https://www.turing.com/kb/brief-introduction-to-transformers-and-their-power#the-transformer-encoder
Decoders pay attention only to the words before them, as opposed to encoders, which pay attention to every word regardless of order. As a result, the prediction for the word at the position, i, only depends on the words preceding it in the sequence.
Based on the search results, scGPT determines "known" and "unknown" genes in the following ways:
1. During Training:
- A random proportion of genes are selected as "unknown" genes.
- The expression values of these "unknown" genes are omitted from the input.
- The remaining genes are considered "known" genes.
2. Attention Masking:
- The model uses a specialized attention mask that only allows attention computation between the "known" genes and the query gene itself.
- "Unknown" genes cannot attend to other "unknown" genes.
3. Iterative Prediction Process:
- The model predicts expression values for "unknown" genes based on information from "known" genes.
- During inference, the process is conducted in K iterative steps.
4. Confidence-based Selection:
- In each iteration, scGPT selects the top 1/K genes from the "unknown" set with the highest prediction confidence.
- These newly predicted genes become "known" genes for the next iteration.
5. Auto-regressive Generation:
- This process creates a form of auto-regressive generation for non-sequential data.
- Gene expressions predicted with highest confidence are generated first and then used to help predict subsequent genes.
6. Cell-prompt vs. Gene-prompt Generation:
- For cell-prompt generation, it starts with a cell embedding representing the cell type condition.
- For gene-prompt generation, it begins with a set of known genes with observed expression values.
This approach allows scGPT to handle the non-sequential nature of single-cell data while still leveraging the power of transformer models for prediction tasks.
Citations:
[1] https://twitter.com/simocristea/status/1676323087959179264
[2] https://www.linkedin.com/pulse/new-generative-ai-tool-predicts-gene-expression-single-colangelo-x7ebf
[3] https://www.the-scientist.com/a-new-ai-tool-predicts-gene-expression-in-a-single-cell-71295
[4] https://www.biorxiv.org/content/10.1101/2023.04.30.538439v2.full
[5] https://www.biorxiv.org/content/10.1101/2023.04.30.538439v1.full
https://chatgpt.com/share/98a22c79-4119-4d07-ac17-9712301d7ae0
BERT and scGPT use different architectures (encoder-only and decoder-only, respectively) due to their specific pre-training objectives and intended applications. Here's an explanation of why each model uses its chosen architecture and how it relates to their pre-training:
BERT (Bidirectional Encoder Representations from Transformers):
1. Encoder-only architecture:
- BERT uses only the encoder part of the transformer architecture.
- This allows for bidirectional processing of input text, meaning it can look at both left and right context simultaneously.
2. Pre-training objective:
- Masked Language Modeling (MLM): BERT is trained to predict masked tokens in a sentence, given both left and right context.
- Next Sentence Prediction (NSP): It also predicts whether two sentences follow each other in the original text.
3. Relation to pre-training:
- The encoder-only architecture is well-suited for these tasks because they require understanding the full context of a sentence or pair of sentences.
- Bidirectional attention allows BERT to capture complex relationships between words in a sentence, which is crucial for predicting masked tokens accurately.
4. Intended use:
- BERT is designed for tasks that require deep understanding of text, such as classification, named entity recognition, and question answering.
- It's not meant for text generation tasks.
scGPT (Single-Cell Generative Pre-trained Transformer):
1. Decoder-only architecture:
- scGPT uses only the decoder part of the transformer architecture.
- This allows for autoregressive prediction, where each token is predicted based on the previous tokens.
2. Pre-training objective:
- Generative pre-training: scGPT is trained to predict gene expression values in a forced sequential manner.
- It uses a specialized attention masking technique to handle non-sequential single-cell data.
3. Relation to pre-training:
- The decoder-only architecture is well-suited for this task because it allows the model to generate expression values one at a time, considering the previously generated values.
- This approach enables the model to capture complex dependencies between genes and learn cell-type specific patterns.
4. Intended use:
- scGPT is designed for tasks related to single-cell multi-omics data, such as cell type annotation, multi-omics integration, and perturbation prediction.
- It can generate unknown gene expression values given known gene values or generate whole genome expression given a known cell type condition.
In summary, the choice of encoder-only for BERT and decoder-only for scGPT is directly related to their pre-training objectives and intended applications. BERT's bidirectional processing is ideal for understanding context in natural language, while scGPT's autoregressive approach is suitable for generating gene expression values in single-cell data. These architectural choices allow each model to excel in its specific domain and tasks.
Citations:
[1] https://vaclavkosar.com/ml/Encoder-only-Decoder-only-vs-Encoder-Decoder-Transfomer
[2] https://datascience.stackexchange.com/questions/65241/why-is-the-decoder-not-a-part-of-bert-architecture
[3] https://ai.stackexchange.com/questions/41505/which-situation-will-helpful-using-encoder-or-decoder-or-both-in-transformer-mod
[4] https://www.linkedin.com/pulse/transformer-architectures-dummies-part-2-decoder-only-bhaskar-t-hj9xc
[5] https://www.reddit.com/r/MLQuestions/comments/l1eiuo/when_would_we_use_a_transformer_encoder_only/
No, scGPT's masked training approach is not similar to BERT's. While both models use masking techniques, they differ significantly in their implementation and purpose:
1. BERT's Masked Language Model (MLM):
- BERT randomly masks tokens in the input sequence and predicts these masked tokens.
