Forem: Ashwani Kumar Shamlodhiya

Probability2: Bayes Theorem

Ashwani Kumar Shamlodhiya — Sat, 27 Sep 2025 18:52:12 +0000

Probability1 : Cheatsheet

Ashwani Kumar Shamlodhiya — Sat, 27 Sep 2025 18:45:41 +0000

Model Creation CNN : A simple example of tensorflow ( Partial dataset of cats vs dogs)

Ashwani Kumar Shamlodhiya — Tue, 02 Sep 2025 04:28:00 +0000

The code : https://github.com/ash322ash422/tut_tensorflow/blob/master/tf_tut3_model_creation_cats_vs_dogs_version1.ipynb

Model Creation CNN : A simple example of tensorflow on colab ( MNIST digit dataset)

Ashwani Kumar Shamlodhiya — Tue, 02 Sep 2025 04:24:43 +0000

The code: https://github.com/ash322ash422/tut_tensorflow/blob/master/tf_tut2_model_creation2_CNN_MNIST_Colab.ipynb

Model Creation ANN: A simple example of tensorflow ( Iris dataset)

Ashwani Kumar Shamlodhiya — Tue, 02 Sep 2025 04:19:27 +0000

The code is available here: https://github.com/ash322ash422/tut_tensorflow/blob/master/tf_tut2_model_creation1_2_ANN_classify_Iris3Class.ipynb

The dataset : https://github.com/ash322ash422/tut_tensorflow/blob/master/iris.csv

Model Creation ANN : A simple example of tensorflow (children - banana)

Ashwani Kumar Shamlodhiya — Tue, 02 Sep 2025 04:12:16 +0000

One of the most common uses of TensorFlow is building and training machine learning models. Let’s create a simple neural network model using TensorFlow's high-level API, Keras.

We'll build a simple model to predict y value for a given x value. I created CSV file where

children: denotes number of children

banana : denotes how many banana are eaten by children

This actually follows a simple relationship: banana = 1 + 2*children . ( That is y = 1 + 2x )

https://github.com/ash322ash422/tut_tensorflow/blob/master/tf_tut2_model_creation1_1_ANN_linearReg_oneVar_childBanana.ipynb

The data file: https://github.com/ash322ash422/tut_tensorflow/blob/master/child_banana.csv

AI - CNN : Convolution Neural Network (part 1 )

Ashwani Kumar Shamlodhiya — Tue, 02 Sep 2025 04:06:06 +0000

ANN 2: Key Terms BatchSize - Iterations - Epochs - LearningRate - LossFunction-Optimizer

Ashwani Kumar Shamlodhiya — Mon, 01 Sep 2025 17:28:10 +0000

ANN 1: Artificial Neural Network-Perceptron-MLP-Neuron-Activation Function-Deep Learning

Ashwani Kumar Shamlodhiya — Mon, 01 Sep 2025 16:53:12 +0000

Some interesting site
https://playground.tensorflow.org/

ML : Convert an image into a tabular data

Ashwani Kumar Shamlodhiya — Mon, 01 Sep 2025 16:37:07 +0000

ML Project Workflow

Ashwani Kumar Shamlodhiya — Mon, 01 Sep 2025 15:35:25 +0000

Data Acquisition / collection (page 2) https://www.kaggle.com/datasets https://archive.ics.uci.edu/datasets Ninja Dataset https://datasetsearch.research.google.com/ https://cloud.google.com/solutions/datasets https://www.kaggle.com/competitions (for competitions) https://ml-cheatsheet.readthedocs.io/en/latest/datasets.html https://catalog.data.gov/dataset/ https://www.kdnuggets.com/2020/03/20-machine-learning-datasets-project-ideas.html https://data.worldbank.org/ https://data.fivethirtyeight.com/ (Topics: Politics,Sports, Science+health, Economics, Culture) https://www.earthdata.nasa.gov/ Databricks : paid https://datarade.ai/data-categories/ai-ml-training-data (paid)

Biggest Data generator are Walmart, Amazon in Retail sector

The difference between the type of problems that are solved by traditional algorithm versus AI/ML: Part2

Ashwani Kumar Shamlodhiya — Sun, 31 Aug 2025 11:40:23 +0000

When do we need AI/ML algorithms ?
We need AI/ML approaches when:

