- Study
Natural Language Processing- & continue learning Image Generation with
Diffusers
- & continue learning Image Generation with
- Apply knowledge to projects
Details: last year's Computer Vision Things I am improving
- Focus on high level libraries like
transformersanddiffusersoverpytorch- Advantage: Few lines to have end to end pipeline
- Advantage: Getting to solution quickly instead of 2 weeks of implementing from scratch
- Do more Kaggle competitions compared to courses. (10 courses & 3 competitions last year, Aiming for 10+ competitions this year)
- Continue coding in pytorch, build code cookbook for experimentation
- Visualize model internals to understand it better. (Didn't do this part last year)
- Understand maths aspect of neural networks. (Didn't do this part last year.)
- Landscape of standard LLM architectures, once memorized, a lot of problems become a lot simpler
- Projects are the best way to learn. Deep Learning is a field of experimentation, not theoratical field. That is why projects are the best way to learn.
- State of the art models require bigger GPUs. Free T4 GPU has 16GB of VRAM, can only store 8 Billion parameter model.
- LLama models are the best models open source AI models, which are as commercial AI models like GPT-4 and Google AI
- Research Papers on diffusion models & transformer models optimizations
- Reinforcement Learning
- Custom Dataset building & fine tuning models
- Model Optimization
NLP Landscape's two entry paths. One slow, long & easier path & other short but steep path
- Long & Slow path: DL Basics -> Simple NN -> CNN -> RNN -> LSTM -> Word2Vec -> Attention -> LLMs
- Short & Steep path: DL Basics -> Attention is all you need -> LLMs
- Everything in NLP & CV is building on top of this single paper. Highest citations, highest used architecture, is most varied kinds of problems.
- Understand this thoroughly, because everything builds on this
| Type | Details | Progress |
|---|---|---|
| 1: Course | Huggingface NLP | |
| 2: Kaggle Competition | Disaster Tweet Classification | |
| 3: Research Paper | Attention is all you need | |
| 4: Research Paper | One Model to learn them all | |
| 5: Youtube Video | 3Blue1Brown: Attention in transformers, visually explained | |
| 6: Youtube Video | 3Blue1Brown: But what is a GPT? Visual intro to transformers | |
| 7: Youtube Video | Campus X: Epic History of Large Language Models | |
| 8: Youtube Video | Campus X: Self Attention | |
| 9: Youtube Video | Campus X: Attention | |
| 11: Udemy Course | LLM Application Dev with Langchain |
- Good Certificate for knowledge validation
- But exam has closed now. No pytorch certification. Currently there is no good gap.
- Studied, learned it, but
Failedbecause things like getting accuracy above a certain number.
| Type | Details | Progress |
|---|---|---|
| 1: Course | Huggingface timm | |
| 2: Course | Huggingface diffusers | |
| 3: Course | Huggingface Community Vision Course | |
| 4: Course | Zero to Mastery Tensorflow |
| Competition | Progress |
|---|---|
| Cats vs Dogs - End to End Pipeline | |
| 10 Small Objects Recognition(CIFAR10) | |
| Imagenet Classification |
- Interesting industries as application
- I love books. So projects / industries around books would be - Book to Illustrated Images or Book's adaption to Film / Tv Series.
- I also love video games. Any application in video game building pipeline
- Found out in further research, Video Game Industry is bigger than Books + Movies + Tv Series + Music.
Harry Potter- immersive book or video game or a filmUnreal Engine 5forVideo GamesScreenplaysforFilm making of Harry Potter- I have loved
Dunetoo. So studiedScreenplayof Dune as well Mahabharat Video Game- Researching
Mahabharat - Reading
Harry Potter series
- Data Engineering for
MTech StudentsforLTI Mindtree - Application Development for
MTech StudentsforLTI Mindtree - How Developers can learn Artificial Intelligence - a
DevRel Talk