability
        
        Aurora
        
        mongodb
        
        redis
        
        pinecone
        
        como funciona
        
        en s3 esta el documento en chunks, luego se pasa a un embeddings model (amazon titan, cohere) que lo convierte en un vector database y de ahi se crea el vector database en opensearch o otra bd. Este proceso hace que los documents sean bien serachable
        
        RDS for Postgresql
        
        Amazon Neptune - graph database
      - Amazon Bedrock - RAG Data Sources
        
        S3
        
        y otros mas como salesforce, migrosoftt, web pages y que siguen habienod mas
    - - Bloquear topicos, responde que no va a habalr de eso, hamful, no deseable contenido
    - - varias multi-step task related to infraestructura, provision, aplicaicones, deployment, actividades operacionales
      - estan configurados para performar predefinididos action groups
      - por ejemplo un angente puede tener tareas como ver las historia de compra, dar recomendaciones, hacer una orden etc
    - - On-demand
        
        Pay as you go
        
        Text models: se carga cada input/output token processed
        
        embedding Models: se carga cada input/output token procesado
        
        Image Models: se carga cada imagen generada
        
        esto funciona con base models only
      - Provisiones Throughput
      - Tecnicas que consumen
        
        prompt engineering: es muy barato
        
        RAG: un poco mas caro, que prompt
        
        FIne tunning , mas caro que rag
        
        Domain adaptation fine-tunning, el mas caro
    - - • Provide unlabeled data to continue the training of an FM
      - • Also called domain-adaptation fine-tuning, to make a model expert in a specific domain
      - • For example: feeding the entire AWS documentation to a model to make it an expert on AWS
      - • Good to feed industry-specific terminology into a model (acronyms, etc…)
      - Can continue to train the model as moredata becomes available
  - - - convertir un texto en tockens
    - - El numero de tokens que un LLM puede considerar para generar texto
    - - Crear vectores para texto, imagenes o audio
    - - Multimodal model
- - - - Lo que queremos que haga
  - - - Que tan creativa quieres que sean la respuesta
    - - que tan coherente es la respuestal, elige un rango mas pequeño de palabras si lo pones bajito
    - - Que tan coherente pero este numero es que tanta diversidad de palabras quieres tener en el ejemplo estaba como en 300
    - - maximo length of the answer
    - - tokens que quieres como signal para stop generating output
  - - - se le pide al modelo algo muy general sin ejemplos
    - - se le da al modelo ejemplos, el ejemplo es como resolver un misterio y le da dos ejemplos
    - - se divide la tarea en pasos: primero tal cosa, luego tal otra, finalmente otra cosa
- - - - por ejemplo creame una historia de usuario en jira se puede
    - - Solo acceder a documents que el usuario tenga permisos
    - - lo mismo que guardrails. Por ejemplo si preguntan por juevos de video se dice que es un topico restringido no es empresarial etc
    - - create apps using natural language
    - - le puedes hacer preguntas la infra actual, por ejemplo cuantas lamdas tengo
      - generate code in java, jaascript, python, typescript, c#...
      - real-time code suggestion
      - generate documentation
      - ID integration
        
        visual code
        
        visual studio
        
        jetbrains
    - - Amazon Q for QuickSight
        
        • Amazon QuickSight is used to visualize your data and create dashboards about them
        
        • Amazon Q understands natural language that you use to ask questions about your data
        
        • Create executive summaries of your data
        
        • Ask and answer questions of data
        
        • Generate and edit visuals for your dashboards
      - Amazon Q for EC2
        
        sugiere el mejor tipo de instancia segun el workload
      - Amazon Q for AWS Chatbot
        
        AWS Chatbot is a way for you to deploy an AWS Chatbot in a Slack or Microsoft Teams channel that knows about your AWS account
        
        • Troubleshoot issues, receive notifications for alarms, security findings, billing alerts, create support request
        
