![DL architecture repair
SE vs DL community
18
[AutoTrainer] Xiaoyu Zhang, Juan Zhai, Shiqing Ma, Chao Shen: AUTOTRAINER: An Automatic DNN Training Problem Detection and Repair System. ICSE, pp. 359-371, 2021.
[Hebo] A. I. Cowen-Rivers, W. Lyu, R. Tutunov, Z. Wang, A. Grosnit, R. R. Gri
ffi
ths, A. M. Maraval, H. Jianye, J. Wang, J. Peters, Haitham Bou Ammar: Hebo: Pushing the limits of
sample-e
ffi
cient hyper-parameter optimisation, JAIR, vol. 74, pp. 1269-1349, 2022.
[Bohb] S. Falkner, A. Klein, and F. Hutter: Bohb: Robust and e
ffi
cient hyperparameter optimization at scale. ICML, pp. 1437-1446, 2018.
Fit surrogate
Sample hyperparameters
Train & evaluate model
Initial hyperparameters
Bayesian hyperparameter optimization [Hebo, Bohb]
Until training
budget is
over
Rule based repair [AutoTrainer]
Symptom → Solution
S1: Add batch normalisation layer
S2: Substitute activation function
S3: Add gradient clip
S4: Substitute initialiser
S5: Adjust batch size
S6: Adjust learning rate
S7: Substitute optimizer
Hyperparameters
H1: activation function
H2: optimizer
H3: batch size
H4: number of epochs
H5: loss function
H6: learning rate
H7: weight initialization
H8: batching (enable/disable)
H9: number of neurons in a layer](https://image.slidesharecdn.com/tonella-icse2023-230522144128-46752760/85/The-Road-Toward-Dependable-AI-Based-Systems-18-320.jpg)
![DL inputs
Auto-encoder validity
23
latent
space
Encoder Decoder
Training set
Anomaly set
[ ]
(e >
𝜃
)?
“All three testing techniques [DeepExplore, DLFuzz,
DeepConcolic] studied produced signi
fi
cant numbers of
[autoencoder] invalid tests”
Input validity approximated as Auto-encoder validity](https://image.slidesharecdn.com/tonella-icse2023-230522144128-46752760/85/The-Road-Toward-Dependable-AI-Based-Systems-23-320.jpg)
The Road Toward Dependable AI Based Systems
- 1. Paolo Tonella Software Institute, Università della Svizzera italiana, Lugano, Switzerland The Road Toward Dependable AI Based Systems 1 https://www.pre-crime.eu
- 2. Precrime ERC project: https://pre-crime.eu 2 Self-assessment Oracles for Anticipatory Testing “Testing the unexpected” • Anticipate failures due to unexpected conditions • Identify unexpected execution contexts at runtime • Generate valid, but unexpected execution scenarios
- 3. Deep neural networks High accuracy on human competitive tasks 3
- 4. Deep neural networks Integrated into complex applications 4
- 5. Deep neural networks Programming model 5 0.9 0.1 Train set Validation set Test set W • Data driven programming/no control logic • Numeric input/numeric output • Non deterministic training process is_red 0.1 0.9 is_red Is this a bug?
- 6. Deep neural networks Learned behaviour vs programmed behaviour 6 Training set Network architecture Hyperparameters Training process Code Code acc = 0.99 x≥0 y・y = x 0.1 0.9 is_red Is this a bug?
- 7. Testing fundamentals Goal of testing Inputs Program Fault Execution Failure Oracle 7 Repair faults, exposed by choosing inputs whose execution violates oracles Repair DL faults and failures Q1: What is a DL fault? Q2: What is a unique DL failure? Q3: What does it mean to repair a DL fault? DL inputs Q4: How to generate DL test inputs? Q5: How to prioritize/select DL test inputs? DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations?
- 8. DL faults 8 Train/validation set Hyperparameters Code Data Architecture IEEE std glossary: fault = “an incorrect step, process or data de fi nition in a computer program” DL faults and failures Q1: What is a DL fault? Q2: What is a unique DL failure? Q3: What does it mean to repair a DL fault?
