m8ta
You are not authenticated, login.
text: sort by
tags: modified
type: chronology
{1569}
hide / / print
ref: -2022 tags: symbolic regression facebook AI transformer date: 05-17-2022 20:25 gmt revision:0 [head]

Deep symbolic regression for recurrent sequences

Surprisingly, they do not do any network structure changes; it’s Vaswini 2017w/ a 8-head, 8 layer transformer (sequence to sequence, not decoder only) with a latent dimension of 512.  Significant work was in feature / representation engineering (e.g. base-10k representations of integers and fixed-precision representations of floating-point numbers. (both of these involve a vocabulary size of ~10k ... amazing still that this works..)) + the significant training regimen they worked with (16 Turing GPUs, 32gb ea).  Note that they do perform a bit of beam-search over the symbolic regressions by checking how well each node fits to the starting sequence, but the models work even without this degree of refinement. (As always, there undoubtedly was significant effort spent in simply getting everything to work)

The paper does both symbolic (estimate the algebraic recurence relation) and numeric (estimate the rest of the sequence) training / evaluation. Symbolic regression generalizes better, unsurprisingly. But both can be made to work even in the presence of (log-scaled) noise!

Analysis of how the transformers work for these problems is weak; only one figure showing that the embeddings of the integers follows some meandering but continuous path in t-SNE space. Still, the trained transformer is able to usually best hand-coded sequence inference engine(s) in Mathematica, and does so without memorizing all of the training data. Very impressive and important result, enough to convince that this learned representation (and undiscovered cleverness, perhaps) beats human mathematical engineering, which probably took longer and took more effort.

It follows, without too much imagination (but vastly more compute), that you can train an 'automatic programmer' in the very same way.

{1556}
hide / / print
ref: -0 tags: concept net NLP transformers graph representation knowledge date: 11-04-2021 17:48 gmt revision:0 [head]

Symbolic Knowledge Distillation: from General Language Models to Commonsense Models

  • From a team at University of Washington / Allen institute for artificial intelligence/
  • Courtesy of Yannic Kilcher's youtube channel.
  • General idea: use GPT-3 as a completion source given a set of prompts, like:
    • X starts running
      • So, X gets in shape
    • X and Y engage in an argument
      • So, X wants to avoid Y.
  • There are only 7 linkage atoms (edges, so to speak) in these queries, but of course many actions / direct objects.
    • These prompts are generated from the Atomic 20-20 human-authored dataset.
    • The prompts are fed into 175B parameter DaVinci model, resulting in 165k examples in the 7 linkages after cleaning.
    • In turn the 165k are fed into a smaller version of GPT-3, Curie, that generates 6.5M text examples, aka Atomic 10x.
  • Then filter the results via a second critic model, based on fine-tuned RoBERTa & human supervision to determine if a generated sentence is 'good' or not.
  • By throwing away 62% of Atomic 10x, they get a student accuracy of 96.4%, much better than the human-designed knowledge graph.
    • They suggest that one way thins works is by removing degenerate outputs from GPT-3.

Human-designed knowledge graphs are described here: ConceptNet 5.5: An Open Multilingual Graph of General Knowledge

And employed for profit here: https://www.luminoso.com/

{1548}
hide / / print
ref: -2021 tags: gated multi layer perceptrons transformers ML Quoc_Le Google_Brain date: 08-05-2021 06:00 gmt revision:4 [3] [2] [1] [0] [head]

Pay attention to MLPs

  • Using bilinear / multiplicative gating + deep / wide networks, you can attain similar accuracies as Transformers on vision and masked language learning tasks! No attention needed, just a in-network multiplicative term.
  • And the math is quite straightforward. Per layer:
    • Z=σ(XU),,Z^=s(Z),,Y=Z^V Z = \sigma(X U) ,, \hat{Z} = s(Z) ,, Y = \hat{Z} V
      • Where X is the layer input, σ\sigma is the nonlinearity (GeLU), U is a weight matrix, Z^\hat{Z} is the spatially-gated Z, and V is another weight matrix.
    • s(Z)=Z 1(WZ 2+b) s(Z) = Z_1 \odot (W Z_2 + b)
      • Where Z is divided into two parts along the channel dimension, Z 1Z 2Z_1 Z_2 . 'circleDot' is element-wise multiplication, and W is a weight matrix.
  • You of course need a lot of compute; this paper has nice figures of model accuracy scaling vs. depth / number of parameters / size. I guess you can do this if you're Google.

Pretty remarkable that an industrial lab freely publishes results like this. I guess the ROI is that they get the resultant improved ideas? Or, perhaps, Google is in such a dominant position in terms of data and compute that even if they give away ideas and code, provided some of the resultant innovation returns to them, they win. The return includes trained people as well as ideas. Good for us, I guess!

{1440}
hide / / print
ref: -2017 tags: attention transformer language model youtube google tech talk date: 02-26-2019 20:28 gmt revision:3 [2] [1] [0] [head]

Attention is all you need

  • Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin
  • Attention is all you need neural network models
  • Good summary, along with: The Illustrated Transformer (please refer to this!)
  • Łukasz Kaiser mentions a few times how fragile the network is -- how easy it is to make something that doesn't train at all, or how many tricks by google experts were needed to make things work properly. it might be bravado or bluffing, but this is arguably not the way that biology fails.
  • Encoding:
  • Input is words encoded as 512-length vectors.
  • Vectors are transformed into length 64 vectors: query, key and value via differentiable weight matrices.
  • Attention is computed as the dot-product of the query (current input word) with the keys (values of the other words).
    • This value is scaled and passed through a softmax function to result in one attentional signal scaling the value.
  • Multiple heads' output are concatenated together, and this output is passed through a final weight matrix to produce a final value for the next layer.
    • So, attention in this respect looks like a conditional gain field.
  • 'Final value' above is then passed through a single layer feedforward net, with resnet style jump.
  • Decoding:
  • Use the attentional key value from the encoder to determine the first word through the output encoding (?) Not clear.
  • Subsequent causal decodes depend on the already 'spoken' words, plus the key-values from the encoder.
  • Output is a one-hot softmax layer from a feedforward layer; the sum total is differentiable from input to output using cross-entropy loss or KL divergence.