Learning, evaluating and explaining (XAI) transferable text representations from grumpy Big Small Data.

Nils Rethmeier*⁺, Prof. Isabelle Augenstein*, Vageesh Saxena⁺

​

NLP with Friends • December 2, 2020

*

+

twitter.com/nils_rethmeier

XAI: explain & interpret defined by [1]

Decision understanding (mostly explain, model = black-box)

• explain input to output/ task relevance
• needs supervised labels
• incidental labels work [2]
• often explain 1 instance
• no idea what the model does

Model understanding (mostly interpret, grey-box methods)

• interpret neuron acitivity
• interpret over entire corpus
• irrelevant of output (unsupervised)
• prune unused neurons [3]
• find redundant neurons [4]

2

[1] Visual Interaction with Deep Learning Models through Collaborative Semantic Inference, 2019, arx:1907.10739

[2] Learning Comment Controversy Prediction in Web Discussions Using Incidentally Supervised Multi-Task CNNs, 2018, acl:W18-6246

[3] TX-Ray: Quantifying and Explaining Model-Knowledge Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, medarx:10.1101/2020.07.21.20157610v1

​

​

Do antidepressants correct the sleep disturbances in these diseases? …

src:lrpserver.hhi.fraunhofer.de/text-classification

[n] shown today, [n] shown not ours

topic label�medicine

disease

vomit

fiever

text

images

[5]

[3]

XAI: explain & interpret defined by [1]

Decision understanding (mostly explain, model = black-box)

• explain input to output/ task relevance
• needs supervised labels
• incidental labels work [2]
• often explain 1 instance
• no idea what the model does

Model understanding (mostly interpret, grey-box methods)

• interpret neuron acitivity
• interpret over entire corpus
• irrelevant of output (unsupervised)
• prune unused neurons [3]
• find redundant neurons [4]

3

[1] Visual Interaction with Deep Learning Models through Collaborative Semantic Inference, 2019, arx:1907.10739

[2] Learning Comment Controversy Prediction in Web Discussions Using Incidentally Supervised Multi-Task CNNs, 2018, acl:W18-6246

[3] TX-Ray: Quantifying and Explaining Model-Knowledge Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, medarx:10.1101/2020.07.21.20157610v1

​

​

Do antidepressants correct the sleep disturbances in these diseases? …

src:lrpserver.hhi.fraunhofer.de/text-classification

[n] shown today, [n] shown not ours

topic label�medicine

disease

vomit

fiever

text

images

[5]

[3]

[3] transfer XAI/ measurement

Why: Limits of supervised probes/ decision understanding XAI

• needs labels - evaluate expected semantics only
• unforseen insights not discoverable,
• probe can mismatch (pre-)training domain

​

What: Model transfer understanding: quantify neuron transfer

• RQ1: self-supervised ‘knowledge’ abstraction (pre-training)
• RQ2: zero-shot ‘knowledge’ transfer to new domain
• RQ3: transfer during supervised adaptation (fine-tuning)

​

How: Adapt self-supervised interpretability [5] for NLP [3]

• visualize what input (features) each neuron prefers (maximally activates on) - activation maximization from RBMs [5]

​

4

[3] TX-Ray: Quantifying and Explaining Model-Knowledge Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

[5] Visualizing Higher-Layer Features of a Deep Network, 2009, http://www.iro.umontreal.ca/~lisa/publications2/index.php/publications/show/247

​

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

Activation maximization, 2 methods

(1) In vision: generate input that maximally activates a neuron bc

​

​

​

we do not know visual primitives (objects, material, color etc.)

​

(2) In NLP:

5

[6] src:

distill.pub/2017/feature-visualization/

Activation maximization, 2 methods

(1) In vision: generate input that maximally activates a neuron bc

​

​

​

we do not know visual primitives (objects, material, color etc.)

​

(2) In NLP: record input (features

that maximally activate a neuron)

→ Neuron := preference distribution

over input tokens/ words

6

_cake

p_k = token activation probability, while training

I like cake better than cookies.

The learning-cake has has cherries inside.

Supervision is the cake’s frosting.

Self-supervision is the cake’s corpus.

The cake is a lie.

_cookies

​

​

Per token, only record the maximally activated (preferred) neuron

neuron 1

[6] src:

distill.pub/2017/feature-visualization/

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

Activation maximization, 2 methods

(1) In vision: generate input that maximally activates a neuron bc

​

​

​

we do not know visual primitives (objects, material, color etc.)

​

(2) In NLP: record input (features

that maximally activate a neuron)

→ Neuron := preference distribution

over input tokens/ words

7

_cake

p_k = token activation probability, while training

I like cake better than cookies.

