Transformer networks

The transformer-powered subtitle for this article is “Our most terrifyingly effective weapon against the forces of evil is our ability to laugh at them.”

December 20, 2017 — June 15, 2024

machine learning
meta learning
neural nets
time series
Figure 1

Well, it’s really terribly simple, […] it works any way you want it to. You see, the computer that runs it is a rather advanced one. In fact it is more powerful than the sum total of all the computers on this planet including—and this is the tricky part—including itself.

— Douglas Adams, Dirk Gently’s Holistic Detective Agency

Transformers are big self-attention networks with some extra tricks — self attention, and a query, key, value mechanism, and usually a positional encoding as well.

I am no expert. Here are some good blog posts explaining everything, for my reference, but I will not write yet another one. This is a fast-moving area and I am not keeping track of it, so if you are on this page looking for guidance you are already in trouble.

These networks are massive (heh) in natural language processing right now.

A key point about such networks seems to be that they can be made extremely large but still remain trainable. This leads to interesting scaling laws.

1 Introductions

So many.

TODO: rank in terms of lay-person-friendliness.

2 Power of

Transformers are pretty good at weird stuff, e.g. automata — see Unveiling Transformers with LEGO (Zhang et al. 2022).

How about Bayesian inference? (Müller et al. 2022)

Can they be an engine of intelligence? What do they do in society? etc. Controversial — see the Stochastic Parrots paper (Bender et al. 2021), and the entire internet commentariat from November 2022 onwards.

Figure 2: Image by Anthrupad

3 As set functions

Transformers are neural set functions (!).


State-space, i.e. recurrent transformers, without (classic) attention. Suggestive connection to S4 models.

RWKV is inspired by Apple’s Attention Free Transformer. (Zhai et al. 2021). …

How to combine the best of transformers and RNNs? The main drawback of transformer-based models is that it can become challenging to run a model with a context window that is larger than a certain value, as the attention scores are computed simultaneously for the entire sequence.

RNNs natively support very long context lengths—only limited by the context length seen in training, but this can be extended to millions of tokens with careful coding. Currently, there are RWKV models trained on a context length of 8192 (ctx8192) and they are as fast as ctx1024 models and require the same amount of RAM.

The major drawbacks of traditional RNN models and how RWKV is different:

  1. Traditional RNN models are unable to utilize very long contexts (LSTM can only manage ~100 tokens when used as a LM). However, RWKV can utilize thousands of tokens and beyond…
  2. Traditional RNN models cannot be parallelized when training. RWKV is similar to a “linearized GPT” and it trains faster than GPT.

By combining both advantages into a single architecture, the hope is that RWKV can grow to become more than the sum of its parts.

Figure 3

5 Tokens as recurrent state

See also RecurrentGPT (Zhou et al. 2023)

GitHub - aiwaves-cn/RecurrentGPT

RecurrentGPT replaces the vectorized elements (i.e., cell state, hidden state, input, and output) in a Long-short Term Memory RNN (LSTM) with natural language (i.e., paragraphs of texts), and simulates the recurrence mechanism with prompt engineering.

At each timestep t, RecurrentGPT receives a paragraph of text and a brief plan of the next paragraph, which are both generated in step t − 1. It then attends to the long-term memory, which contains the summaries of all previously generated paragraphs and can be stored on hard drives, and relevant paragraphs can be retrieved with semantic search.

RecurrentGPT also maintains a short-term memory that summarizes key information within recent timesteps in natural language and is updated at each time step. RecurrentGPT combines all aforementioned inputs in a prompt and asks the backbone LLM to generate a new paragraph, a short plan for the next paragraph, and updates the long-short term memory by rewriting the short-term memory and appending the summary of the output paragraph to the long-term memory.

6 Practicalities

For you and me, see AI democratizateion.

7 Embedding vector databases

See embedding vector databases.

  • LMQL: Programming Large Language Models: “LMQL is a programming language for language model interaction.” (Beurer-Kellner, Fischer, and Vechev 2022)

    LMQL generalizes natural language prompting, making it more expressive while remaining accessible. For this, LMQL builds on top of Python, allowing users to express natural language prompts that also contain code. The resulting queries can be directly executed on language models like OpenAI’s GPT models. Fixed answer templates and intermediate instructions allow the user to steer the LLM’s reasoning process.

  • How does in-context learning work? A framework for understanding the differences from traditional supervised learning

    TL;DR — In-context learning is a mysterious emergent behavior in large language models (LMs) where the LM performs a task just by conditioning on input-output examples, without optimizing any parameters. In this post, we provide a Bayesian inference framework for understanding in-context learning as “locating” latent concepts the LM has acquired from pretraining data. This suggests that all components of the prompt (inputs, outputs, formatting, and the input-output mapping) can provide information for inferring the latent concept. We connect this framework to empirical evidence where in-context learning still works when provided training examples with random outputs. While output randomization cripples traditional supervised learning algorithms, it only removes one source of information for Bayesian inference (the input-output mapping).

  • What Are the Different Approaches for Detecting Content Generated by LLMs Such As ChatGPT? And How Do They Work and Differ?

