Guiding Transformers to Process in Steps

29 Sep 2021  ·  Simas Sakenis, Stuart Shieber ·

Neural networks have matched or surpassed human abilities in many tasks that humans solve quickly and unconsciously, i.e., via Kahneman's “System 1”, but have not been as successful when applied to “System 2” tasks that involve conscious multi-step reasoning. In this work, we argue that the kind of training that works for System 1 tasks is not sufficient for System 2 tasks, propose an alternative, and empirically demonstrate its effectiveness. Specifically, while learning a direct mapping from inputs to outputs is feasible for System 1 tasks, we argue that algorithmic System 2 tasks can only be solved by learning a mapping from inputs to outputs through a series of intermediate steps. We first show that by using enough intermediate steps a 1-layer 1-head Transformer can in principle compute any finite function, proving the generality of the approach. We then show empirically that a 1-layer 1-head Transformer cannot learn to compute the sum of binary numbers directly from the inputs, but is able to compute the sum when trained to first generate a series of intermediate results. This demonstrates, at a small scale, how a fixed-size neural network can lack the expressivity to encode the direct input-output mapping for an algorithmic task and yet be fully capable of computing the outputs through intermediate steps. Finally, we show that a Frozen Pretrained Transformer is able to learn binary addition when trained to compute the carry bits before the sum, while it fails to learn the task without using intermediates. These results indicate that explicitly guiding the neural networks through the intermediate computations can be an effective approach for tackling algorithmic tasks.

PDF Abstract
No code implementations yet. Submit your code now

Tasks


Datasets


  Add Datasets introduced or used in this paper

Results from the Paper


  Submit results from this paper to get state-of-the-art GitHub badges and help the community compare results to other papers.

Methods