mg98
commented
1 year ago Ideas...
- "Perfect search" as an instance of unsupervised online learning-to-rank (OL2R) (using deep neural networks). The ML model should use user signals (i.e., clicks/downloads) as a measure to learn query-document relevancy and accordingly adapt its weights.
- I think the employment of NN makes sense as we have very limited metadata/understanding about each torrent, much unlike is the case with web pages
- State-of-the-art for OL2R: Duel Bandit Gradient Descent (DBGD) (listwise approach), (placeholder)
- Potential research lines:
- Analysis of pointwise, pairwise, & listwise OL2R approaches for their application on distributed or decentralized systems; How do the tradeoffs added by listwise approaches scale with the constraints of robustness and efficiency in a dist./dec. network? Or more specific: how would DBGD work out in this setting?
- Alternatively, ...
- Decentralized AI-powered information retrieval using something as simple as, e.g., BM25 ranking (that is, without NN)
- Decentralized embeddings for semantic search
synctext
+ Something we called crowdsourcing 20 years ago: RLHF, https://huyenchip.com/2023/05/02/rlhf.html
+ https://github.com/Tribler/science/issues/42
+ https://github.com/openlm-research/open_llama
+ https://github.com/togethercomputer/RedPajama-Data/blob/main/data_prep/arxiv/arxiv_cleaner.py
+ 2011 [P2Prec: A P2P Recommendation System for Large-Scale Data Sharing](https://dl.acm.org/doi/abs/10.5555/2028190.2028194)
+ 2013 [A peer-to-peer recommender system for self-emerging user communities based on gossip overlays](https://doi.org/10.1016/j.jcss.2012.05.011)
+ simpel sprint idea: IPv8 community with trivial primitive. 3 items in a request, 2 recommendations + signed semantic similarity in return. ToDo: tag-based.
Suggested sprint: you successfully encoded ```sin()```. Next, take 4 images, try to [embed them in a machine learning model using useful overfitting](https://towardsdatascience.com/how-to-create-a-concise-image-representation-using-machine-learning-20156c1e0c19). Meerkat [docs example](https://github.com/HazyResearch/meerkat/blob/main/docs/source/examples/image-search.rst) of ```embed()```. Scientific primitive, unbounded storage of [unstructured data](https://hazyresearch.stanford.edu/blog/2023-03-01-meerkat) for decentralised learning. Key question, how much storage space for 1 million static thumbnails?
```
import meerkat as mk
mk.search(
df,
query=mk.embed("A photo of a person", engine="clip"),
by=mk.embed(df["image"], engine="clip")
)
``` 
devos50
< Placeholder > timeline: April 2023 - April 2027. ToDo: 6 weeks hands-on Python onboarding project. Learn a lot and plan to throw it away. Next step is then to start working on your first scientific article. You need 4 thesis chapters and you then completed your phd.
One idea: towards trustworthy and perfect metadata for search (e.g. perfect memory retrieval for the global brain #7064 ). Another idea: Gradient decent model takes any keyword search query as input. Output is a limited set of vectors. Only valid content is recommended. Learning goal is to provide semantic matching between input query and output vector. General background and Spotity linkage possible dataset sharing
Max. usage of expertise: product/market fit thinking
Background reading:
pointwise approach, broadly speaking, each historic impression with a click is a positive training example, and each impression without a click is a negative training example.https://towardsdatascience.com/learning-to-rank-a-primer-40d2ff9960afWe will implement a character-level sequence-to-sequence model, processing the input character-by-character and generating the output character-by-character. Another option would be a word-level model, which tends to be more common for machine translation., https://blog.keras.io/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.htmlVenues:
Possible scientific storyline: SearchZero a decentralised, self-supervised search engine with continuous learning