Hacker News Reader

Comment by skunkworker | Hacker News Reader

It's to the point now where using a basic LSTM is almost discouraged compared to using a Transformer. GPT-2 wouldn't have been possible without these recent innovations.

Here are some good guides on Transformers [1] and attention/ multi headed attention [2], as well as the paper that proposed the transformer model "Attention is all you need" [3]. GPT-2 heavily relies upon the advancements that transformers brought [4]

[1] http://jalammar.github.io/illustrated-transformer/

[2] https://towardsdatascience.com/attention-for-time-series-cla...

[3] Attention is all you need. https://arxiv.org/abs/1706.03762

[4] http://jalammar.github.io/illustrated-gpt2/

View on HN

It's to the point now where using a basic LSTM is almost discouraged compared to using a Transformer. GPT-2 wouldn't have been possible without these recent innovations.

So, but that "point" could just be that being a published academic means always using the latest thing. That doesn't demonstrate how much better these are, much less demonstrate that they represent fundamental steps forward.

This is mostly true for supervised and unsupervised learning models, but for reinforcement learning the LSTM is king because of the convenient fact that it can be evaluated one time step at a time, instead of just outputting a sequence like a transformer. For things like robotic control, etc, attention-based models are pretty nonsensical.

Not true, a transformer can be used in models without any lookahead, for example how it is used in gpt-2.! The real difference is the complexity of the model and the large increase in computational cost.

Except GPT-2 is irrelevant when it comes to time series specifically.

GPT-2 handles natural language which is specifically a time-series set, a sequence of natural word tokens. It is especially relevant because success in NLP tasks requires tracking and learning long-term relationships - not just between words in a phrase or two, but between words and phrases in different paragraphs, in different parts of the text, so that a general coherence of the whole document is kept. This exactly what is done with long-term relationships.

Then where is the publication that uses GPT-2 for a long-term NLP task, like, for example, reconstructing the rules of the great vowel shift from a corpus of all pre 20th century English?

Synchronously parsing the meaning a single text has at one moment in time involves no time series at all.

The parent comment isn't thinking time-series like you and I are.

Being able to follow multiple agents and correctly deduce their relationships at a given time t is very hard.

NLP "time-series" does a fine job at making back references within a text, but wouldn't be able to have multiple representations of a word or character through the years.

It's very hard to get the computer to say "ah, the context is 16th century, so here are the relationships" without fudging it / tailoring models via tailored corpuses.

Would adding a context vector for "16th century" be "fudging it"?

Depends. In the articles Uber example, the failure wasn't with detection, it as with context switching.

The detection kept changing, and so the model kept going "oh, new object, restart decision process."

Lacking the ability to generate and maintain it's own context is an area where a human would do better. We might not know what the object was, but our "slow down" response wouldn't keep resetting depending on what we classified the object as.

Same as words switching meanings within a piece or sentence. It's hard, but most humans can pickup when the usage changes

How about a model that can detect the semantic shift of "Amazon" from "river" to "company"? https://arxiv.org/abs/1703.00607

They're not even using transformers, just simple word embeddings.

I know that NLP sequences are analogous to time series but GPT-2 can't be applied to say financial time series for example. That's what I meant.

Or predicting the performance of a sports team over time.

From what I've seen, It breaks down at the embedding layer, because while the teams "remain the same" in name / dictionary, their actual relative relationship to each other varies season to season / week to week.

My response here is purely intuition, since I have never worked much with time series.

But wouldn't capturing that relationship require periodic retraining or other components to the network regardless? It may suggest that end-to-end training of a transformer is not suitable for these tasks, but that it might still capture the prediction of the long-scale time-series, if provided with extra data at each timestep in addition to the embeddings?

It does generate a long-term representation, the issue being that even with context and timestep specific data, that embedding is too general to make a good representation.

Sports are particularly problematic because almost all teams and statistics regress to the average at some point, meaning your generated future timestep context clues don't really help modify the embedding.

You're also dealing with variation within a season (injuries, better play, etc) and between seasons (personnel changes, rule changes, new stadiums, etc). So a team might have 4 seasons of above average performance, and then abruptly be the worst team in the league the next because they lost their coaches and star players.

Reply on news.ycombinator.com