- It uses bidirectional context, looking at both left and right contexts to predict the masked tokens.
- This approach is designed for understanding the full context of text in natural language processing tasks.
2. scGPT's Specialized Attention Masking:
- scGPT uses a unique attention masking technique designed specifically for non-sequential single-cell data[1].
- It supports both gene-prompt and cell-prompt generations in a unified way.
- The masking defines the order of prediction based on attention scores, not on sequential order as in text.
- It allows attention computation only between embeddings of "known genes" and the query gene itself[1].
Key differences:
1. Purpose: BERT's masking is for understanding context in text, while scGPT's is for generating gene expression values in non-sequential data.
2. Directionality: BERT is bidirectional, while scGPT uses a form of forced sequential prediction in non-sequential data.
3. Prediction target: BERT predicts masked tokens, while scGPT predicts gene expression values.
4. Context use: BERT uses full context, while scGPT uses known genes to predict unknown genes iteratively.
5. Architecture: BERT is an encoder-only model, while scGPT is a decoder-only model, more similar to GPT architectures[1].
In summary, while both use masking, scGPT's approach is specifically tailored for single-cell genomic data and is fundamentally different from BERT's text-based masking approach.
Citations:
[1] https://www.biorxiv.org/content/10.1101/2023.04.30.538439v1.full
[2] https://www.reddit.com/r/deeplearning/comments/17gmtxr/bert_explained_training_masked_language_model/
[3] https://softteco.com/blog/bert-vs-chatgpt
[4] https://blog.invgate.com/gpt-3-vs-bert
[5] https://www.biorxiv.org/content/10.1101/2023.04.30.538439v2.full
Quantum Vision Transformers
quantum implementation of attention mechanism
Recent work on representing “Feynman diagrams as computational graphs” has sparked an intriguing idea: Let’s map AI computation to Feynman diagrams to visualize and optimize AI architectures.
๐ก By leveraging Meta’s LLM Compiler, we can create a powerful interpreter between quantum field theory techniques and AI model design.
๐๐๐ซ๐'๐ฌ ๐ก๐จ๐ฐ ๐ข๐ญ ๐ฐ๐จ๐ซ๐ค๐ฌ:
1. Represent AI models as Feynman-like diagrams, with nodes as computation units (e.g., transformer blocks) and edges showing data flow.
2. Use the LLM Compiler to analyze these diagrams, suggesting optimizations based on both structure and underlying computations.
3. Instead of integrating traditional LLVMs we swap it out for Meta’s LLM compiler for a multi-level optimization approach:
- ๐๐ข๐ ๐ก-๐ฅ๐๐ฏ๐๐ฅ: LLM-driven architectural changes
- ๐๐ข๐-๐ฅ๐๐ฏ๐๐ฅ: Standard compiler optimizations
- ๐๐จ๐ฐ-๐ฅ๐๐ฏ๐๐ฅ: Hardware-specific tweaks
๐๐ก๐ข๐ฌ ๐๐ฉ๐ฉ๐ซ๐จ๐๐๐ก ๐จ๐๐๐๐ซ๐ฌ ๐ฌ๐๐ฏ๐๐ซ๐๐ฅ ๐ค๐๐ฒ ๐๐๐ฏ๐๐ง๐ญ๐๐ ๐๐ฌ:
1. ๐๐ง๐ก๐๐ง๐๐๐ ๐ข๐ง๐ญ๐๐ซ๐ฉ๐ซ๐๐ญ๐๐๐ข๐ฅ๐ข๐ญ๐ฒ: Feynman diagrams provide a visual language for complex AI systems, crucial for debugging and regulatory compliance.
2. ๐๐ซ๐จ๐ฌ๐ฌ-๐๐จ๐ฆ๐๐ข๐ง ๐ข๐ง๐ฌ๐ข๐ ๐ก๐ญ๐ฌ: The LLM's capabilities to compile and optimize models inspired by QFT principles.
3. ๐๐๐ซ๐๐ฐ๐๐ซ๐-๐๐ฐ๐๐ซ๐ ๐๐๐ฌ๐ข๐ ๐ง: Optimizations can be tailored to specific GPU or TPU architectures, improving efficiency.
4. ๐๐ญ๐๐ซ๐๐ญ๐ข๐ฏ๐ ๐ซ๐๐๐ข๐ง๐๐ฆ๐๐ง๐ญ: Continuous learning from optimization patterns leads to increasingly sophisticated improvements over time.
Of course, there are challenges. Representing very deep networks or handling the complexity of recurrent connections could be tricky. But I believe the potential benefits outweigh these hurdles.
๐ก Now, here's where we can take it to the next level: Combine this Feynman diagram approach with LLM-based intelligent optimization, like Meta's LLM Compiler. We could create a powerful system where both human designers and AI systems work with the same visual language.
๐ช Imagine an LLM analyzing these AI Feynman diagrams, suggesting optimizations, and even generating or modifying code directly. This could bridge the gap between high-level model architecture and low-level implementation details, potentially leading to more efficient and interpretable AI systems.
This approach could be particularly powerful in domains like hashtag#explainableAI and hashtag#AIsafety, where understanding the decision-making process is crucial.
I'm incredibly excited about this direction. It could be a major leap towards more intuitive and powerful ways of developing AI, bringing together experts from physics, AI, and visual design.