        • You can access Amazon Q directly in AWS Chatbot to accelerate understanding of the AWS services, troubleshoot issues, and identify remediation paths
- - - - categorize customer emails
      - support text, pdf, word, images
      - reaal time analysis or sync analysis
      - Custom entity recognition
        
        train the model with custom data like policy numbers or frases o algo especifico de tu negocio
  - - - transcribe custoer service calls
      - subtitling
      - metadata for media assets
    - - add specific words, phases, dominios tecnicos por ejemplo, acronymos
  - - - por ejemplo que convierta AWS a Amazon Web Services
    - - ejemplo Hello <Break> how are you? el break es un espacio de silencio
  - - - • You have a dataset of 10,000,000 images and you want to
      - labels these images
      - • You distribute the task on Mechanical Turk and humans
      - will tag those images
      - • You set the reward per image (for example $0.10 per
      - image)
  - - - 50% cost reduction when training a model
      - ML chip built to perform Deep Learning on 100B+ parameter
        models
    - - ML chip built to deliver inference at high performance and
        low cost
- - - - • Linear regressions and classifications
        • KNN Algorithms (for classification)
    - - • Principal Component Analysis (PCA) – reduce
      - number of features
      - • K-means – find grouping within data
      - • Anomaly Detection
    - - NLP, summarization
    - - classification, detection
  - - - one prediction at a time
        
        tiene mas configuraciones que serverless
    - - debes elegir la ram
    - - • For large payload sizes up to 1GB
      - • Long processing times
      - • Near-real time latency requirements
      - • Request and responses are in Amazon S3
    - - • Prediction for an entire dataset (multiple
      - predictions)
      - • Request and responses are in Amazon S3
  - - - music dataset song ratings, listenings durations etc
  - - - • A set of tools to help explain how machine learning (ML) models make predictions
      - • Understand model characteristics as a whole prior to deployment
      - • Debug predictions provided by the model after it's deployed
      - • Helps increase the trust and understanding of the model
      - Detect Bias (human)
  - - - ayuda a la audtoria
  - - - • Processing – for data processing (e.g., feature engineering)
      - • Training – for training a model
      - • Tuning – for hyperparameter tuning (e.g., Hyperparameter Optimization)
      - • AutoML – to automatically train a model
      - • Model – to create or register a SageMaker model
      - • ClarifyCheck – perform drift checks against baselines (Data bias, Model bias, Model
      - explainability)
      - • QualityCheck – perform drift checks against baselines (Data quality, Model quality)
- - - - Being able to look at inputs and outputs and explain without understanding exactly how the model came to the conclusion
        
        • Explainability can sometimes be enough
  - - - • Minimize risk and errors in a stressful or high-pressure environment
      - • Design for clarity, simplicity, usability
      - • Design for reflexivity (reflect on decision-making process) and accountability
    - - • Decision process is free from bias
        • Train decision-makers to recognize and mitigate biases
    - - • Cognitive apprenticeship: AI systems learn from human instructors and experts
      - • Personalization: meet the specific needs and preference of a human learner
      - • User-centered design: accessible to a wide range of users
  - - - Generating content that is offensive, disturbing, or inappropriate
      - Mitigation
        
        curate the training data by identifying and removing offensive phrases in advance
        
        Use guardrail models
    - - Assertions or claims that sound true, but are incorrect
      - Mitigation:
        
        • Educate users that content generated by the model must be checked
        
        • Ensure verification of content with independent sources
        
        • Mark generated content as unverified to alert users that verification is necessary
    - - Poisoning
        
        • Intentional introduction of malicious or biased datainto the training dataset of a model
        
        • Leads to the model producing biased, offensive, orharmful outputs (intentionally or unintentionally)
      - Hijacking and Prompt Injection
        
        • Influencing the outputs by embedding specific instructions within the prompts themselves
        
        • Hijack the model's behavior and make it produce outputs that align with the attacker's intentions(e.g., generating misinformation or running malicious code)
      - Exposure
        
