In Search for a SAT-friendly Binarized Neural Network Architecture
Authors: Nina Narodytska, Hongce Zhang, Aarti Gupta, Toby Walsh
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We focus on the improvements we proposed for the neuron level of granularity. We use three datasets from (Narodytska et al., 2018; Baluta et al., 2019; Khalil et al., 2019). For each, we experiment with two tasks. First, we check if there is an untargeted adversarial attack (Goodfellow et al., 2015). Second, we compute an approximate number of adversarial examples for a given image. We used 4 hidden layers and one output layer. The input of the network is an image of size 28 28. The network has the following dimensions for all layers: [784, 500, 300, 200, 100, 10]. This gives 623K parameters, which is 3 to 10 times bigger than the networks in (Narodytska et al., 2018; Baluta et al., 2019). We used a full-precision trained network to seed weights before binarized training (Alizadeh et al., 2019). As the first layer inputs are reals, we used an additional BN + sign layer after the |
| Researcher Affiliation | Collaboration | Nina Narodytska VMware Research Palo Alto, USA nnarodytska@vmware.com Hongce Zhang Princeton University Princeton, USA hongcez@princeton.edu Aarti Gupta Princeton University Princeton, USA aartig@princeton.edu Toby Walsh UNSW Sydney, Data61 Sydney, Australia tw@cse.unsw.edu.au |
| Pseudocode | No | The paper describes algorithms and mathematical formulations but does not include any clearly labeled pseudocode blocks or algorithm listings in a structured format. |
| Open Source Code | No | The paper does not provide any statement about releasing source code for the methodology described, nor does it include a link to a code repository. |
| Open Datasets | Yes | We use three datasets from (Narodytska et al., 2018; Baluta et al., 2019; Khalil et al., 2019). For each, we experiment with two tasks. ... BNNs MNIST FASHION MNISTBG % #prms % #prms % #prms Vanilla 96.5 623K 82.1 623K 74.3 623K Sparse 96.4 32K 84.1 37K 78.2 41K Sparse+Stable 95.9 32K 83.2 37K 78.3 38K Sparse+L1 96.0 20K 83.7 35K 78.4 36K Sparse+L1+Stable 95.2 20K 82.9 37K 80.0 34K |
| Dataset Splits | No | The paper frequently mentions "test accuracy" and refers to "trained networks", but it does not specify any explicit validation dataset splits (e.g., percentages, counts, or a dedicated validation set) used during training. |
| Hardware Specification | No | The paper mentions using a SAT solver, an approximate model counting tool, and an ILP solver, but it does not specify any particular hardware components such as GPU models, CPU types, or memory sizes used for the experiments. |
| Software Dependencies | No | The paper mentions several software tools like "Py SAT tool (Ignatiev et al., 2018)", "Glucose as a SAT solver (Audemard & Simon, 2018)", "Approx MC3 (Soos & Meel, 2019)", and "CPLEX solver". However, it does not specify explicit version numbers for these software dependencies, only citing the papers they are described in. |
| Experiment Setup | Yes | We used 4 hidden layers and one output layer. The input of the network is an image of size 28 28. The network has the following dimensions for all layers: [784, 500, 300, 200, 100, 10]. ... We used the value of T = 0.03 for MNISTBG and FASHION and T = 0.04 for MNIST using 60%80% quantile depending on the dataset and the accuracy of the resulting network. ... We used five values of epsilon, ϵ {1, 3, 5, 10, 15, 20}. ... We used default parameters of Approx MC3. |