- 9. Taxonomy of Real Faults in Deep Learning Systems MODEL (29+45) GPU USAGE (10+1) API (20+0) TENSORS & INPUTS (53+20) TRAINING (37+160) Model Type & Properties (6+20) Layers (23+25) Activation Function (3+2) wrong model initialisation (1+2) wrong weights initialisation (1+0) wrong selection of model (2+1) multiple initialisations of CNN (1+0) suboptimal network structure (1+15) wrong network architecture (0+2) wrong type of activation function (1+2) missing softmax activation function (1+0) missing relu activation function (1+0) wrong input sample size for linear layer (5+0) wrong defined input shape (2+0) wrong amount & type of pooling in convolutional layer (0+1) wrong defined output shape (3+1) wrong defined input & output shape (1+0) wrong filter size for a convolutional layer (1+1) layers' dimensions mismatch (0+9) suboptimal number of neurons in the layer (0+6) bias needed in a layer (1+0) missing destination GPU device (1+0) incorrect state sharing (1+0) wrong reference to GPU device (2+0) missing transfer of data to GPU (1+0) wrong data parallelism on GPUs (1+0) calling unsupported operations on CUDA tensors (1+0) conversion to CUDA tensor inside the training/test loop (1+0) wrongly implemented data transfer function (CPU-GPU) (0+1) wrong position of data shuffle operation (1+0) deprecated API (1+0) wrong usage of image decoding API (1+0) wrong usage of placeholder restoration API (1+0) missing argument scoping (1+0) wrong tensor transfer to GPU (1+0) missing global variables initialisation (3+0) wrong API usage (10+0) missing API call (1+0) wrong reference to operational graph (1+0) Wrong Tensor Shape (21+5) Wrong Input (32+15) wrong tensor shape (missing squeeze) (5+0) wrong tensor shape (wrong indexing) (2+0) wrong tensor shape (wrong output padding) (1+0) wrong tensor shape (other) (13+3) tensor shape mismatch (0+2) Wrong Shape of Input Data (22+7) Wrong Type of Input Data (5+3) Wrong Input Format (5+5) wrong shape of input data for a method (6+0) wrong shape of input data for a layer (16+2) wrong shape of input data (0+5) wrong type of input data for a method (4+0) wrong type of input data for a layer (1+0) wrong type of input data (0+3) wrong input format (1+5) wrong input format for RNN (2+0) wrong format of passed weights (1+0) incompatible tensor type (1+0) Hyperparameters (10+26) suboptimal number of epochs (2+4) data batching required (2+7) suboptimal batch size (2+1) wrongly implemented data batching (1+0) missing regularisation (loss and weight) (0+1) suboptimal learning rate (1+5) suboptimal hyper- parameters tuning (2+8) Loss Function (7+16) missing masking of invalid values to zero (1+0) wrong loss function calculation (5+4) missing loss function (0+1) wrong selection of loss function (1+11) Validation/Testing (2+4) missing validation set (1+0) incorrect train/test data split (0+3) wrong performance metric (1+1) Preprocessing of Training Data (13+37) Missing Preprocessing (11+22) missing preprocessing step (subsampling, normalisation, input scaling, resize of the images, oversampling, encoding of categorical data, padding, ... skip ..., data shuffling, interpolation) Wrong Preprocessing (2+15) wrong preprocessing step (pixel encoding, padding, text segmentation, normalisation, ... skip ..., positional encoding, character encoding) Optimiser (3+3) wrong optimisation function (1+3) epsilon for Adam optimiser too low (2+0) Training Data Quality (2+60) low quality training data (0+11) not enough training data (0+14) overlapping output classes in training data (0+1) too many output categories (0+1) small range of values for a feature (0+2) unbalanced training data (0+11) wrong selection of features (1+6) wrong labels for training data (1+12) discarding important features (0+2) Training Process (0+14) model too big to fit into available memory (0+5) reference for non-existing checkpoint (0+1) missing data augmentation (0+3) redundant data augmentation (0+1) wrong management of memory resources (0+4) missing dropout layer (1+1) missing normalisation layer (1+0) missing average pooling layer (0+1) missing softmax layer (1+0) redundant softmax layer (1+0) wrong layer type (1+2) wrong type of pooling layer (0+1) missing dense layer (1+0) missing flatten layer (1+0) Missing/Redundant/ Wrong Layer (7+5) Layer Properties (13+18) GPU tensor is used instead of CPU tensor (1+0) Taxonomy of real faults Repository mining + interviews with developers 9 Stack Over fl ow: 477 discussions Github: 271 issues and pull requests Github: 311 commits Interviews: 20 researchers and practitioners Validation Survey: 21 researchers and practitioners
- 13. DeepCrime Mutation testing of DL systems 13 Training set Network architecture Hyperparameters DeepCrime Re-train Accuracy = 85% Accuracy = 95% https://github.com/deepcrime-tool/DeepCrime
- 14. DeepCrime Mutation operators 14 Training Data Change labels of training data TCL Remove portion of training data TRD Unbalance training data TUD Add noise to training data TAN Make output classes overlap TCO Hyperparameters Change batch size HBS Decrease learning rate HLR Change number of epochs HNE Disable data batching HDB Activation Function Change activation function ACH Remove activation function ARM Add activation function to layer AAL Regularisation Add weights regularisation RAW Change weights regularisation RCW Remove weights regularisation RRW Change dropout rate RCD Change patience parameter RCP Weights Change weights initialisation WCI Add bias to a layer WAB Remove bias from a layer WRB Loss function Change loss function LCH Optimisation Function Change optimisation function OCH Change gradient clipping OCG Validation Remove validation set VRM Uses of DL mutation Evaluate adequacy of a test set Compare test input generators Benchmark repair tools Assess test ordering produced by prioritization techniques
- 15. DL failures Failure counting is not trivial 15 ❌ 6 ❌ ❌ ❌ ❌ 6 6 6 6 Diversity matters! DL faults and failures Q1: What is a DL fault? Q2: What is a unique DL failure? Q3: What does it mean to repair a DL fault?
