ML Model Performance, Optimization, & Regularization

Regularization

• The main goal of regularization techniques like L1 and L2 is to reduce model complexity.
• Strategies to reduce complexity include increasing learning rate or dataset size.

Loss Functions

• Squared Error, Cross Entropy, and Huber Loss are appropriate loss functions for regression problems.

Optimization Algorithms

• Stochastic Gradient Descent(SGD) is used with Log Loss or reducing dimensionality.
• Momentum is used with Huber Loss or calculating feature importance.

Model Performance

• Overfitting means the model performs well on training data but poorly on test data.
• The curse of dimensionality arises from having too many dimensions, causing overfitting.
• Activation functions introduce non-linearity, optimize the loss function, and reduces overfitting.
• Sigmoid is for binary classification tasks.
• ReLU is most appropriate for multi-classification tasks.

ROC Curve

• An ROC curve measures True Positive Rate vs. False Positive Rate.

Optimization Protocols

• Stochastic Gradient Descent is an algorithm used to find the global minimum in optimization problems.
• Techniques to prevent overfitting include increasing bias or changing data preprocessing.

Ensemble Methods: Bagging vs. Boosting

• The primary goal of hyperparameter tuning is to improve model performance.
• Bagging reduces variance, while boosting reduces bias.

Transfer Learning

• Transfer learning involves reusing pre-trained models on new tasks.
• PCA(Principal Component Analysis) is a technique for reducing data dimensionality.

Gradient Issues

• Gradient explosion refers to gradients becoming very large.
• The vanishing gradient problem occurs in deep learning when gradients become very small and Batch Normalization addresses the small gradients.

Batch Normalization

• Batch normalization stabilizes the learning process in neural networks.

Performance Metrics

• ROC-AUC (Area Under the Receiver Operating Characteristic Curve) is ideal for imbalanced datasets.

Supervised Learning

• Linear Regression is a supervised learning algorithm.

Clustering

• The purpose of clustering is to group similar data points together.
• K-means is a type of unsupervised learning technique used for clustering.
• PCA is used for dimensionality reduction.

Optimization Problems

• Gradient descent is used for regression problems.
• Loss function controls adding more information to the speed of learning.
• A local minimum in gradient descent is a suboptimal solution.

Cost Function

• The cost function in gradient descent measures prediction error.

Gradient Descent Types

• Batch gradient descent updates weights using the entire dataset
• Batch is different from stochastic gradient descent because it uses entire datasets rather than subsets of data.

Supervised vs Unsupervised Learning

• Supervised learning uses labeled data, while unsupervised learning uses unlabeled data.

Objectives of Gradient Descent

• The objective of gradient descent is to minimize the cost.
• A learning rate that is too high can cause the cost function to oscillate or the model may diverge.

Convergence

• “Convergence” in gradient descent means the cost function reaches a minimum.
• The learning rate controls the step size of the algorithm.

Underfitting vs Overfitting

• Overfitting is when a model learns the training data too well leading to poor data outcomes on new data.

Data Dimensions

• PCA reduces the dimensions of a dataset.
• PCA address overfitting.

Unsupervised Learning

• The main objective of unsupervised learning is to discover hidden patterns.
• K-means algorithm is primarily used for clustering.

Activation Functions

• The activation function introduces non-linearity.
• If the model has high bias and low variance, a Decision Tree works well and is appropriate for both regression and classification tasks.

Feature Scaling

• Feature scaling normalizes feature ranges and balances class distributions.

Epochs

• An "epoch" in machine learning is one iteration over the entire dataset.

Gradient Descent Implementation

• Feature Scaling is done by minimizing learning rate, adding regularization, feature complexity, and calculating change.

Dimensionality Reduction

• PCA reduces data dimensions.
• Feature scaling balances features, finds clusters, and increases variance.

Min-Max Normalization

• Min-max normalization scales data to a range between 0 and 1.
• For dealing with missing values: Imputation finds null values.

One-Hot Encoding

• The purpose of one-hot encoding is to convert categorical data.

Feature Extraction

• Feature extraction creates new features, combines features, and normalizes labels.