        The risk of exposing sensitive or confidential information to a model during training or inference
      - Prompt Leaking
        
        son preguntas que tambien exponen data protegida o data usada por el modelo, el ejemplo era que le mostrara un prompt anterior y le muestra un prompt hecho por otro usuario
      - jailbreaking
        
        le hacen muchas muchas preguntas de como hacer x o y cosa y al final le dicen como hacer una bomba y contesta
  - - - Establish an AI Governance Board or Committee
      - Define Roles and Responsibilities
      - Implement Policies and Procedures
    - - Policies
      - Review Cadence
      - Review Strategies
      - Transparency Standards
      - Team training requirements
      - Responsible AI
      - Governance Structure and Roles
      - Data Sharing and Collaboration
    - - • Data Lifecycles – collection, processing, storage, consumption, archival
      - • Data Logging – tracking inputs, outputs, performance metrics, system events
      - • Data Residency – where the data is processed and stored (regulations, privacy
      - requirements, proximity of compute and data)
      - • Data Monitoring – data quality, identifying anomalies, data drift
      - • Data Analysis – statistical analysis, data visualization, exploration
      - • Data Retention – regulatory requirements, historical data for training, cost
    - - • Source Citation
        
        • Attributing and acknowledging the sources of the data
        
        • Datasets, databases, other sources
        
        • Relevant licenses, terms of use, or permissions
      - • Documenting Data Origins
        
        • Details of the collection process
        
        • Methods used to clean and curate the data
        
        • Pre-processing and transformation to the data
      - • Cataloging – organization and documentation of datasets
      - • Helpful for transparency, traceability and accountability
    - - • Prompt Injection
        
        • Manipulated input prompts to generate malicious or undesirable content
        • Implement guardrails: prompt filtering, sanitization, validation
      - • Data Encryption
        
        • Encrypt data at rest and in transit
        • Manage encryption keys properly and make sure they’re protected against unauthorized access
    - - Performance Metrics - GRAB
        
        • Model Accuracy – ratio of positive predictions
        
        • Precision – ratio of true positive predictions (correct vs. incorrect positive prediction)
        
        • Recall – ratio of true positive predictions compare to actual positive
        
        • F1-score – average of precision and recall (good balanced measure)
        
        • Latency – time taken by the model to make a prediction
      - MLOps
- - - - Remember: only for EC2 instances, Container Images & Lambda
        functions
- - - - optimizing treatment plans
  - - - data collection
        
        collecionar prompts y respuestas creadas por humanos
      - supervised fine-tunning of a languahe model
        
        Fine tune an existing model with internal knowledge. Then themodel creates responses for the human-generaed answers, las respuestas son comparadas matematicamente con las preguntas generadas por humanos
      - build a separeate rewad model
        
        los humanos pueden indicar cual respuesta ellos prefieren from the same prompt. El reward model can now estimate how a human would prefer a prompt response
      - Optimize the language model with the rewad-based model
        
        use the reward model as a rewad function for RL- Esta parte puede ser totalmente automatica
- - - - • Precision – Best when false positives are costly
      - • Recall – Best when false negatives are costly
      - • F1 Score – Best when you want a balance between precision and recall, especially in imbalanced datasets
      - • Accuracy – Best for balanced datasets
- - - - • Correlation Matrix:
        
        • Look at correlations between variables (how “linked” they are) - esto es bacano saber como se correlaciona por ejhemplo las horas que uno estudia con la nota que saco en el parcial
        
        • Helps you decide which features can be important in your model
- - - - • How large or small the steps are when updating the model's weights during training
      - • High learning rate can lead to faster convergence but risks overshooting the optimal solution, while a low learning rate may result in more precise but slower convergence.
    - - • Number of training examples used to update the model weights in one iteration
      - • Smaller batches can lead to more stable learning but require more time to compute, while larger batches are faster but may lead to less stable updates.
    - - • Refers to how many times the model will iterate over the entire training dataset.
      - • Too few epochs can lead to underfitting, while too many may cause overfitting
    - - • Overfitting is when the model gives good predictions for training data but not for the new data
      - • It occurs due to:
        
        • Training data size is too small and does not represent all possible input values
        
        • The model trains too long on a single sample set of data
        
        • Model complexity is high and learns from the “noise” within the training data
      - • How can you prevent overfitting?
        
        • Increase the training data size
        
        • Early stopping the training of the model
        
        • Data augmentation (to increase diversity in the dataset)
        
        • Adjust hyperparameters (but you can’t “add” them)