Finetuning Pretrained Transformers into Variational Autoencoders
This work addresses a bottleneck in text VAE research by making Transformer-based VAEs more accessible without extensive computing, though it is incremental as it builds on existing techniques.
The paper tackles the problem of posterior collapse in text variational autoencoders (VAEs) by proposing a two-phase finetuning method to convert pretrained Transformers into VAEs, achieving competitive results with massively pretrained models in some metrics while requiring fewer resources.
Text variational autoencoders (VAEs) are notorious for posterior collapse, a phenomenon where the model's decoder learns to ignore signals from the encoder. Because posterior collapse is known to be exacerbated by expressive decoders, Transformers have seen limited adoption as components of text VAEs. Existing studies that incorporate Transformers into text VAEs (Li et al., 2020; Fang et al., 2021) mitigate posterior collapse using massive pretraining, a technique unavailable to most of the research community without extensive computing resources. We present a simple two-phase training scheme to convert a sequence-to-sequence Transformer into a VAE with just finetuning. The resulting language model is competitive with massively pretrained Transformer-based VAEs in some internal metrics while falling short on others. To facilitate training we comprehensively explore the impact of common posterior collapse alleviation techniques in the literature. We release our code for reproducability.