Data Standardization

• Standardization in preprocessing involves scaling data to zero & unit variance, split data, reduce noise and removes labels.

Tokenization

• Tokenization splits text into tokens, removes stopwords, and encodes text.

Data Preprocessing

• The main goal of data preprocessing is to improve model performance.

Data Scaling

• Scaling data finds clusters and normalizes weights.
• The data preprocessing technique used for handling missing data is imputation (1, 0).

Normalization

• The purpose of data normalization is to adjust feature ranges to (0, 1.0).
• Handling categorical features uses one-hot encoding.

Principal Component Analysis (PCA)

• PCA is a preprocessing step that reduces multicollinearity in a dataset by resampling or data standardization.

AI Agents

• The main purpose of an LLM agent is to combine LLM capabilities with reasoning and action.
• A ReAct agent focuses on using LLM-based reasoning and action.
• Agents handle tool selection using either random selection or machine learning.
• Agent memory helps maintain context to speed up processing.
• Action Plan handles agent planning.
• Agents handle task decomposition by breaking down tasks into subtasks.
• Tool augmentation helps gather additional information about actions.

LangChain(s)

• Chunking is implemented in LangChain using Recursive Character Text Splitter.
• Chroma is used for vector storage in Retrieval-Augmented Generation (RAG).

Coding Selection

• Fixed size chunks Splitter is selected with Recursive Character Text Splitter.
• FAISS is a library used for similarity search and implements prompt templates using string formatting.

Similarity in FAISS

• Similarity search (used with index.search()).

Temperature Settings

• Temperature in LLM calls is implemented using a temperature config.
• The max_tokens parameter improves output quality and determines the quantity of words in responses.

Rate Limiting

• Exponential backoff is for retry logic in LLM coding or using throttling.

LLM Response Streaming

• Streaming in LLM responses aims to increase speed but not to reduce costs.

Parameter Efficient Fine-Tuning (PEFT)

• PEFT is a parameter efficient fine-tuning that enhances existing models and not a new model architecture.
• Low-Rank Adaptation is used in LoRA.

QLORA

• It reduces catastrophic forgetting using training algorithms.
• Prompt tuning serves to increase model accuracy, especially in large LLM training.
• Adapter tuning works by changing trainable parameters but not by replacing entire layers.

Few-Shot Learning

• Few-shot learning refers to learning with full dataset training (no training required).

Layer Normalization

• Layer normalization aims to improve inference and reduce memory.

Positional Encoding

• Positional encoding in transformers is used to capture sequential data ordering.

Transformer Components

• Feed-forward Network handles context understanding in a transformer.
• Multi-head attention allows for parallel processing of different parts of the input sequence.
• Residual connections aim to prevent gradient vanishing and increase inference.
• Decoder attention masking improves the functionality of all 3: multi-head attention, embedding layer, and output layer.

Fine-Tuned Models

• Instruction-tuned models exhibit improved task-following capabilities.
• Vision-language models can process both images and text.
• Base models are pre-trained foundation models with larger model size or enables convolutional use.
• Multimodal models enable converting text coding.
• Instruction-refined models are specific to data analysis.
• Code generation is made possible using a decoder-only model.
• Text-to-SQL models are specific to data analysis.

Tokenization

• Tokenization in NLP splits text into meaningful units (tokens).

NLP Techniques

• Word embedding is a more technique, not stemming or data encryption.
• Stop word removal eliminates words that add little meaning.
• TF-IDF (Term Frequency-Inverse Document Frequency) measures word importance.
• Byte Pair Encoding (BPE) breaks words adds prefixes that address vocabulary.

NLP Tools

• Named entity recognition identifies categories (e.g., people, organizations).
• Dependency parsing is used to analyze grammatical structure.

(RAG) Retrieval-Augmented Generation

• RAG generates text using external knowledge sources.
• Embedder is the element to find relevant documents in the process.

Vector Similarity

• Vector similarity search is used to speed up finding relevant documents and improves code functionality.
• BERT-base is a common embedding model in Retrieval-Augmented Generation RAG implementations.
• Context window includes combining keyword search.

Reranking

• Reranking is a method for improving evaluation quality.
• Claude includes all 3: limited modality, strong capacity, and small model.

