Transformer LMs

David Strohmaier, ALTA Institute, University of Cambridge

[email protected]

"We offer no explanation as to why these architectures seem to work; we attribute their success, as all else, to divine benevolence." ~ Shazeer (2020)

Goals of this Presentation

• Provide an introduction to transformer language models (LMs)
• Understand basic architecture and its variations
• Understand sources of failure
• Highlight details of interest for philosophers and cognitive scientists

Generative AI

• Models that generate text, images, videos etc. that can be directly consumed.
• In the case of text: typically sequence-to-sequence (seq2seq)
• Model architectures that can be used for generation:
  • Recurrent Neural Networks
  • Diffusion models (images)
  • Transformers ←today
  • ...

Language Models

• A language model (LM) provides a probability distribution over tokens given a context
• $ p(word|context) \approx \Phi(context) $
• Old approach: Counting occurrences
• Neural language models: Around since the early 2000s

Transformer Models

• Introduced by Vaswani et al. (2017)
• A type of neural architecture based on the attention mechanism
• Efficient for processing and producing sequences

(image taken from Vaswani et al. (2017) and modified)

Simplifying! (cf. Geva et al. 2022 for feedforward layers)

Attention Mechanism

• A way to contextualise representations
• Representations are combined based on their similarity (dot-product)

---

• "Attention Is all You Need" (Vaswani et al. 2017)
• "Attention Is not all You Need" (Dong et al. 2021)
• "Attention is Turing Complete" (Pérez et al. 2021) ← under certain assumptions
• https://www.isattentionallyouneed.com/

Image source: Fig 3 in Bahdanau et al., 2015 https://arxiv.org/pdf/1409.0473

• Attention is used for interpretation of models.
• Some attention heads specialise on specific syntactic phenomena (e.g. Clark et al 2019, also Rogers et al. 2020).
• But it's not just importance.

Semantic Space

• Embeddings/activations: The vector of numbers a layer puts out
• Embeddings/activations live in a cartesian space
• Geometric interpretation of embeddings
• Space exhibits interesting regularities
  • Even when reduced to 2D
  • Why would that be?

Generation

• How to get from state to vocabulary
• Different decoding strategies
• Sampling from the probabilitiy distribution:
  • Temperature: Adjusts probability distribution

Next-word vs. masked-language-model

• Predict one (sub-)token at a time, left to right.
• Autoregressive language modelling (i.e. taking its own output as input)

---

• Predict (sub-)token anywhere in the sequence.
• Usually a single token or whole word.

• Surprising differences!

The world is everything that is the ___.

The world is everything that is the ___.

Training

• Large amounts of data
• In batches: predict target word (masked or next) + backpropagation
• Models might be undertrained

Backpropagation

1. Pass batch of data through
2. Calculate loss: Divergence between output and target
3. Use gradient to adjust weights.

Some bold claims

• The neural network is not a blank slate
  • Architectural bias
  • Random initialisation, but which distribution?
  • Lottery ticket hypothesis (Frankle & Carbin 2018)
• The training is not just gradual
  • Especially when it comes to compositional skills

What I have not covered

• Prompting
• Hyperparameter search
• Finetuning
• RLHF
• Multi-modality
• Lots of architectural details: activation, position....
• ...

References

• Clark, K., Khandelwal, U., Levy, O., & Manning, C. D. (2019). What Does BERT Look at? An Analysis of BERT’s Attention. In T. Linzen, G. Chrupała, Y. Belinkov, & D. Hupkes (Eds.), Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP (pp. 276–286). Association for Computational Linguistics. https://doi.org/10.18653/v1/W19-4828
• Dong, Y., Cordonnier, J.-B., & Loukas, A. (2021). Attention is not all you need: Pure attention loses rank doubly exponentially with depth. Proceedings of the 38th International Conference on Machine Learning, 2793–2803. https://proceedings.mlr.press/v139/dong21a.html
• Frankle, J., & Carbin, M. (2018, September 27). The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks. International Conference on Learning Representations. https://openreview.net/forum?id=rJl-b3RcF7
• Geva, M., Caciularu, A., Wang, K., & Goldberg, Y. (2022). Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 30–45. https://aclanthology.org/2022.emnlp-main.3

• Pérez, J., Barceló, P., & Marinkovic, J. (2021). Attention is Turing-Complete. Journal of Machine Learning Research, 22(75), 1–35.
• Rogers, A., Kovaleva, O., & Rumshisky, A. (2020). A Primer in BERTology: What We Know About How BERT Works. Transactions of the Association for Computational Linguistics, 8, 842–866. https://doi.org/10.1162/tacl_a_00349
• Shazeer, N. (2020). GLU Variants Improve Transformer (arXiv:2002.05202). arXiv. https://doi.org/10.48550/arXiv.2002.05202
• Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is All you Need. Advances in Neural Information Processing Systems, 30. https://papers.nips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html