The learning-cake has has cherries inside.

Supervision is the cake’s frosting.

Self-supervision is the cake’s corpus.

The cake is a lie.

_cookies

​

​

Per token, only record the maximally activated (preferred) neuron

_I

_than

neuron 1

[6] src:

distill.pub/2017/feature-visualization/

neuron 6

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

Activation maximization, 2 methods

(1) In vision: generate input that maximally activates a neuron bc

​

​

​

we do not know visual primitives (objects, material, color etc.)

​

(2) In NLP: record input (features

that maximally activate a neuron)

→ Neuron := preference distribution

over input tokens/ words

8

_cake

I like cake better than cookies.

The learning-cake has has cherries inside.

Supervision is the cake’s frosting.

Self-supervision is the cake’s corpus.

The cake is a lie.

_cookies

_I

_than

neuron 1

neuron 6

[6] src:

distill.pub/2017/feature-visualization/

neuron prefers stop words

neuron prefers sweets

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

9

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

10

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

FLAIR NLP pos tagger used on Wikitext-2 (2M tokens) → noisy tags

POS word cluster activation (normalized/ sum to 1)

unique words average activation (not normalized)

POS clusters: bc otherwise x-axis has 100k tokens → unreadable

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

11

• larger change epoch 1 vs 48.�no change epoch 48. vs 49.
• neuron well initialized?!
• very specific neuron (1 word only)

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

POS word cluster activation (normalized/ sum to 1)

unique words average activation (not normalized)

FLAIR NLP pos tagger used on Wikitext-2 (2M tokens) → noisy tags

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

12

little change (same word + POS)

→ high transfer

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

​

large change

→ little transfer

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

13

Post-hoc reasoning (not in paper):

self-supervision as next word prediction → messy feature preference

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

​

large change

→ little transfer

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

14

Post-hoc reasoning (not in paper):

self-supervision as next word prediction → messy feature preference

We often only care about transfer to supervised task i.e.

• representation has to transfer, not be intelligble

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

​

large change

→ little transfer

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

15

Post-hoc reasoning (not in paper):

self-supervision as next word prediction → messy feature preference

We often only care about transfer to supervised task i.e.

• representation has to transfer, not be intelligble

How neuron preference changes during pre-training (50) epochs:

​

1 vs 48. or 49. - loss converges at epoch 49

​

large change

→ little transfer

Uncomfortable question: Are learning objectives that produce intelligble representations better and would more self-supervised evaluation help?!

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ1: pretraining ‘builds’ knowledge

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

16

LSTM LM pre-training learns POS during early epochs - also shown 2018 by [6]

[6] Language Models Learn POS First, 2018 acl:W18-5438

​

​

​

via FLAIR tagger

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ2: apply knowledge to new inputs

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

17

feed new text Xend to frozen pre- trained encoder (1)

zero-shot transfer

no neuron transfer if

large change vs.

​

neuron transfers if small change vs.

​

zero-shot preference distribution

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ2: apply knowledge to new inputs

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

18

feed new text Xend to frozen pre- trained encoder (1)

zero-shot transfer

no neuron transfer if

large change vs.

​

neuron transfers if small change vs.

​

zero-shot preference distribution

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ3: transfer to + from supervision

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

19

feed new text Xend to the pre-trained encoder (1)

zero-shot transfer

supervised transfer

fine-tune encoder (1) with labels (supervised) only

Pretrain → supervise

• pretrain lang. model on Wikitext-2 (1)
• fine-tune on IMDB binary review labels (3)

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ3: transfer to + from supervision

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

20

pruning effects on supervised performance

zero-shot transfer

supervised transfer

Pretrain → supervise

• pretrain lang. model on Wikitext-2 (1)
• fine-tune on IMDB binary review labels (3)

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ3: transfer to + from supervision

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

21

zero-shot transfer

supervised transfer

Pretrain → supervise

• pretrain lang. model on Wikitext-2 (1)
• fine-tune on IMDB binary review labels (3)

pruning effects on supervised performance

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

RQ3: ‘forgetting’ due to supervision

How does each neuron abstract/ transfer knowledge during

RQ1: pretraining RQ2: zero-shot appl. RQ3: superv. fine-tune

22

zero-shot transfer

supervised transfer

Pretrain → supervise

• pretrain lang. model on Wikitext-2 (1)
• fine-tune on IMDB binary review labels (3)

pruning effects on supervised performance

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

TX-Ray take aways

can explore generalization and specialization at neuron-level

• self-supervised pre-training builds general knowledge
• preference spread across many neurons
• zero-shot application matches model knowledge vs new data
• preference less spread -- partial generalization
• supervised fine tuning, sparsifiers (concentrates) activation
• preference peaked and cocentrated -- domain over-specialization supervision forces to forget pretraining

