Muhammad Ahsan

11/09/2026

Continuing my learning journey in Generative AI and Natural Language Processing through the IBM AI Engineering Professional Certificate Program.

Recently, I’ve explored topics including word embeddings, N-grams, RNNs, encoder-decoder models, attention and self-attention, positional encoding, multi-head attention, Transformers, and text evaluation metrics.

Excited to keep learning, practicing, and turning these concepts into real AI projects.

05/09/2026

04/09/2026

Another step forward in my journey of learning Natural Language Processing and Generative AI.

As part of the IBM AI Engineering Professional Certificate program, I’ve recently been exploring how machines move from simply representing words as numbers to understanding context, processing sequences, generating text, and evaluating the quality of those generations.

My recent learning covered Word2Vec and CBOW, Recurrent Neural Networks (RNNs), GRUs and LSTMs, Sequence-to-Sequence models, Encoder-Decoder architectures, and evaluation metrics such as Perplexity, Precision, Recall, F1 Score, BLEU, and ROUGE.

What I find most interesting is seeing how these concepts connect rather than learning them in isolation:

Word Embeddings → Sequential Context → Memory → Encoder-Decoder → Text Generation → Evaluation

Understanding these foundations is helping me build a clearer picture of how NLP systems evolved and how the building blocks behind modern language models work.

Still learning, experimenting, and building my AI engineering skills one concept at a time.

25/08/2026

Continuing my learning journey in NLP and Generative AI.

In this stage, I explored how raw text is converted into numerical representations and then used for document classification and language modeling. I learned about tokenization, vocabularies, one-hot encoding, embeddings, embedding bags, neural networks, cross-entropy loss, optimization, and finally N-gram models for next-word prediction.

From converting words into features to classifying documents and predicting the next word, these concepts are helping me build a stronger foundation in NLP and understand how modern language models work.

08/08/2026

Data quality and diversity for effective LLM training
The quality and diversity of data are foundational to developing robust and inclusive large language models (LLMs). As models advance in sophistication, well-curated data is more critical than ever. Below are key aspects of preparing high-quality data and why they are crucial for effective LLM training.

Data quality
Data quality refers to the accuracy, consistency, and completeness of the dataset used for training. Poor-quality data introduces noise that can lower a model's accuracy and reliability. Here are practices to ensure high data quality:

Noise reduction: Remove irrelevant or repetitive data to help the model focus on significant patterns and linguistic structures. For example, clean datasets by removing typos and irrelevant information, such as forum tags, to enhance quality.

Consistency checks: Regularly verify consistency to prevent conflicting or outdated information from confusing the model. Consistency is essential for entities, like public figure names or technical terms, ensuring uniform usage throughout the dataset.

Labeling quality: For labeled datasets, accurate labeling is crucial to avoid misleading the model. Clear guidelines for human annotators can reduce subjective errors and improve labeling quality.

Diverse representation
A diverse dataset enhances a model’s inclusivity, ensuring it responds accurately to varied cultural, demographic, and regional inputs. Without diverse representation, models may reflect narrow views, leading to unintentional biases. Here’s how to achieve meaningful diversity:

Inclusion of varied demographics: Incorporate text from various demographic groups to avoid over-representing a single perspective. Include sources in multiple languages or dialects and represent diverse cultural norms to improve global applicability.

Balanced data sources: Draw a balanced dataset from sources such as news, social media, literature, and technical documents. This broadens the model’s knowledge base and reduces dependence on any single source.

Regional and linguistic variety: Including datasets from diverse regions and languages expands the model’s linguistic and cultural context, enhancing accuracy in multilingual contexts and better supporting translation tasks.

Regular updates
Language is constantly evolving, with new terminologies and shifts in usage patterns emerging frequently. Regular updates to datasets are essential to keep a model relevant and accurate. Here’s why updates matter:

New vocabulary and trends: Capture evolving language trends, such as “selfie” or “cryptocurrency,” through regular data updates to reflect current terminology.

Cultural and social norms: As societal perspectives shift, language models should adapt accordingly. An LLM trained on outdated data may inadvertently reinforce outdated norms or stereotypes.

Model retraining: Periodically updating the model with fresh data helps maintain alignment with contemporary knowledge and societal standards.

Ethical considerations in data collection
Ethics in data collection are essential to protect user privacy and ensure fair representation. Ethical data practices are fundamental to building trust and reducing biases:

Data privacy: Use anonymized data to protect personal information, especially in datasets containing sensitive or identifiable information.

Fair representation: Ensure the inclusion of marginalized voices to avoid bias that can reinforce societal inequalities.

Transparency in data sources: Disclose data sources used for model training. This transparency fosters user trust and allows for understanding the foundation of the model’s knowledge.

Conclusion
Focusing on data quality, diversity, and ethical practices contributes to developing LLMs that are accurate, inclusive, and socially responsible. With a well-prepared dataset, your LLM will perform more effectively while helping bridge gaps in AI fairness and representation.

04/08/2026

Today I learned about Vision Transformers (ViTs) and it completely changed the way I think about how AI understands images. 🤖🖼️

Unlike traditional CNNs, Vision Transformers process an image by breaking it into small patches, almost like reading a sentence word by word. Here's what I found most interesting:

✅ Images are divided into fixed-size patches.

✅ Each patch is converted into an embedding that the model can understand.

✅ Positional Encoding helps the model know where every patch belongs in the original image.

✅ Self-Attention allows every image patch to interact with every other patch, helping the model understand both local details and the overall context.

✅ A special **CLS Token** collects information from all patches and is used to make the final prediction.

✅ Instead of training from scratch, we can use **Transfer Learning** with pre-trained Vision Transformers and fine-tune them on smaller datasets.

✅ Combining **CNNs + Vision Transformers** creates a powerful hybrid model that captures both local features and global relationships.

One thing I really enjoyed learning is that Vision Transformers don't necessarily replace CNNs—they complement them. Each architecture has its own strengths, and choosing the right one depends on the problem you're solving.

I'm documenting my AI & Machine Learning journey publicly because every concept I learn brings me one step closer to becoming a better AI/ML Engineer. 🚀

More deep learning and computer vision topics coming soon!