- 16. Feature maps Failures grouped by features 16 Luminosity Min Radius Orientation Mean Lateral Position MNIST BeamNG 16 Number of feature map cells fi lled with misbehaviours Number of failures
- 17. DL faults and failures Fault repair 17 DL faults and failures Q1: What is a DL fault? Q2: What is a unique DL failure? Q3: What does it mean to repair a DL fault? Training Data Change Fix labels of training data TCL Remove Augment training data TRD Unbalance training data TUD Add Remove noise to from training data TAN Make Fix output classes that overlap TCO Hyperparameters Change Fine-tune batch size HBS Decrease Fine-tune learning rate HLR Change Fine-tune number of epochs HNE Disable Fine-tune data batching HDB Activation Function Change Change activation function ACH Remove Add activation function ARM Add Remove activation function to layer AAL Regularisation Add Remove weights regularisation RAW Change Change weights regularisation RCW Remove Add weights regularisation RRW Change Fine-tune dropout rate RCD Change Fine-tune patience parameter RCP Weights Change Fine-tune weights initialisation WCI Add Remove bias to a layer WAB Remove Add bias from a layer WRB Loss function Change Change loss function LCH Optimisation Function Change Change optimisation function OCH Change Change gradient clipping OCG Validation Remove Add validation set VRM
- 18. DL architecture repair SE vs DL community 18 [AutoTrainer] Xiaoyu Zhang, Juan Zhai, Shiqing Ma, Chao Shen: AUTOTRAINER: An Automatic DNN Training Problem Detection and Repair System. ICSE, pp. 359-371, 2021. [Hebo] A. I. Cowen-Rivers, W. Lyu, R. Tutunov, Z. Wang, A. Grosnit, R. R. Gri ffi ths, A. M. Maraval, H. Jianye, J. Wang, J. Peters, Haitham Bou Ammar: Hebo: Pushing the limits of sample-e ffi cient hyper-parameter optimisation, JAIR, vol. 74, pp. 1269-1349, 2022. [Bohb] S. Falkner, A. Klein, and F. Hutter: Bohb: Robust and e ffi cient hyperparameter optimization at scale. ICML, pp. 1437-1446, 2018. Fit surrogate Sample hyperparameters Train & evaluate model Initial hyperparameters Bayesian hyperparameter optimization [Hebo, Bohb] Until training budget is over Rule based repair [AutoTrainer] Symptom → Solution S1: Add batch normalisation layer S2: Substitute activation function S3: Add gradient clip S4: Substitute initialiser S5: Adjust batch size S6: Adjust learning rate S7: Substitute optimizer Hyperparameters H1: activation function H2: optimizer H3: batch size H4: number of epochs H5: loss function H6: learning rate H7: weight initialization H8: batching (enable/disable) H9: number of neurons in a layer
- 19. DL architecture repair Are we there yet? 19 IR = Improvement Rate The random baseline produces comparable or better patches than other repair techniques, but the e ff ectiveness of tools varies depending on the fault, which justi fi es the need for future work to fi nd more e ffi cient ways of exploring the hyperparameter space. IR = Mpatch − Mfault Mfix − Mfault
- 20. DL faults and failures Take-away messages 20 DL faults and failures Q1: What is a DL fault? Q2: What is a unique DL failure? Q3: What does it mean to repair a DL fault? DL faults a ff ect data, model architecture, hyper-parameters, training process, not just the code. When exposing/counting failures, diversity in the feature space matters. Exposing more failures does not necessarily mean higher e ff ectiveness. The landscape of repair operations is not well understood and rule-based/bayesian techniques are not substantially superior to random exploration of such a space.
