Do large language models really need all those layers?
Finding that 70% of attention heads and 20% of feed-forward networks can be excised with minimal effect on in-context learning suggests that large language models are undertrained.
Large language models (LLMs) have been around for a while but have really captured the attention of the public this year, with the advent of ChatGPT. LLMs are typically pretrained on massive volumes of data; recent variants are additionally tuned to follow instructions and incorporate human feedback using reinforcement learning.
A fascinating ability that these LLMs demonstrate is in-context learning, where a model can learn to perform a task just by following a few (or sometimes even zero) good examples provided along with a new input. Following this paradigm of learning, larger LLMs also proved more capable of performing a wide variety of tasks than smaller ones, when the amount of pretraining data was fixed.
In a paper we’re presenting at this year’s meeting of the Association for Computational Linguistics (ACL), we investigate the importance of model scale for in-context learning, from the perspective of architectural interpretability. We specifically ask the question Are all LLM components really needed to perform in-context learning?
We conducted our investigation as a case study of the OPT-66B model, a 66-billion-parameter LLM that was open-sourced by Meta last year to serve as an open replica of GPT-3 (and was the largest publicly available decoder-only LLM at the time of our study). We found that a significant portion of the model could be discarded without affecting performance, indicating that OPT-66B and quite likely other prominent LLMs are undertrained.
We believe our findings are useful in helping build more powerful LLMs by identifying (or more generally providing methods to identify) architectural elements that may need to be trained better.
LLM building blocks
Modern LLMs use the Transformer architecture, which depends on an attention mechanism: the model learns to predict which prior tokens in the sequence it should attend to when predicting the current token.
Specifically, LLMs use multihead attention, meaning that they apply multiple attention mechanisms, or heads, in parallel. OPT-66B has 64 layers with 72 attention heads in each layer. The output of the multihead attention passes through a separate feed-forward network (FFN) at each layer.
Our first method for analyzing OPT-66B was to assign a score to each attention head and FFN indicating how important they were to a given task. On the basis of those scores, we then pruned the model.
We found that important attention heads are primarily clustered in the model’s intermediate layers, and important FFNs are primarily in later layers. The ability to perform zero-/few-shot in-context learning on 14 different natural-language-processing (NLP) datasets/tasks stayed nearly intact when up to 70% (~15.7B parameters in OPT-66B) of the attention heads are removed.
The attention heads that are important (and unimportant) for in-context learning also seemed to overlap across tasks and shots. This indicates that a common task-agnostic subset of the attention heads is responsible for in-context learning. We also found that up to 20% of the FFNs (~8.5B parameters) can be removed with minimal decline in performance on zero-/few-shot in-context learning.
Our second analytic technique was to quantify the capacity of all attention heads in OPT-66B to perform a pair of task-agnostic primitive operations associated with in-context learning. Those primitives are prefix matching and copying: explicitly searching for a prior occurrence of the current token in context and copying over the token that succeeded it (its suffix).
Heads specialized for these two operations were first discovered by the machine learning research company Anthropic and termed induction heads. We found that a small set of heads in OPT-66B have nontrivial scores for both primitives. We also found that these heads overlap (to varying degrees) with the heads important for specific tasks identified earlier. This indicates that induction heads are capable of more sophisticated behaviors associated with in-context learning, such as latent concept matching, but are not the only heads with such capabilities.
Our overarching observation that only a core nucleus of attention heads and FFNs seem to be important for in-context learning indicates that OPT-66B and quite likely other prominent LLMs are undertrained. This also reinforces recent research that questions the efficacy of keeping the amount of pretraining data fixed when scaling models up, suggesting that the amount of pretraining data seen must be scaled hand-in-hand with the models themselves to attain optimal performance. It would be interesting to see how newer variants of LLMs released since the publication of our study, such as those tuned to follow instructions, fare in such analyses.