Distinct LLM model from OpenAl

• GPT-4 includes audio processing and faster use.
• Anthropic’s Claude is recognized for strong reasoning and limited context.
• PaLM (Pathways Language Model) is distinguished by its efficient performance and single modality.
• Google focuses on functional models.
• Mistral AI models are characterized by efficient performance.

Blood Relations

• If A is the mother of B and C is the son of B, then A is C's grandmother. "He is the brother of my father(brother’s son)” means that it is Ravi himself.
• The phrase, "My father is your father's only son," indicates the speaker is the listener's child.
• If A is the mother of B, and C is the son of B's daughter, then the question is incomplete, as the relationship between A and C, requires that the entire statement be examined.

Family Trees

• The girl who is the speaker's father's brother's daughter is the speaker's cousin

Jan 1, 2001

• If January 1, 2001, was a Monday, then May 1, 2001, will be a Tuesday.

Clock and Calendar

• Clock hands overlap 11 times in 12 hours(22 in 24 hr).
• The angle between clock hands at 10:30 is 135 degrees.
• At 3:15 PM, the angle between the clock hands is 7.5 degrees.

APPLE = ZKKOV, ORANGE = LIZMTV

• If APPLE is coded as ZKKOV, then MANGO is coded as NZMTV.

BIRD coded

• If TREE is coded as GVIIV, FLOWER is coded as KOTDVN.
• If BIRD is coded as DKTF, then FISH is coded as HTUK.

Directional Reasoning

• A man starts 8m north, 6m west. Therefore, he has travelled 10 metres

• A person walks 10m north, 5m west, 6m south, and 5m east. Therefore, they are 4 metres north

• walks 12m north, 5m west, and 13m south. Distance from the start is 5

• walks 15m north, 10m east, 5m south, 10m west. The net distance from start is 10m N.

• faces east, turns 270 clockwise, 90 counterclockwise. Facing south

• walks 30m N, right 20m S Distance from start is 50m

• faces east, turns 135" clockwise 12m N/ 5m E. Shortest distance between the points is 24m E.

• turns right 5m, right again 10m The person does not go South, but West instead

• Shortest distance between the points is 24 metres

Logic tests

• If there 8 balls and one is heavier than the other, weigh balls against one another in order to determine the heaviest (3 trials).
• You are stuck on an island and only have a frying pan, you cannot drink the see water.
• There is no way to measure 45 minutes using two ropes that each burn in one hour.
• Bringing the goat across the river is the appropriate action.
• Touch for Heat is the correct option to identify the light-switch.
• It five machines can five widgets in five minutes, then 100 machines would make one widget in 100 minutes.
• If a farmer has 17 sheep, they would have 17 sheep since the farmer had none carried away.
• If the price of the bat + ball is one dollar more than of that of the cap, the correct cost is $0.05 cents less.
• A man spends 1/3 child, 1/6 teen, 1/2 adult then reaches 72 yrs, the teen years total 12
• Light up one switch is the correct option to identify the bulb-switch.
• Two trains are 120 km apart with speeds 50 & 70 km/hr while a bird flies between them has been used to express a long-distance.
• A man with 2 sons each having 2 children makes for 7 family members including the wife.
• A 3L jug is to be filled using a 5L, and both are initially empty, fill 5L fully.
• 12 cows graze a field in 10 days. Grazing days for 24 cows= 10 days
• The 45-degree angle is a right angle
• If 5 cats catch 5 mice in 5 minutes then 100 cats catch 100 mice in 5 minutes.

Number Series

• The missing value in 15,24, 33,40
• The missing value in 1, 3, 9, 27, 81, 243
• The missing term in 2, 6, 12, 20, 30, 42, 56
• The odd number in 64, 125, 216, 343, 512 is 41.
• The missing number i 5, 11, 23, 47, 95
• The odd one out: 49, 64, 81, 100, 121 or 83 is 83.
• If your team is unable to meet a deadline, Identity the reasons for the delay.
• When a colleague is not contributing to the report, Discuss the issue with the colleague.
• A new manager assigns additional work with an unrealistic deadline, Politely explain the challenge and negotiate.
• Confront or discuss a conflict between two colleagues respectfully to act as a mediator.

