Prioritized Training on Points that are Learnable, Worth Learning, and Not Yet Learnt

Training on web-scale data can take months. But most computation and time is wasted on redundant and noisy points that are already learnt or not learnable. To accelerate training, we introduce Reducible Holdout Loss Selection (RHO-LOSS), a simple but principled technique which selects approximately those points for training that most reduce the model's generalization loss. As a result, RHO-LOSS mitigates the weaknesses of existing data selection methods: techniques from the optimization literature typically select 'hard' (e.g. high loss) points, but such points are often noisy (not learnable) or less task-relevant. Conversely, curriculum learning prioritizes 'easy' points, but such points need not be trained on once learned. In contrast, RHO-LOSS selects points that are learnable, worth learning, and not yet learnt. RHO-LOSS trains in far fewer steps than prior art, improves accuracy, and speeds up training on a wide range of datasets, hyperparameters, and architectures (MLPs, CNNs, and BERT). On the large web-scraped image dataset Clothing-1M, RHO-LOSS trains in 18x fewer steps and reaches 2% higher final accuracy than uniform data shuffling.

Mitigating harm in language models with conditional-likelihood filtration

Language models trained on large-scale unfiltered datasets curated from the open web acquire systemic biases, prejudices, and harmful views from their training data. We present a methodology for programmatically identifying and removing harmful text from web-scale datasets. A pretrained language model is used to calculate the log-likelihood of researcher-written trigger phrases conditioned on a specific document, which is used to identify and filter documents from the dataset. We demonstrate that models trained on this filtered dataset exhibit lower propensity to generate harmful text, with a marginal decrease in performance on standard language modeling benchmarks compared to unfiltered baselines. We provide a partial explanation for this performance gap by surfacing examples of hate speech and other undesirable content from standard language modeling benchmarks. Finally, we discuss the generalization of this method and how trigger phrases which reflect specific values can be used by researchers to build language models which are more closely aligned with their values.

Prioritized training on points that are learnable, worth learning, and not yet learned

We introduce Goldilocks Selection, a technique for faster model training which selects a sequence of training points that are "just right". We propose an information-theoretic acquisition function -- the reducible validation loss -- and compute it with a small proxy model -- GoldiProx -- to efficiently choose training points that maximize information about a validation set. We show that the "hard" (e.g. high loss) points usually selected in the optimization literature are typically noisy, while the "easy" (e.g. low noise) samples often prioritized for curriculum learning confer less information. Further, points with uncertain labels, typically targeted by active learning, tend to be less relevant to the task. In contrast, Goldilocks Selection chooses points that are "just right" and empirically outperforms the above approaches. Moreover, the selected sequence can transfer to other architectures; practitioners can share and reuse it without the need to recreate it.

Add a SideNet to your MainNet

As the performance and popularity of deep neural networks has increased, so too has their computational cost. There are many effective techniques for reducing a network's computational footprint (quantisation, pruning, knowledge distillation), but these lead to models whose computational cost is the same regardless of their input. Our human reaction times vary with the complexity of the tasks we perform: easier tasks (e.g. telling apart dogs from boat) are executed much faster than harder ones (e.g. telling apart two similar looking breeds of dogs). Driven by this observation, we develop a method for adaptive network complexity by attaching a small classification layer, which we call SideNet, to a large pretrained network, which we call MainNet. Given an input, the SideNet returns a classification if its confidence level, obtained via softmax, surpasses a user determined threshold, and only passes it along to the large MainNet for further processing if its confidence is too low. This allows us to flexibly trade off the network's performance with its computational cost. Experimental results show that simple single hidden layer perceptron SideNets added onto pretrained ResNet and BERT MainNets allow for substantial decreases in compute with minimal drops in performance on image and text classification tasks. We also highlight three other desirable properties of our method, namely that the classifications obtained by SideNets are calibrated, complementary to other compute reduction techniques, and that they enable the easy exploration of compute accuracy space.