​

​

23

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

TX-Ray take aways

can explore generalization and specialization at neuron-level

• self-supervised pre-training builds general knowledge
• preference spread across many neurons
• zero-shot application matches model knowledge vs new data
• preference less spread -- partial generalization
• supervised fine tuning, sparsifiers (concentrates) activation
• preference peaked and cocentrated -- domain over-specialization supervision forces to forget pretraining

​

​

24

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

TX-Ray take aways

can explore generalization and specialization at neuron-level

• self-supervised pre-training builds general knowledge
• preference spread across many neurons
• zero-shot application matches model knowledge vs new data
• preference less spread -- partial generalization
• supervised fine tuning, sparsifiers (concentrates) activation
• preference peaked and cocentrated -- domain over-specialization supervision forces to forget pretraining

​

​

25

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

XAI: further (newer) reading

XAI conceptualizing neurons/ NN components:

[7] Understanding the role of individual units in a deep neural network, PNAS,

2020. arxiv:2009.05041 (newer)

[8] Activation Atlas, 2019 distill.pub/2019/activation-atlas/ - Gave idea

that neurons should be explorable (but only worked for supervised learning). TX-Ray explores unsupervised resprentations → removes probing task limitation. Delay expectation bias.

​

Pruning:

[9] Movement Pruning: Adaptive Sparsity by Fine-Tuning, 2020 (newer)

arxiv:2005.07683 - pruns neurons that change during training/ transfer in

NLP, no XAI

26

Open questions/ caveats

Pruning:

• is forced forgetting → I’d expect to lose generalization beyond the test set
• when predicting many classes, with a small model,

pruning always hurts (MSc. thesis: Saxena)

So, is pruning only for overparameterized models?

​

XAI:

• Explainability is sensitive to supervised learning quality
• TX-Ray (interpretability) + ‘its’ pruning are sensitive to self- supervised learning quality (see early episodes)
• However, if sensitivity means worse abstractions, then could better abstractions indicate better self-supervised learning -- i.e. allow evaluation without probes?

27

[3] TX-Ray: Quantifying and Explaining Model-Knowledge

Transfer in (Un-)Supervised NLP, 2020, arx:1912.00982

TX-Ray in patient care prediction [4]

Evaluate preferred data source and model component for a task

28

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings,

2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

​

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

Task: will patient die in 24h?

activation

1 hour bucket pooling -- low layer

pooling over hour buckets -- higher layers

limited training data: 1. low-resource pretraining + 2. XAI, since no XL pretrained model available

TX-Ray in patient care prediction [4]

Evaluate preferred data source and model component for a task

29

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

Insights:

1. output events data not used

Task: will patient die in 24h?

activation

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings,

2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

​

limited training data: 1. low-resource pretraining + XAI, since no XL pretrained model available

TX-Ray in patient care prediction [4]

Evaluate preferred data source and model component for a task

30

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

Insights:

1. output events data not used

​

pooling over hour bucket matters

​

1 hour bucket pooling -- low layer

Task: will patient die in 24h?

activation

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings,

2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

​

limited training data: 1. low-resource pretraining + XAI, since no XL pretrained model available

TX-Ray in patient care prediction [4]

Evaluate preferred data source and model component for a task

31

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings, 2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

Insights:

1. output events data not used

​

pooling over hour bucket matters

1. for pooling over time,

pooling matters most

1 hour bucket pooling -- low layer

Task: will patient die in 24h?

pooling over hour buckets -- higher layers

activation

[4] EffiCare: Better Prognostic Models via Resource-Efficient Health Embeddings,

2020, AMIA Symposium, medarx:10.1101/2020.07.21.20157610v1

​

limited training data: 1. low-resource pretraining + XAI, since no XL pretrained model available

Evaluation: approaches, questions

Supervised: “Catch them all” supervised benchmarks

• example: [10] - 18 tasks word embedding benchmark
• (complete) autoencoder retrofits word embeddings
• better single task + all task performance
• more improvement on smaller pre-training data

Problem:

• probe model (depth) limits performance [10, 11]

32

32

[10] MoRTy: Unsupervised Learning of Task-specialized Word Embeddings by Autoencoding, 2020, acl:W19-4307, github.com/NilsRethmeier/MoRTy

[11] Designing and Interpreting Probes with Control Tasks, 2019, acl:D19-1275

Evaluation: approaches, questions

Supervised evaluation via probe design problems continued

• probe models limit performance [11] -- deep probes are better

Nagging question 2: How much do probe benchmarks tell us then?