- 21. DL inputs Test input generation 21 DL inputs Q4: How to generate DL test inputs? Q5: How to prioritize/select DL test inputs? Inputs classified as 5 Frontier of behaviours Validity Dom ain Frontier Input Pair 5 5 5 9 21 6 9 6 4 6 1 Validity Domain Validity Domain LQ HQ LQ HQ DeepJanus
- 22. DL inputs Input validity 22 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
- 23. DL inputs Auto-encoder validity 23 latent space Encoder Decoder Training set Anomaly set [ ] (e > 𝜃 )? “All three testing techniques [DeepExplore, DLFuzz, DeepConcolic] studied produced signi fi cant numbers of [autoencoder] invalid tests” Input validity approximated as Auto-encoder validity
- 24. DL inputs Validity of automatically generated inputs 24 Test Generator Comparison Auto-encoder validity Human validity Label preservation Automated vs Human Test Generators Root causes of invalidity: Raw input generators: corruption of large number of pixels Latent space generators: lack of continuity in latent space Model based generators: low quality model All generators: label preservation not ensured
- 25. DL inputs Test input selection and prioritization 25 DL inputs Q4: How to generate DL test inputs? Q5: How to prioritize/select DL test inputs? Train set Train model Active set Prioritize and select Selected data Label data
- 26. DL inputs Uncertainty estimation 26 pip install uncertainty-wizard github.com/testingautomated-usi/uncertainty-wizard
- 27. DL inputs Evaluation metrics for test input prioritization 27 6 9 8 9 6 5 1 2 3 APFD = 9/18 6 6 5 8 9 9 1 2 3 3 APFD = 15/18, +50% 3 3 1 1 1 { Any permutation gives the same APFD APFD is actually measuring just top-3 accuracy Diversity matters!
- 28. DL inputs Take-away messages 28 Generating DL inputs that trigger failures is easy, but ensuring their validity and assigning them a valid label is not. Model uncertainty can be used to select and prioritize test cases. Traditional test case prioritization evaluation metrics, like APFD, do not account for DL failure diversity. DL inputs Q4: How to generate DL test inputs? Q5: How to prioritize/select DL test inputs?
- 29. DL oracles System level quality 29 DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations? lateral position headway position steering braking traf fi c signs O(x) = m1(x) ≤ t1 ∧ … ∧ mn(x) ≤ tn Oracles depend on thresholds ∀x ∈ X, mD 1 (x) ≤ t* 1 ∧ … ∧ mD n (x) ≤ t* n No FP in (manually) veri fi ed executions X t* 1 , …, t* n = arg max t1,…,tn |{j ∈ M|∃x, m Mj 1 (x) > t1 ∨ … ∨ m Mj n (x) > tn} No/few FN when executing on faulty models (max mutants killed) 0.1 0.9 is_red
- 30. DL oracles Runtime oracle 30 DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations? Autonomous driving Supervisor Safe disengagement Train set latent space Encoder Decoder Histogram and theoretical densities data Density 0.02 0.04 0.06 0.08 0.10 0.12 0 10 20 30 40 50 gamma CDF 10 0.12 Q−Q plot 𝜃 Train set Auto-encoder based anomaly detection
- 31. DL oracles Runtime oracle: eutoencoders 31 DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations? (e > 𝜃 )? Nominal Original Reconstructed Anomalous l l
- 32. DL oracles Runtime oracle: uncertainty estimation 32 DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations? github.com/testingautomated-usi/uncertainty-wizard
- 33. DL oracles Take-away messages 33 DL oracles should capture system level quality, not just extreme cases of misbehaviours. Runtime supervisors are needed to deal with rare or unexpected conditions in which misbehaviours are possible. DL oracles Q6: When is a DL program correct? Q7: How to prevent runtime oracle violations? Thresholds for system level oracles should ensure no FP and few FN on mutants.
- 34. A framework for DL testing Dev-to-production data shift 34 Goal Repair faults, exposed by choosing inputs whose execution violates oracles, to increase reliability measured in production Inputs Program Fault Execution Failure Oracle Repair Production test set Production environment Reliability Boolean Correctness ⟹ System Reliability
- 35. A framework for DL testing Implications for researchers 35 E ff ectiveness of test input generation, test prioritization and DL model repair should be measured as reliability increase, on production data We need ways to collect/simulate production data We need techniques to estimate reliability in a sample e ffi cient way Correctness, i.e. no error on the test set or 100% accuracy, is not a realistic goal. Production time reliability should be estimated instead.
- 36. A framework for DL testing Putting all together 36 Inputs Program Fault Execution Failure Oracle Repair Production test set Production environment Reliability DL faults a ff ect data, model architecture, hyper-parameters, training process, not just the code. When exposing/counting failures, diversity in the feature space matters. Exposing more failures does not necessarily mean higher e ff ectiveness. Generating DL inputs that trigger failures is easy, but ensuring their validity and assigning them a valid label is not. DL oracles should capture system level quality, not just extreme cases of misbehaviours. Runtime supervisors are needed to deal with rare or unexpected conditions in which misbehaviours are possible. E ff ectiveness of test input generation, test prioritization and DL model repair should be measured as reliability increase, on production data
- 37. Paolo Tonella Software Institute, Università della Svizzera italiana, Lugano, Switzerland The Road Toward Dependable AI Based Systems 37