Vortrag: Are neural networks all you need?

A critical perspective on neural networks and evaluation benchmarks

In the past few years we have seen rapid growth of neural-network based architectures solving a variety of NLP tasks with impressive benchmarks. However, very little has been said about the "pipeline code" enabling these improvements. This talk examines why neural networks by themselves are not the singular drive of improvement.

This talk will be streamed online and shown in the indicated room.

While neural networks have solid applications, when we talk about neural architectures, the code that is used to enable training objectives is seldomly discussed. In this talk we will examine a number of pre-training and post-training steps of word2vec, GloVe and SentenceTransformer, and how they can be used to improve traditional (e.g BOW) models. We will see how we can use "clever tricks" found while experimenting with neural networks and apply them directly to distributional problems such as word and sentence embeddings. We will also take a critical look at the underlying evaluation benchmarks and how neural networks perform well.

Info

Tag: 28.10.2023
Anfangszeit: 12:00
Dauer: 00:25
Raum: NIG Raum 2
Track: Computational Linguistics
Sprache: en

Links:

Feedback

Uns interessiert Ihre Meinung! Wie fanden Sie diese Veranstaltung?

Gleichzeitige Events