• SotA scores depend on:
• hardware budget [12] or seed hacking [13]
• using ever more pre-training data [14, 15]

Nagging question 3: are new pretraining models better or just use more data -- lack of controlled experiments!

​

1990-2010: learning is easy when you have enough (supervised) data

• simple, small models will do

2018-20xx: learning is easy when you have web-scale (self-supervised) data

• simple, XL models [14, 15]

33

33

[11] Designing and Interpreting Probes with Control Tasks, 2019, acl:D19-1275, [12] The Hardware Lottery, 2020, arx:2009.06489

[13] Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping, 2020

[14] A Survey on Contextual Embeddings, 2020,

[15] How can we accelerate progress towards human-like linguistic generalization, 2020. acl:http2020.acl-main.465.

Evaluation: approaches, questions

Supervised evaluation via probe design problems continued

• probe models limit performance [11] -- deep probes are better

Nagging question 2: How much do probe benchmarks tell us then?

• SotA scores depend on:
• hardware budget [12] or seed hacking [13]
• using ever more pre-training data [14, 15]

Nagging question 3: are new pretraining models better or just use more data -- lack of controlled experiments!

​

1990-2010: learning is easy when you have enough (supervised) data

• simple, small models will do

2018-20xx: learning is easy when you have web-scale (self-supervised) data

• simple, XL models [14, 15]

34

34

[11] Designing and Interpreting Probes with Control Tasks, 2019, acl:D19-1275, [12] The Hardware Lottery, 2020, arx:2009.06489

[13] Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping, 2020

[14] A Survey on Contextual Embeddings, 2020,

[15] How can we accelerate progress towards human-like linguistic generalization, 2020. acl:http2020.acl-main.465.

Learning from small data

Evaluation idea: addressing [16-20]

• evaluate small data pre-training for few and zero-shot learning
• no large external pre-training
• pre-train on a few GPUs (~1-4)
• long-tailed (imbalanced) data
• tail is always few-shot to zero-shot and tail grows with data size

​

Problem: how to learn a text encoder at small scale?

1. We need zero-shot + long-tail capabilities
• e.g. label-embedding CNNs [22], where labels = word embeddings
• [23] label-embedding BERT needs ELMO word embeddings for labels
2. CNNs/ LSTMs [20] outperform transformers when

using small pre-training data [16, 18-21]

Considering 1+2: we chose CNNs

35

35

[16] Learning and evaluating general linguistic intelligence., 2020, arx:2010.01061 [17] How can we accelerate progress towards human-like linguistic

generalization, 2020. acl:http2020.acl-main.465. [18] Mogrifier LSTM, 2020, [19] Transformer on a Diet, 2020, [20] Character-aware neural language models. 2016. abs/1508.06615, [21] Pointer Sentinel Mixture Models, 2016, [22] Multi-task label embedding for text classification. 2018

[23] X-BERT: eXtreme Multi-label Text Classification with BERT,

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Self-supervised CNN pre-training (e.g. as SSL transformer alternative)

36

36

feed words ‘measuring..’ + positive labels + negative labels through embedding layer (1)

feed word embeddings trough text encoder NN → text encoding T

and label(-word)-embedings through label encoder NN → label encoding L

concatenate each label encoding L with a copy if the text encoding T -- T^ktorch.repeat(T, |L₁ … L_k|)

match label-encodings L to text-copies T^k -- matcher = learned similarity function with sim = 0..1

• matcher is a single class that predicts k label to text ‘similarities’

train with binary cross entropy over k labels * batchsize (k = num pos + neg labels)

• (a) labels = supervision labels or (b) words as pseudo labels (sampled from text in batch)

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Supervised and self-supervised training objectives

​

• noise contrastive estimation NCE
• NCE is BCE with 1 positive and k negative samples
• we use g positives + b negatives: N = g + b

​

​

​

• previous work used supervised labels g, and optinally b [24]
• we add positive w⁺ + negative input words w^- as pseudo-labels
• N = g + b for supervised learning mode
• N = w⁺ + w^- for self-supervised learning
• i.e. contrastive partial autoencoding

​

​

detail we use a CNN and deeper encoders than [25]

37

37

[24] Multi-task label embedding for text classification. 2018

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Pseudo-label embedding NCE (PLENCE) properties:

​

• self-supervised (SSL) pretraining → allows SSL zero-shot learning

​

• NCE allows infite label sets and vocabulary sets
• no softmax → one less bottleneck
• ‘text-to-text’ becomes ‘embedding-to-embedding’
• networks does not have to translate to anonymous {0,1} labels
• labels now have meaning (semantics) and are words (like for humans)
• we pretrain word embeddings, ergo, labels are also pretrained :)
• unknown labels can be inferred via FastText, or Attentive Mimicking [25]

38

38

[25] Learning Semantic Representations for Novel Words: Leveraging Both Form and Context. 2018

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Evaluate zero-shot ZS, few-shot + long-tail learning via PLENCE:

​

allowing less pre-training text

75% … 10%

• takes longer to train
• but reaches comparable ZS performance

3.3x self-supervision pseudo-labels

• better/ faster ZS performance

Using larger net + more pseudo-labels

• much better ZS performance

​

​

39

39

Insight: when large, external data (signal) is unavailable, increase self-supervison

works despite long-tail problem worsening with few-shot

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Evaluate zero-shot ZS, few-shot + long-tail learning via PLENCE:

​

allowing less pre-training text

75% … 10%

• takes longer to train
• but reaches comparable ZS performance

3.3x self-supervision pseudo-labels

• better/ faster ZS performance

Using larger net + more pseudo-labels

• much better ZS performance

​

​

40

40

Insight: when large, external data (signal) is unavailable, increase self-supervison

works despite long-tail problem worsening with few-shot

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Pre-training from small data

Evaluate zero-shot, few-shot FS + long-tail learning via PLENCE:

​

with self-supervised pretraining (PLENCE)

• FS performance is much better
• trains fast, without pseudo plateaus
• epoch 1 performance is much higher -- practical

without self-supervised pretraining:

• few-shot performance is bad
• we take long (200 epochs to) learn
• FS learning pseudo plateaus

​

​

41

41

Insight: pseudo-label SSL pretraining boost data efficiency, learning speed, learning stability and few-shot end performance -- even on 10% data

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

pseudo plateaus

stable learning curves

Pre-training from small data

Evaluate zero-shot, few-shot FS + long-tail learning via PLENCE:

​

with self-supervised pretraining (PLENCE)

• FS performance is much better
• trains fast, without pseudo plateaus
• epoch 1 performance is much higher -- practical

without self-supervised pretraining:

• few-shot performance is bad
• we take long (200 epochs to) learn
• FS learning pseudo plateaus

​

​

42

42

Insight: pseudo-label SSL pretraining boost data efficiency, learning speed, learning stability and few-shot end performance -- even on 10% data

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

pseudo plateaus

stable learning curves

Pre-training from small data

Evaluate zero-shot ZS, few-shot + long-tail learning via PLENCE:

​

43

43

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

Evaluating a long-tail:

• 5 class frequency balanced buckets
• i.e. each bucket has same amount of positive (real) labels
• tail buckets have many more labels
• classes per bucket still very imbalanced

y-axis class label frequency in log -- extreme imbalance, 1305 classes

Pre-training from small data

Evaluate zero-shot ZS, few-shot + long-tail learning via PLENCE:

​

44

44

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

SL Sup label-embeddings (LE)

• better performance than [0,1]-BCE, esp. on long-tail

SSL prertrained LE:

• slightly better yet

XL SSL pretrained LE with larger net + 3.3x labels:

• strong long-tail performance boost

Pre-training from small data

Evaluate zero-shot ZS, few-shot + long-tail learning via PLENCE:

​

45

45

[24] Long-Tail Zero and Few-Shot Learning via Contrastive Pretraining on and for Small Data, 2020, arx:2010.01061

​

SSL pretraining label embs:

• learn long-tail from epoch 1

XL SSL pretrained LE:

• learns long-tail even faster
• need 3x less epochs compared to

Insight: pseudo-label SSL pretraining boost long-tail learning.

Larger model + more SSL signal boost long-tail further.

Summary

self-supervised XAI:

• for self-supervised (transfer learning) evaluation
• without limiting to supervised probing (expected) semantics

challenging, low-resource evaluation:

• zero + few-shot reveal pretraining effects
• low-resource (LR), long-tail (LT) setups show learning efficiency
• long-tail problem grows with data size
• in the tail we are always in the low-resource setting

data efficient learning:

• more self-supervision boosts LR few + zero-shot performance

→ Evidence that instead of large external data pretraining, we can

just boost SSL learning signal on small data

​

​

46

46

FIN

47

47

XAI: further/ newer reading

[7] Understanding the role of individual units

in a deep neural network, PNAS, 2020.

arxiv:2009.05041

​

​

​

​

​

​

​

48

An image classification CNN

Redundancy in represenations

49