Assessment

• When you criticize yourself constructively Take it personally. Assess the feedback.
• A critical task a delayed, follow up and collaborate until issue is solved.
• A junior's mistake, help the partner learn.
• When over worked, Explain priorities to the manager and ask for help and prioritization
• Unfairly criticized, act professionally.
• A similar working style, try to find it.
• Vague work demand, ask questions.
• If a client wants it delivered in an unrealistic time, talk with your partner.

Basic Programming Algorithms

• Finding the samllest and unsorted means complexity algorithm.
• Dynamic programming is for overlapping and memory issues.
• Merge sort is the correct complexity.
• The traveling salesman = finding routes.
• The master theorem = codes and math.

Cache Coherence

• Invalid cache access explains false sharing.
• Faster queries is what Cors is primarily for.
• Event notifications for the Observer pattern.
• Separate interfaces via pattern.
• The inversion pattern is program flow.
• The strategy pattern is code.
• The facade pattern accesses code faster
• Microservice is a small program.
• eventsourcing = data tracking
• The interface is building design.
• Domain is a language.

Programming Strings

• gnimmargorp is programming reversed.

• The answer is syntax errors.

• The dining philosphers problem illustrated deadlock not coherance.

• Lock-free algorithms uses the ABA memory corrupting algorithm.

• Concurrent is running at the same time so paralleiism takes more sources.

Basic Coding

• Deadlocks are code locks and wait. Memory is limited (buffer).

• Semaphores = counting.

• Data = the tree, array, and graph and the log scale code.

• Key primary = query.

• Multiversion is database history.

• Substitution = all areas of memory.

Normal Tables

• Foreign = indexes.
• Sharing = organizes data
• Cohesion and coupling are opposites
• regression testing = repeating tests after release

Other programming principles

• Changing code = improving codes
• Automation is integrating and checking
• Test code is test-driven.
• Eventually = distributed
• Theorem = protocols
• Full memory is how garbage is collected

IP Addresses

• Mac = Media Access
• Slow start = protocols.
• Protocol = Transfer.
• Handshake = Connecting.
• Switching systems is a lot processes and can be hard like thrshing of systems.
• Kernel = a crashing state code.
• A compiler = translates codes.
• Object orientation = object oriented code.

Security

• No new coding is lazy behavior.

• Fibornacci is an antivious code.

• Attacks = social engineers

• Days = exploits

• Man in middle = communications

• Over flow = memory excessive

• Cross site is code remotes commands

• Infrastructure encryption codes.

• Out authentication protocols are coded

• Secret timing passwords.

• Cycomelic measured the test design.

English Grammar

• Know = does not.

• The cat... The cat is seeing the tree.

• Has = used for gone to the store

• Are = used for kids playing.

• Neither = Have

• Every = The teacher

• Doesnt= Not

• Last time = 6 years

• She = enjoy.

Communications

• Opposite = decline

• Enhanced = is better

• Correct plural form from the list = children

• He will eat = has

• Cold = this winter

• By the the ear = the teacher will go there

• She = has to

• Finish = They have

• If...= good at = Saw =

• Righting can be more... the stronger one= More

• Movie ... the best choice = bored/enjoued

• Siblings + dont / children = play

• The best option = Have'nt /since

Neither = were

• There is many people = The number of
• Right = Writes right
• Has'nt tired=
• Both the teachers
• Each of the boys have = Have
• The bell rang=
• To = would

Writing skills

= Would she would...=

• Correct uses can make your ideas sound as clear, and concise as possible.

• More well = better

• Insisted= Plain

• John= likes those

• Had the best for what I can follow?

• Wish she... Wish she were home.

• Always he tries... As he is the boss.

• Who never knows....

• Never you give what...

• Each of the what she could. Best as can. Follow

• "Bestest" = the best of two options

• Home and its doing it = DID.

• Began= to rain

Words

• The opposite = Compassion
• That is Ubiquitous? = everywhere
• Mebily= closest to sad
• Unique = plain
• Opposite = natural
• Chat = Not what you want
• Ambiguous = vague. is the best.