About this event

How does model architecture, pre-training objective, the side of the dataset and parameter count affect model’s linguistic abilities? They don’t 🤯. Or at least not as directly we thought. Evaluation of generated text remains a significant issue. Recently-introduced model-based metrics have shown promising results compared to n-gram-based metrics like BLEU, yet they still suffer severe drawbacks (http://arxiv.org/abs/2205.10696).

Speaker

Vladislav Lialin is a computer science PhD student at University of Massachusetts Lowell advised by Anna Rumshisky. His research areas include continual learning for large language models, multimodal learning, and model analysis. In particular, he is hyping large-scale models and thinks that every task is a language modeling task if you try hard enough. He is currently interning at Amazon Alexa AI.