  • Large Language Models as General Pattern Machines

    We observe that pre-trained large language models (LLMs) are capable of autoregressively completing complex token sequences—from arbitrary ones procedurally generated by probabilistic context-free grammars (PCFG), to more rich spatial patterns found in the Abstract Reasoning Corpus (ARC), a general AI benchmark, prompted in the style of ASCII art. Surprisingly, pattern completion proficiency can be partially retained even when the sequences are expressed using tokens randomly sampled from the vocabulary. These results suggest that without any additional training, LLMs can serve as general sequence modelers, driven by in-context learning. In this work, we investigate how these zero-shot capabilities may be applied to problems in robotics—from extrapolating sequences of numbers that represent states over time to complete simple motions, to least-to-most prompting of reward-conditioned trajectories that can discover and represent closed-loop policies (e.g., a stabilizing controller for CartPole). While difficult to deploy today for real systems due to latency, context size limitations, and compute costs, the approach of using LLMs to drive low-level control may provide an exciting glimpse into how the patterns among words could be transferred to actions.

  • karpathy/nanoGPT: The simplest, fastest repository for training/finetuning medium-sized GPTs.

8 References

Bahdanau, Cho, and Bengio. 2015. Neural Machine Translation by Jointly Learning to Align and Translate.” In.
Bender, Gebru, McMillan-Major, et al. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.
Beurer-Kellner, Fischer, and Vechev. 2022. Prompting Is Programming: A Query Language For Large Language Models.”
Brown, Mann, Ryder, et al. 2020. Language Models Are Few-Shot Learners.” arXiv:2005.14165 [Cs].
Bubeck, Chandrasekaran, Eldan, et al. 2023. Sparks of Artificial General Intelligence: Early Experiments with GPT-4.”
Cao. 2021. Choose a Transformer: Fourier or Galerkin.” In Advances in Neural Information Processing Systems.
Celikyilmaz, Deng, Li, et al. 2017. Scaffolding Networks for Teaching and Learning to Comprehend.” arXiv:1702.08653 [Cs].
Choy, Gwak, Savarese, et al. 2016. Universal Correspondence Network.” In Advances in Neural Information Processing Systems 29.
Din, Karidi, Choshen, et al. 2023. Jump to Conclusions: Short-Cutting Transformers With Linear Transformations.”
Ergen, Neyshabur, and Mehta. 2022. Convexifying Transformers: Improving Optimization and Understanding of Transformer Networks.”
Freeman. 2019. How to Communicate Evidence to Patients.” Drug and Therapeutics Bulletin.
Huang, Vaswani, Uszkoreit, et al. 2018. Music Transformer: Generating Music with Long-Term Structure.”
Katharopoulos, Vyas, Pappas, et al. 2020. Transformers Are RNNs: Fast Autoregressive Transformers with Linear Attention.” arXiv:2006.16236 [Cs, Stat].
Korbak, Perez, and Buckley. 2022. RL with KL Penalties Is Better Viewed as Bayesian Inference.”
Li, Wallace, Shen, et al. 2020. Train Large, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers.” arXiv:2002.11794 [Cs].
Merrill, and Sabharwal. 2022. Transformers Implement First-Order Logic with Majority Quantifiers.”
Müller, Hollmann, Arango, et al. 2022. Transformers Can Do Bayesian Inference.”
Nguyen, Brandstetter, Kapoor, et al. 2023. ClimaX: A Foundation Model for Weather and Climate.”
Ortega, Kunesch, Delétang, et al. 2021. Shaking the Foundations: Delusions in Sequence Models for Interaction and Control.” arXiv:2110.10819 [Cs].
Phuong, and Hutter. 2022. Formal Algorithms for Transformers.”
Piantadosi, and Hill. 2022. Meaning Without Reference in Large Language Models.”
Radford, Wu, Child, et al. 2019. “Language Models Are Unsupervised Multitask Learners.”
Rafailov, Sharma, Mitchell, et al. 2023. Direct Preference Optimization: Your Language Model Is Secretly a Reward Model.”
Ramsauer, Schäfl, Lehner, et al. 2020. Hopfield Networks Is All You Need.” arXiv:2008.02217 [Cs, Stat].
Serrano, Brumbaugh, and Smith. 2023. Language Models: A Guide for the Perplexed.”
Vardasbi, Pires, Schmidt, et al. 2023. State Spaces Aren’t Enough: Machine Translation Needs Attention.”
Vaswani, Shazeer, Parmar, et al. 2017. Attention Is All You Need.” arXiv:1706.03762 [Cs].
Wang, Gangavarapu, Yan, et al. 2024. MambaByte: Token-Free Selective State Space Model.”
Willig, Zečević, Dhami, et al. 2022. Can Foundation Models Talk Causality?
Wu, Tan, Wang, et al. 2024. Beyond Language Models: Byte Models Are Digital World Simulators.”
Yang, and Hu. 2020. Feature Learning in Infinite-Width Neural Networks.” arXiv:2011.14522 [Cond-Mat].
Zhai, Talbott, Srivastava, et al. 2021. An Attention Free Transformer.”
Zhang, Backurs, Bubeck, et al. 2022. Unveiling Transformers with LEGO: A Synthetic Reasoning Task.”
Zhou, Jiang, Cui, et al. 2023. RecurrentGPT: Interactive Generation of (Arbitrarily) Long Text.”
Zou, Wang, Kolter, et al. 2023. Universal and Transferable Adversarial Attacks on Aligned Language Models.”