The problem is too complex to explicitly program all possible scenarios
Example: Self-driving cars
You can’t manually code every possible road condition, pedestrian behavior, or traffic situation. ML models learn from vast driving data to handle unpredictable situations.
Patterns are subtle or hidden within vast amounts of data
Example: Detecting fraudulent transactions in banking
Fraud patterns are often rare, dynamic, and subtle — ML models can learn from millions of examples and detect anomalies that aren’t apparent to humans or rule-based systems.
Example: Predictive maintenance: Machines give off subtle signals (vibration, temperature, steady increase in current usage ,etc.) before failure. ML models can detect trends and predict breakdowns earlier than rule-based thresholds
Rules are difficult to articulate but examples are plentiful
Example: Facial recognition
We can’t write explicit rules to define every possible human face, but we can train a model on millions of labeled images to recognize faces effectively.
Example: Natural Language Processing (NLP): Understanding context, sentiment, and meaning in text. It's nearly impossible to write deterministic rules for sarcasm, sentiment, grammar, or slang
The problem requires prediction based on incomplete information
Example: Medical diagnosis based on symptoms and patient history
Symptoms may vary or be incomplete. ML models can make informed predictions based on patterns learned from prior patient data.
The environment is dynamic and continually changing
Example: Stock market trading algorithms
The market shifts constantly based on news, sentiment, and global events. Adaptive ML models can continuously retrain and adjust to changes better than static logic.
The solution needs to improve with experience
Example: Personalized recommendation systems (e.g., Netflix, Amazon)
The system gets better over time by learning user preferences and adjusting recommendations accordingly.
Approximation is acceptable and preferable to no solution
Example: Language translation (e.g., Google Translate)
It's often better to have an approximate translation than none at all, especially across languages where perfect translation is impossible.
Example: Speech recognition: Converting spoken language to text may be 90% accurate.

Real-World Examples Where AI/ML Beat Traditional Algorithms

Image Recognition and Computer Vision
Initial Challenge: "We need to identify objects in images for our security system."
Why ML Won: Traditional computer vision algorithms using edge detection and feature extraction failed at:
Handling variable lighting conditions
Recognizing objects from different angles
Dealing with partial occlusion
Generalizing across diverse environments
Real-World Success: Deep learning models like CNNs revolutionized applications such as:
Face recognition systems (Facebook's DeepFace, Apple's FaceID)
Medical imaging analysis (detecting tumors in CT scans)
Manufacturing quality control (spotting defects invisible to rule-based systems)
Autonomous vehicle perception systems (Tesla's Autopilot)
Natural Language Processing
Initial Challenge: "We need to understand and respond to customer inquiries."
Why ML Won: Rule-based NLP systems became unwieldy because:
Human language is incredibly complex and ambiguous
Context and intent are critical but difficult to encode in rules
Grammar exceptions are countless
Slang, idioms, and linguistic variations are ever-changing
Real-World Success: ML approaches transformed:
Machine translation (Google Translate improved dramatically with neural networks)
Customer service chatbots (companies reduced call center volume by 30-40%)
Sentiment analysis for brand monitoring (detecting subtle customer dissatisfaction)
Speech recognition (Amazon's Alexa, Google Assistant)
Fraud Detection
Initial Challenge: "We need to identify fraudulent transactions in real-time."
Why ML Won: Rule-based systems failed because:
Fraud patterns constantly evolve
False positives damaged customer experience
Rules became too complex to maintain
New attack vectors couldn't be anticipated
Real-World Success: ML models delivered:
PayPal's fraud detection system (reduced fraud rates by 10% while improving customer experience)
Credit card companies' real-time fraud alerts (detecting subtle pattern changes)
Insurance claim fraud detection (identifying connections between seemingly unrelated claims)
Recommendation Systems
Initial Challenge: "We need to suggest products customers will actually want."
Why ML Won: Traditional approaches using demographic rules or simple collaborative filtering couldn't:
Handle the "cold start" problem effectively
Capture complex preference patterns
Adapt quickly to changing consumer tastes
Balance exploration and exploitation
Real-World Success:
Netflix credits its recommendation engine with saving over $1 billion annually through improved retention
Amazon generates 35% of its revenue from its recommendation system
Spotify's Discover Weekly created a significant competitive advantage
Predictive Maintenance
Initial Challenge: "We need to predict when factory equipment will fail."
Why ML Won: Traditional threshold-based monitoring failed because:
Equipment failures have complex, subtle early indicators
Interrelated factors create patterns humans can't detect
Operating conditions vary significantly
Failure modes evolve over equipment lifetime
Real-World Success:
GE reduced unplanned downtime by 20% in power plants using ML-based predictive maintenance
Mining companies reduced equipment failures by 25% using sensor data and ML
Airlines reduced delays and cancellations
Medical Diagnosis
Initial Challenge: "We need better tools to assist doctors in diagnosis."
Why ML Won: Rule-based expert systems couldn't match ML because:
Disease presentations vary dramatically between patients
Multiple conditions create complex interactions
Subtle patterns in test results may indicate problems
Medical knowledge continues to expand beyond what can be manually encoded
Real-World Success:
Google's DeepMind created systems detecting eye disease from scans with accuracy matching specialists
ML systems for mammogram analysis found cancers missed by radiologists
IBM Watson for Oncology helps identify treatment options for complex cancer cases
In each case, the ability of ML systems to find patterns in complex data, adapt to changing conditions, and improve with experience provided capabilities that traditional algorithmic approaches simply couldn't match.